Gait Recognition System For Human Identification - Report Writing - Assessment Answer

Gait Recognition System Assignment

Assignment Task

Write a report on the topic :-

Gait recognition system for human identification

Introduction

1.1Background Biometric Systems and Gait Recognition System

Biometrics is a well known and emerging technological research field using which a particular person can be identified by using his behavioural or psychological traits. Hence, there is high demand for efficient and effective automatic authorization and authentication of human in environments that are security sensitive.Resources of biometrics such as signature, fingerprints, iris, hand geometry and speech have been studied by various researchers in - depth and used in wide range of applications in security domain. The above mentioned resources of biometrics have their own benefits and limitations and one of the significant limitations is the failed authentication because of images of low resolution. Also, another significant limitation is that these systems require user / target co – operation to obtain effective results. In case of forensic domain, if the target uses gloves and mask, then capturing of fingerprints and face is not possible. Also, if the target turns his / her face opposite to the position of camera, then in such situations, face cannot be captured effectively.However, cameras have the capability of recording gait of the target individual effectively. Also, effective capturing of human gait is possible from a permissible distance and also does not require co – operation from the target individual. Human gait consists of information about the target individual’s physical condition and his / her psychological situation. Hence, in certain situations, information on gait is more than sufficient to identify the target individual and thus making it difficult for them to hide, fake or steal.

However, identification of humans from a certain permissible distance is an extremely difficult and challenging work to carry out and has been a very popular topic for the researchers in the domain of computer vision. The sequences of gait of different people is can be completely different and hence, this feature of gait of human body characteristic makes it a significant factor that can be used effectively in the identification of human from a certain distance.The term “gait recognition” clearly shows the significance of identification of a person by making use of video sequence when the person is walking. Thus, the significant feature in the world of security is the unique identity of an individual. Gait is defined to be an individual’s unique way of walking.Thus, the significant features of gait are the camera placed at a certain distance and without the target person’s knowledge. Human gait is hence considered to be very complex yet distinctive / unique locomotive pattern used in Biometrics that consists of synchronized movements of various human body parts including the joints and interaction among the parts.In the year 1973, Johansson conducted a psychological research discovery and clearly proved that humans have the capability to recognize their walking friends that is dependent on the light markers that are fixed to them.Since then, many researchers have been continuously researching on this topic of gait analysis and were successful to prove that the feature of gait is extremely useful to recognize people. Also, in addition, gait is considered to be an obstructive type of biometric that can be captured from a certain distance and without the need of any kind of intervention from the user.

There are several covariate factors which directly affect the overall performance and efficiency of the gait recognition system. They are – duration, light illumination, camera view – point, load carrying, apparel of subject and the person’s speed of walking. Therefore, the above mentioned covariate factors make the gait recognition technology a very challenging research topic. Recently, most of the researches based on gait recognition system focused on making use of view - invariant gait recognition system. Considering the realistic situation of surveillance, the subjects / targets are expected to walk in different directions in order to reach their destination. Thus, human identification from different random viewpoints is a very important pre – requisite for the different tasks that include access control, covert security and perceptual interfaces that is meant for intelligent environments etc. In order to achieve optimal efficiency and performance, the gait recognition system should make use of maximum number of cues as possible and then combine them altogether in meaningful ways.

1.2An overview of Gait – Biometric System

From the perspective of technology, the biometric – based gait recognition system can be combined into three different groups namely – Machine Vision (MV), wearable sensor and floor sensor based. According to the research statistics, various firms tend to prefer the Machine Vision (MV) – based recognition systems because of its features that include cost effectiveness, simplicity, non – intrusive and continuous authentication. Every individual possess unique and idiosyncratic pattern of walking, which can be easily identified and understood from the viewpoint of biomechanics. Basically, the human ambulation contains synchronized and integrated movements of various joints and muscles. Although, basic patterns of bipedal movement are similar for all humans, there is significant variation in the gait of persons in few details that include magnitudes and relative timing. Figure – 1 shows the basic block diagram for biometric authentication process.

As already mentioned in the previous section of Introduction, there are basically two different types of biometric characteristics. They are – (1) behavioural characteristics and (2) psychological characteristics. Psychological characteristics include – DNA, face, fingerprints, iris etc and whereas, the behavioural characteristics are gait and voice. As the results obtained from using psychological characteristics are not very effective, gait recognition is preferred over these approaches. Recognition of a person using gait is done by identifying him / her by the way they walk. Gait recognition system technology also provides effective results for images of low resolution.

1.1Applications of Gait Recognition System

Following are some of the important applications of using Gait Recognition System for Human identification –

• Automated systems for person identification for the purpose of monitoring applications and visual surveillance in various environments that are security – sensitive such as – airports, parking lots, banks etc
• Gait recognition system also finds extensive applications in the field of forensics. For instance considering a real – time situation where in Denmark reported a major bank robbery case, the court made use of gait analysis and its system as a significant tool to identify the robber
• The Gait Recognition system technology not only finds applications in the field of forensics, but also it can be used in medical diagnostics. For instance, basic recognition of walking pattern of a person can be considered to be an early sign of Parkinson’s disease, Normal Pressure Hydrocephalus (NPH) and multiple sclerosis
• Also, Gait Analysis can be used in professional sports training so as to optimize and enhance the athletic performance
• Gait analysis also find applications in recognition of genders
• Gait recognition system can also be extremely useful in identifying the exact direction of a walking person.

2.Literature Review

This section of the Research Thesis focus on discussing various important aspects related to Research thesis such as research on human gait, human gait versus other types of biometric traits and analysis of gait system as a multi – biometric component etc.

2.1 Analysis of Human gait VS different types of biometric traits

Considering various biometric traits, researchers found that gait possess some very unique features. As a unique and well known biometric trait, the highlighting characteristic of gait is the obstructiveness; because in comparison to other biometric traits, gait of the target subject can be captured easily from a certain distance and does not even require consent of the target subject.Other popular biometric types such as face recognition, finger print recognition, iris, signature, voice recognition and hand geometry etc compulsorily need physical contact or close distance with the target subject in order to capture the feature and thus, gait recognition also has the major benefit of difficulty to steal, fake or hide. Various parameters based on kinesiology which defines the gait forms an important basis for identification. But, there are certain significant disadvantages as it becomes very difficult to identify, capture and analyse the parameters that affect gait recognition system. Thus, the system of gait recognition has to be dependent on video sequence that is captured in uncontrolled or controlled environments. Various researchers who conducted study on gait recognition system said that, the gait features changes with respect to time and also there are certain significant factors that affect the quality of gait, such as clothes, walking speed, footwear, emotional condition and the surface on which the target subject walks. The above mentioned factors are the limitations to accuracy of the entire gait system and thus, put a question on its deployment as a distinctive biometric system. Before choosing gait system for human recognition, it is important to understand the application type and its purpose and then compare the gait recognition system with other well – known biometric methods based on certain factors such as appropriateness, invariability and uniqueness. Analysis for the comparison of gait system with other biometrics must be carried out under various subjects and also changing conditions to obtain a wide range of results. Thus, for this reason gait technology is not used as the only means for human identification especially in large database applications and hence is considered to have high potential as a valuable technology in the multi – modal type of biometric system.

2.2 Analysis of gait system as a multi – biometric component

Study conducted so far by the Researchers in the gait recognition field has clearly shown that the gait recognition system when combined with biometric systems can be a reliable system in human identification. Considering iris, fingerprint and palm methods as a different class to biometrics, the range of biometric systems that can be used in combination with gait recognition system as a multi – biometric system are foot pressure and face. In order to measure ground reaction force as the foot pressure, a specialised device will be required.

Foot pressure and gait has been easily used to narrow the DB (database) of subjects in a typical multi – biometric system. Early researches (year 2005) on gait system suggested on using three different biometric systems that are – gait, face recognition and foot pressure and form a multi – biometric system that is used for human identification. Few researches carried out in past focused on utilizing combination of face recognition and gait system and few other researches utilized the combination of ground floor reaction and gait system to achieve effective outcomes. These results showed that gait recognition system for human identification yielded effective results when it was combined with facial features directly in a multi – modal system. Enhanced performance was observed in researches that utilized combination of gait and face recognition and gait alone. Thus, previous researches concluded that effective and efficient outcomes are obtained by combining gait system with other biometric systems.

Figure – 3 shows various stances that are typically observed in a gait cycle. The silhouettes in Figure – 3 have been taken from the database of CMU’s MoBo. Though there are significant differences in the walking styles, but there is similarity in the walking process for all human beings. Generally, most researches focused on choosing only four walking stances – (1) right double support, (2) right mid – stance (3) left double support and (4) left – mid stance. And using these stances, a gait cycle is described as the interval that is recorded successfully between two stances (either left / right). Half cycle is defined as the interval that is successfully recorded in between any two mid – stances consecutively. And thus, the gait period is defined as the interval of time specifically in which a gait cycle is carried out. The frequency of walking is hence called as the frequency of fundamental gait.

2.3 Research on Human Gait

Computer vision was basically the first and very popular method that was utilized for the measurement of gait. During the initial days of research, a vision approach was developed and that proved to be a practical method to track humans in big and open spaces such that cameras are fixed at higher distances above the ground. Another research project made use of Markov models in order to properly sequence the several movements of a human while he is walking. Computer vision was also used as a people counting system. The limitation of using computer vision was physical attribute or requirement of a particular camera vantage point. Hence, in situations where it is not possible to meet the requirements of computer vision, other methods are more significant.

This gave rise to one such approach that made use of continuous wave radar. The major benefit of using wave radar was to identify the human presence or absence in a particular location. Few researches made use of accelerometers as a secondary non – vision approach. Hence, many associated research projects made use of accelerometers to recognize a human based on the gait system.One such research approach was slightly different and in which the accelerometer device was physically attached to the human / target subject at the back or leg. It is important to note that each method of biometry is sensitive to various attacks and the type of attack actually plays role in identifying what type of biometric signature to be used. Ultimately factors that determine the type of method to be used for an application are the application environment and the specific requirements.

In order to develop various methodologies for human gait identification, researchers initially focused their study on understanding the basics of human gait and its anatomy. Human gait is defined to be periodic movement and the period between walking is defined as the gait cycle. Figure – 4 shows a gait cycle.

Referring to Figure – 4, it is clear that the directional forces shown are only Y and Z because the directional force in x – axis is the right or left axis which passes through the figure. Every gait cycle consists of unique features of human’s walk and this unique feature or characteristic forms the gait signature. Thus, researchers mainly focused on developing different effective methods to identify a human’s gait signature and extract the target subject’s gait cycles which is obtained from a continuous signal of data.

2.4 Research on Gait Recognition System

User authentication is considered to be the first significant step towards avoiding unauthorized access. The process of confirming or verifying an individual’s identity is known as user authentication. User authentication using passwords was the traditional approach but had the limitation of easily hacking the passwords and thus, it was replaced by biometric authentications.Biometrics is a technological domain that makes use of automated methods in order to identify and verify an individual. User verifications and authentications are very critical and compulsory in various real – time applications such as airports and banks etc and thus, requirement of effective and efficient biometric identification methods in real – time applications are used. As discussed in Introduction section, characteristics of biometrics are of two variations. They are – (1) Physiological and (2) Behavioural.

• Physiological – Physiological biometrics are the ones that are obtained from direct measurement of part of target subject’s body. The most significant types of these measures are fingerprints, iris, face recognition, DNA, palm print etc. Basically these are related to human’s body. ^xiii
• Behavioural – Gait and voice are the features that are associated with the behaviour of an individual. Extract features that are based on an action carried out by an individual, form an indirect measure of feature of an individual. The significant feature of behavioural biometric is the usage of time as the main metric. Well – known and successful measures include speech patterns and keystroke – scan and thus, the biometric identification process is designed and developed in a way to make it an automated process. Extraction of features manually is both time consuming and undesirable as acquiring huge amount of data and then processing the data to develop a specific biometric signature is not feasible. Inability to extract the required features automatically which would make the process unrealistic and infeasible on real – time data sets in a real – time application. ^xiii

Gait Analysis is the process of carrying out a systematic study of human movement that is basically supported by instrumentation for the purpose of measuring movements of body, muscle activities and body mechanics.Following are the significant advantages of using gait – based recognition system in applications that are based on video surveillance.

• Human recognition that make use of gait technology does not need any cooperation from the target subject or the user
• An individual’s gait features can be captured effectively from a certain distance
• Very high quality images are not required in gait recognition and thus, good results are obtained even in low resolution images

Researchers have carried out in – depth study on the various methods or approaches that can be implemented and used in gait recognition. Few basic approaches that are used for gait recognition are – gait recognition based on moving video, gait recognition system based on floor sensor, gait recognition based on wearable sensor etc. in the first approach that is gait recognition based on moving video, gait of the target subject or user is captured by making use of a video – camera and is captured from a certain distance. Methods that are generally adopted to extract the features of gait for human identification purpose are image processing and video processing for instance, static body parameters, cadence and stride etc. In the second approach named, gait recognition system based on floor sensor, there is implementation of a set of force plates or sensors on the floor and the function of these sensors is to calculate the features related to gait when the target subject walks on them. Factors that are calculated using this method are maximum value of heel strike for both time and amplitude. Lastly, in the third approach named gait recognition based on wearable sensor, features of gait are gathered by making use of body worn sensors of the type motion recording (MR). The motion recording (MR) sensors record the gait acceleration and this is used for the process of user authentication.

Generally, there are four steps in the gait recognition system. They are - (1) Background Subtraction (2) Pre – Processing (3) Feature Extraction and (4) Recognition. Following is the detailed description of the four basic steps that are involved in the gait recognition system.

• Background Subtraction – In this method, initially the objects that are moving from the background are identified. After the identification of background moving objects, some techniques that are based on background subtraction are applied to it. The traditional approach that is generally carried out is background subtraction which clearly finds out the moving objects from a part of video frame which is different from the one in background model. Background subtraction give rise to features known as binary silhouettes that are black and white type of binary images (moving pixels). Background subtraction is included in a group of methods that is specifically meant for the segmentation of objects of interest which finds application in domains like as surveillance. Generating an efficient algorithm for background subtraction is very challenging. Few challenges associated with generation of effective background subtraction algorithm are – firstly, the background subtraction algorithm should be highly robust against the illumination changes. Secondly, the algorithm must avoid the detection of background objects (non – stationary) such as rain, snow, moving leaves and shadows of the objects that are in motion. Lastly, the internal background model must be capable of quickly responding to background changes for example stopping and starting vehicles.
• Pre – Processing – Initial step in the recognition of gait feature is the segmentation of silhouette. Pre – processing is carried out on the video frames in order to decrease the presence of noise and later some filters are effectively applied and upon which, the image frames are blurred and hence, it helps in effective removal of shadow after carrying out the motion detection type of pre – processing. This helps in delineating the foreground in image from background. As mentioned in previous step, background subtraction give rise to binary images which consists of white and black moving pixels and hence post processing is carried out in order to acquire normalized type of silhouette images consisting of less noise. This process makes use of morphological operators such as erosion and dilation to cover up small spaces within the silhouette and to filter small noise that is present in the background. In order to decrease the computational cost, researchers developed innovative silhouette representation framework that make use of only few pixels that are present on the contour.
• Feature Extraction – The Feature Extraction step is an innovative form process that is meant for dimensionality reduction. Considering the situation where in the input data dimension is too large to be processed and is considered to be redundant in nature, then the input data is successfully transformed into features vector, which is the representation of reduced features set. The process of transformation of the input data into the features set is known as Feature Extraction.
• Recognition – Recognition is the final step that is carried out in human recognition by making use of gait. In this process, comparison is carried out between input videos and the sequences that are saved in DB (database). Wide range of classifiers is used in the process of recognition that includes MDA, which stands for Multi – linear Discriminant Analysis and LDA (Linear Discriminant Analysis). MDA method is basically used to carry out the optimization of gait features segregation.

Gait Feature Recognition – The Gait Recognition System is used in recognition of unauthorized target subject (individual) and then carries out the comparison of gait of target subject with the sequences that are saved prior and hence recognize him. The most commonly used approach in gait recognition system is the background subtraction method. By making use of background subtraction method, pre – processing is carried out to decrease the noise content. Background subtraction method is again segregated into two different techniques. They are – (1) recursive approach and (2) non – recursive approach. Non – recursive method makes use of an approach called sliding window for background subtraction, but whereas, the recursive methods make use of Gaussian mixture model and single Gaussian method. Basically there are two different parts in the gait recognition method. They are – (1) Training part and (2) Testing part. Figure – 5 shows the gait recognition system in the form of a block diagram.

Figure – 5 Basic Gait Recognition System Block Diagram

Laboratory of gait analysis consists of various cameras of the type infrared or video which are mounted around the treadmill are connected directly to a PC. The target subject has certain markers that are placed at different points of the body such as ankle malleolus and pelvis spines. When the target subject moves through the treadmill, the trajectory of all markers is calculated immediately in three different directions by the PC. In order to calculate the bones movement, a specific model is used.

In general there are two categories of gait recognition systems that are commonly used. They are – (1) model – based methods and (2) appearance – based methods. Among the two above mentioned methods, the appearance – based method has a certain limitation of frequent changes in the appearance which entirely changes the viewing directions or walking directions.Whereas in case of model – based methods, the movement of human body is extracted by means of fitting the models to the respective input images. Model – based methods are scale and view invariant which clearly reflects on the kinematic features of walking manner. Basically, a gait is considered to be made up of a sequence of a human’s kinematic features and almost all systems recognize its existence by finding the similarity of these features.Human recognition using gait have various unique benefits over other traditional biometrics such as iris, fingerprint, face, palm prints etc. These benefits include perceivable at a certain distance, non – invasive and non – contact. Thus, the innovation based on gait characteristics is seen as a high – end technology for future in domains like an intelligent surveillance security system that is video – based. Gait possesses its own pros and cons but despite that it is considered to be an effective technology that can be implemented in security domain for the purpose of human identification from a certain distance.

1.Detailed description of various approaches on gait recognition system

In this section of Thesis Report, we have presented a detailed description of various methods or approaches designed and implemented by researchers on Gait Recognition System for Human Identification. Various methods discussed in this section are as follows:

• Use of modified type of independent component analysis for gait recognition
• Gait recognition system design using Neural and Fuzzy Logic concept
• Gait recognition system using the concept of inner angle of triangle
• Use of extraction and fusion of Global motion features for gait recognition

1.1Use of modified type of independent component analysis for gait recognition

In this specific research project, a system for gait recognition has been proposed by the researchers for the purpose of human identification by making use of a concept of MICA (Modified Independent Component Analysis). In this research project, three important modules were presented. They are – (a) Human detection module with tracking feature (b) Training module by making use of Modified ICA and (c) Human recognition module. The algorithm was thoroughly tested on a database of the type NLPR gait which consisted of images of target subjects who were walking in different directions in an outdoor surrounding.

The gait recognition system that is proposed in this system features gait that is in terms of gait signature that is directly computed from the silhouettes sequence. The entire system is considered to be a general type of pattern recognizer that is made up of three most important modules as mentioned above.Primarily, the target subjects (humans) are segregated and tracked in every frame of the provided video sequence, which is the tracking module. Identification of the person’s identity is done by carrying out training and as well as testing processes by using MICA on the feature vectors that are extracted from the module of pattern recognition. Figure – 6 shows the proposed block diagram of the Gait Recognition System.

• Detection of Human and process of Tracking

The first step in the process of gait recognition is the detection and tracking of target subjects (humans) that is done using a video sequence. The system that is being proposed in this section functions by assuming that the captured video sequence is from a non – moving camera and the subjects that are in motion in the video sequence are humans (target subjects). After acquiring a sequence of video from a non – moving camera, the function of this module is to detect and track the silhouettes that are in motion. Again the process is divided into two sub modules. They are – (a) Foreground Modelling and (b) Human tracking by making use of a process known as skeletonization operation.

Foreground Modelling – The approach that has been applied extensively in the process of foreground modelling is Background Subtraction, where in a static camera is used to acquire moving scenes. It is quite critical to reliably develop the background image from the sequences of video. The system that is proposed in this section adopted a less complex method of motion detection that is dependent on median value. Using this approach, the background is modelled using the video sequence. In order to develop the formula for approach, let P denote a sequence of video consisting of N image frames. Thus, the formula developed for construction of background P (x, y) is –

P (x, y) = median [P1 (x, y), P2 (x, y) ... P_N(x, y)]............................ (1)

The particular P (x, y) value defines the brightness of background which has to be manipulated in pixel location denoted by (x, y) and the median value is obtained from median. In the proposed system of gait recognition, the median value is calculated instead of calculating pixel intensity’s mean value over N video frames, because,

• Distortion of mean values is observed for major change in pixel intensities when there is motion encountered by the target subject and
• Comparing the median value computation and values of least mean squares, median value computation is quite faster. Both the conditions mentioned here hold good with the assumption in research that the target subject moves around continuously over the video frames. Factors that are given for the foreground modelling are the original image frames and extracted background. The role of algorithm of background subtraction is to subtract the background from the image frames in order to acquire the foreground objects in motion.

1. Human Tracking – The next step in gait recognition process is to effectively track the silhouettes that are in motion of the target subject which is obtained from the foreground image. In order to carry out human tracking, researchers adopted the morphological type of skeleton operator. The process of reducing the foreground regions that are present in the binary image to a skeletal form of remnant is called Skeltonization.
2. Training using MICA-Researchers made use of an approach called MICA (Modified Independent Component Analysis) in order to extract the gait characteristics and also train them. The aim of providing training to the skeletonised type of silhouettes with the MICA or Modified ICA is to achieve a specific range of independent components in order to effectively represent the features of original gait from a measurement space of high – dimension to an Eigen space of low – dimension. The ICA concept can be regarded as the developmental of PCA (Principal Component Analysis) and its primary objective is to efficiently define a group of variables by making use of basic functions, where in the factors are independent either statistically or as much as possible. The aim of ICA is to find out the vectors which describe information and data in terms of reproducibility, but the above defined vectors might not consists of any kind of information that is required for effective classification and may not include information of discriminative type.

• Human Recognition System

Considering the trained set of MICA in hand, the last step in the research was to examine the effectiveness of the entire system of gait recognition. The traditional problem associated with the gait recognition system is the pattern classification that can be effectively solved by manipulating and comparing the similarities that exists between the instances of test database and the training database. Gait can also be defined as the motion pattern of spatiotemporal type. Hence, the gait video sequence that is coming from the input side is transformed into equal magnitude of parametric eigen space by making use of modified ICA feature.Then, the gait recognition is achieved by measuring the similarity which is computed between the test sample in eigen space and reference patterns. Researchers made use of L2 Norm Distance to calculate the similarity between the two obtained gaits. The formula to calculate the L2 Norm Distance is as follows –

D = ||P_R– P_N||²₂, where, || ||²define the L2 norm, P_Ris defined to be the reference pattern and the new pattern of video sequence is defined by P_N.

1.1Gait recognition system design using Neural and Fuzzy Logic concept

In the second research project, the researchers proposed a new system for gait recognition. In the proposed system, initially the binary silhouette of target subject who is walking is effectively identified from each frame. Next step involved extracting the feature from every frame by making use of the operation of image processing. The important characteristics considered in the research are – step size length, center of mass and cycle length. Finally, the neural network is used in the research in order to train and test the proposed system. Researchers were successful in creating various model of neural networks based on certain important factors such as – setting of different parameters for training purpose, hidden layer and choice of training algorithm. All the experimentation processes were carried out using the gait database. Different sets of datasets of training and testing provide different outcomes.

A group of inter – connected neurons is defined as a neural network and finds extensive application in the universal approximation. An inter – connection of artificial neurons forms the artificial neural networks. Artificial neural networks can be extensively used for acquiring an understanding of the neural networks of biological type or can also be used for calculating the problems of artificial intelligence without developing the actual system model of a real biological system. It is already known that the real biological system is extremely complex and thus, the artificial neural network based algorithms step in to modulate this complexity and hence aim to hypothetically understand what matters most in an information processing domain.

3.2.1 Artificial Neural Network and its Architecture

The general artificial neural network’s architecture contains three different types of layers of neurons. They are – input layer, hidden layer and the output layer. In the networks of feed – forward type, the direction of signal flow is strictly from input side to the output side and is in a feed – forward direction only. The process of data processing can be extended to various layers of units but there is absence of feedback connections.Basically in the feed – forward type of networks, the network’s dynamic properties are extremely important. In a few situations, the unit’s activation values go through a process of relaxation in a way that the network is made to emerge to a stable state in which there are no changes in the activations anymore.

3.2.3 Feed forward type of neural networks

Feed – forward type of artificial neural networks (ANNs) permit the signals to traverse in a single direction, which is from input to output. Loops of feedbacks are not present, which means that the output pertaining to any layer has no affect on the same layer. Thus, the artificial neural networks abbreviated as ANNs are considered to be straight forward type of networks that link inputs with the outputs. Artificial neural networks (ANNs) find extensive applications in pattern recognition. These kinds of arrangements are also known as top – down or bottom – up methods. Kinds of feed – forward types of neural networks are – radial basis function, multi – layer perceptron and single – layer perceptron.

3.2.4 Single – layer perceptron

A single – layer type of perceptron is the simplest type of neural network that contains only one layer of output nodes; the output nodes are directly fed by the input nodes via a range of weights. Thus, it is said to be the simplest form of neural network of the feed – forward type. The weights product is added to the input and this calculation is done in each node and if the final value in each node is above a certain threshold (like 0), then the neuron fires and updates its value to activated one (like 1), otherwise it takes the value which is -1 (deactivated). Neurons having the features of activation functions are also known as linear threshold units or artificial neurons. In the research report, researchers made use of the word perceptron for the networks which contains only one of the units. Development of a perceptron can be done by making use of any values for the states of activation and deactivation as long as the values of threshold lie between the mentioned limits. The output values of most of the perceptrons is either 1 or -1, having a value of threshold as 0 and researchers have presented an evidence that the training of such perceptrons can be done in a short span of time when compared to the networks that are developed from the nodes having different values for activation and deactivation. Delta rule is the basic learning algorithm that can be applied to train the perceptrons. The delta rule estimates the presence of errors between the sample output data and the calculated output and hence makes use of this value to develop an adjustment to the network weights. In this way, a gradient descent is implemented. Using the single – unit perceptrons, it is only possible to learn patterns that are linearly separable.

3.2.5 Multi – layer neural networks

The class of Multi – layer neural networks contains computational units in multiple layers that are basically inter – connected in the pattern of feed – forward type. Every single neuron that is present in one layer is directly connected to the neuron node in subsequent layers. In various applications the units of multi – layer neural networks the activation signal applied is a sigmoid function. The theorem developed on universal approximation for neural networks defines that, each continuous function which effectively maps the real numbers intervals can be manipulated closely from a perceptron of multi – layer with a single hidden layer. This output result is applicable for a few restricted classes of functions of activation, for instance, sinusoidal functions. The multi – layer networks make use of wide range of learning techniques and the most commonly used technique is back – propagation

The values obtained from the network output are compared with the precise answer to calculate the values of few pre – defined error functions. By using wide range of techniques and methodologies, the error is further fed back through the multi – layer neural network. Using the available information, the algorithm effectively manipulates the weights that are present at each nodal connection so as to minimize the value of error function by some delta amount. If this process is continuously repeated for a huge number of training cycles, the multi – layer neural network basically tend to converge to a stable state where in the error value and its related calculations are minimized. In such situation, one can clearly say that the network has effectively learned a specific target function. In order to manipulate the weights properly, one must apply a basic method for non – linear type of optimization which is named as gradient descent. To achieve this, the error function’s derivative is calculated with respect to the network weights and then the weights are altered in such a way that the error decreases. It is thus quite clear that the algorithm of back – propagation can only be simulated on the multi – layer neural networks that consist of differentiable type of activation functions.

3.3. Gait recognition system using the concept of inner angle of triangle

The system of gait recognition is a multi – stage system that is meant to carry out various tasks in various steps as shown in Figure – 8.

The first step is to capture the video sequence that is captured by the camera. The position of camera is maintained such that the target subjects (humans) are covered and also another important factor to be considered is that the installed camera must be able to capture one gait cycle minimum as shown in Figure 9. The next step in the gait recognition system is the conversion of video into images which is done for one cycle known as gait cycle. The third important step in the gait recognition system is the feature extraction after which the classifier is taken into account. The role of classifier is to compare the gait based results that are stored in the database and display it if the match is correct and print the result. If the results do not match, then the classifier has the option to store the new person’s gait image in the database.

Figure – 9 shows the Gait Cycle. Referring to Figure – 9, it is clear that every gait cycle is segregated into two different portions that are – swing and stance. The gait portion that is covered by stance is 60 percent that starts from the initial contact till the area of toe off. And whereas, the portion covered by the swing phase in gait image is 40 percent that starts from the toe off area till the initial contact area. Thus, there is continuous repetition of these cycles.

Figure – 9 The Gait Cycle that clearly defines the Stance and Swing portions

In the proposed system of gait recognition, researchers chose the lower part of the target subject body for the process of recognition. As described in the Figure – 10, researchers chose both the feet and hand and after the process of feature extraction a triangle was developed as shown in Figure – 11.

Figure – 11 The triangle ABC that was developed after the process of feature extraction

Following is the nine – step algorithm adopted and implemented by the researchers in the gait recognition system using the concept of inner triangle –

Step – 1: An input video must be taken for a single cycle

Step – 2: The video must be converted into coloured frames

Step – 3: Every coloured frame must be converted into a grey scale

Step – 4: The pixel values namely – (x1, y1), (x2, y2) and (x3, y3) must be extracted and these denote the pixel values for hand and right foot and left foot respectively

Step – 5: A logical triangle must be developed in between the pixel values (x1, y1), (x2, y2) and (x3, y3) of every frame desired

Step – 6: The angles of logical triangle must be calculated using the suitable formula

Step – 7: The mean of every angle (?1, ?2 and ?3) must be calculated for one cycle

Step – 8: The mean’s result must be printed

Step 9: Exit the process

Here ABC is a triangle which is made using the co – ordinate points (x1, y1), (x2, y2) and (x3, y3). Here – x1, x2, x3 and y1, y2 and y3 define the pixel values. The angles in the triangle ABC are: ACB, ABC and BAC. The hand point is represented by point A, left feet is represented by point B and right feet is represented by point C. Values for the triangle edges are AB, BC and CA which is represented by a, b and c respectively. Following are a set of formulae to calculate the triangle edges a, b and c:

• a = sqrt ((x2 – x1) ^2 + (y2 – y1) ^2)
• b = sqrt ((x3 – x2) ^2 + (y3 – y2) ^2)
• c = sqrt ((x3 – x1) ^2 + (y3 – y1) ^2)

The angles BAC, ABC and ACB are respectively represented by ?1, ?2 and ?3 and by using sine and cosine formulae, the angles of triangle ABC are calculated as shown below –

Using the cosine formula:

?2 = cos ^-1((a2 + c2 – b2) / 2 * a * c ------------------- Calculation of second angle

By using the sine formula:

a / sin ?1 = b / sin ?2 = c / sin ?3

?3 = sin ^-1((c * sin ?2) / b) ------------------- Calculation of third angle

By making use of the property of triangle, ?1 + ?2 + ?3 = 180

Thus, ?1 = 180 – ?2 – ?3 ------------------- Calculation of first angle

Following is the Figure – 12 which shows flow chart proposed by the researchers for gait recognition system using the concept of inner triangle:

Figure – 12 Flow chart proposed by the researchers for gait recognition system using the concept of inner triangle

3.4 Use of extraction and fusion of Global motion features for gait recognition

3.4.1 Introduction

There are significant differences in the way every human walks and these differences are very important in understanding the individual. Considering only a person walking in a video sequence, there is high generation of valuable data in the correlated frames. It is assumed that the person walking in a particular video sequence which is effectively segmented, frame by frame, develops only a single class, where every frame is a component of this class. Thus, the main objective of this research is to develop a framework that has the capability of recognizing an individual from the way he walks. Human body motion can easily be identified by using various techniques of image processing. The proposed system in this research must be capable of acquiring global information about an individual’s movement. This is achieved by using four different models of video images of segmented type, before actually combining all the outcomes into a single model known as GBM (Global Body Motion). This model must enhance the biometric recognition rates. Researchers presented the following key contributions of the proposed system:

• A new framework developed for the human gait recognition – that consists of the GBM (Global Body Motion) and the proposed work, that conduct analysis on the information about target subject movements which is captured by four unique segmentation of video based images. As all individuals have their own unique features, the main objective of this research is to only capture the global information on body movement
• Combination of features – individual recognition is carried out by effectively combining four models of individual human movement. The function of each model is to process different information such as binary images (2 – D), greyscale images (3 – D), movement and motion of joints, the contour of the human body’s silhouette. The proposed fusion process in this research adds up the similarity indexes acquired from the individual ratings of every model.

3.4.2 Proposed Methodology

1. a) The GBM Model (Global Body Motion)

Figure – 13 below shows the GBM Model (Global Body Motion) that is presented in this research paper. The required features extracted from the model develop new independent models such as SSW, SGW, SEW and SBW of the global body motion and are individually compared. The proposed system takes into account a group of frames that are segmented of every target subject as a different class and every frame is considered as an object of this class. Also, the complete segmented object is effectively used to develop vector features, but in various methods the gait recognition is carried out by segmenting the target subject’s body parts.

In order to exclude the motion’s consequent segmentation and the background, the algorithm or law that was based on GMM (Gauss’ Mixture Model) was implemented. Utilizing this segmentation, two different types of images are created. They are –

• The first type of image refers to the one taken in segmented movement called greyscale and its sequence is named as Silhouette – Grey (SG)
• The second type of image developed is acquired from binary mask that is created by GMM and its sequence is named as Silhouette – Binary (SB)

Figure – 13 shows the general outline of the proposed GBM Model (Global Body Motion).

Figure – 13 The general outline of the proposed GBM Model (Global Body Motion)

In every segmented frame, a pre – process is conducted that is in a person’s images who is walking without an actual background, to filter the noise content and to fill small holes in the image. Figure – 14 shows the outcomes of segmentation by making use of GMM.

Figure – 14 Outcomes of segmentation by making use of GMM

Erosion and dilation types of morphological operators that are present in the OpenCV library are used for this purpose. The frames are to be efficiently centralized and for which primarily the system finds the center of mass by using the equation (1). The image size has dimensions n x m. Co – ordinates (x, y) defines the center of mass and the image intensity function is described using F (i, j).

(x, y) = 1 / n, m summation (n, i = j = 0) {(i, j). F (i, j)} --------------- (1)

Every image is properly framed to the required size, considering the window center as the image’s center of mass, having the dimensions of 124 x 240 pixels.

Scale Reduction Process – (Haar WT)

In order to decrease the scale, three different alternatives or options were tested. They are – the Wavelet Transform, an averaging filter and the Gaussian filters. In order to achieve benefits and better cost in the algorithm’s operation speed and complexity, researchers decided to decrease the scale by using Wavelet Transform. The co – efficient image (approximation) that is scaled by WT, maintains the silhouette’s low frequency information.

Among all the orthogonal type of wavelet transforms, the basic one is Haar WT (Wavelet Transform). O (n) is the complexity function of Haar WT. By the application of Haar WT, the image can be easily divided into 4 different sub – bands consisting of different information both in terms of detail and content. For every decomposition level, there is creation of four new images and each image has the scale and spatial resolution half of it. Every level generates an output image that is resulted from the low – pass type of filtering process and three other images are developed from the high – pass filtering process. The low – pass filters generates a low resolution type of approximation and the system images obtained from the high – pass filters are the ones with details of horizontal, diagonal and vertical. For this research, the co – efficient image (of the approximation type) is used as it consists of all the required information about shape of the image and variations in grey level.

Considering that the original image (segmented one) consists of all important data and information about the human body’s global movement while in motion and there is no significant change in the information with scale, and hence, at two different levels, the Haar WT is applied to the segmented sequences. After the application of Haar Wave Transform (WT) to the SG sequence, there is generation of three sequences. They are – SGW sequence, SB sequence and SBW sequence. The segmented sequences that are made up of subjects of each class, generate output images of dimensional pixels of 31 x 60, with the target subject walking centralized in every frame. Basically the scale reduction involves decreasing the magnitude of data without reducing the global information amount that is present in the movement. In this way, the computational effort of gait recognition is optimized. Thus, the two different models of the system of gait recognition are obtained. They are: SGW model and SBW model. The SG sequence derives the SGW model after the Haar Wave Transform (WT) is applied at two levels and whereas, the SBW model is obtained by the application of Haar Wave Transform (WT) to the SB sequences.

Movement of the skeleton and contour

In order to effectively acquire the global variations of the human body movement or motion, that is present only in the silhouette’s contour, there is application of Canny edge detector in the sequences of SBW to obtain the sequences of SEW as output. Figure – 15.a shows the resulting classes that are generated from the contour movement. By applying the algorithm developed by Lam et al, skeletonization of SBW sequences is achieved. This process successfully generates SSW which is the skeleton form of sequence classes. This is depicted in Figure – 15.b. The major benefit of using the above mentioned methods, is that they effectively decrease the amount of information of redundant type without affecting the information quality significantly regarding the global movements of human body.

Figure – 15.a Resulting classes that are generated from the contour movement

Figure – 15.b Generation of SSW which is the skeleton form of sequence classes

Thus, researchers were successful in generating 4 complete models. They are described as follows –

• The SGW Model – SGW model is described as Silhouette – Gray – Wavelet model. Each class in this model is defined by the sequence of a silhouette greyscale by making use of Haar Wavelet Transform (WT) that is applied to objects in motion which are segmented using GMM. The SGW model consists of all the important information about the human gait’s global movement in three – dimension with the variations of greyscale, but with light variations there is quite increase in sensitiveness.
• The SBW Model – SBW model is described as Silhouette – Binary – Wavelet model. In this model, each class is defined by a sequence consisting of binary silhouettes that were created using Haar Wavelet Transform (WT) and applied to objects in motion that are generally segmented by GMM. The significance of SBW model is that it provides a detailed information on the global movement of human body’s silhouette in two – dimension while the target subject is walking. The limitation of sensitivity to light variation is overcome by the SBW model but can be affected by the clothes worn by the target subjects and thus clothes are considered to be a variable factor which can negatively affect the performance of gait recognition system.
• The SEW Model – The SEW model is called as Silhouette – Edge – Wavelet model. In this model, the representation of each class is done using a sequence of silhouettes of edge images that are acquired from SBW model. The main function of SEW model is to carry all the necessary information contours global behaviours while in motion. Compared to the above two models, the SEW model is comparatively more immune to the variations in light. Also, it is important to note that the information that is available in the contour is not satisfactory for effective gait recognition.
• The SSW Model – The SSW model is called as Silhouette – Skeleton – Wavelet model. In this particular model, representation of each class is done by making use of a sequence of silhouettes of skeleton that are acquired from SBW model. Important information that is contained in the SSW model is global movement of target subject’s joints (human body) and their typical behaviour while in motion.

Figure – 16 shows a clear illustration of all the above mentioned 4 classes. They are – SSW, SBW, SEW and SGW. Every model generates output information that is associated with the particular features of the global movement and which are affected by various factors such as – walking angle, clothes worn by target subjects, light conditions etc. Also, it is very important to mention that the methodology proposed for the human gait recognition in this particular research consider all the correlations that are related to the present data output from each model (that are SSW, SBW, SEW and SGW) in the images (segmented), for consecutive frames and for the same target subject.

Figure – 16 A Clear illustration of all the four models: Extracting the sequence of SGW from SG sequence, similarly extracting the SBW sequence from SB sequence and obtaining the sequences of SSW and SEW from SGW sequence

Eigen Gait – Feature Extraction

For processing the correlated data, it is very important to apply few approaches for a process called data decorrelation. This is specifically applicable for human gait images as they involve substantial correlations among the few consecutive samples. PCA approach, which stands for Principal Component Analysis is the one that is extensively used for the purpose of data decorrelation. Elimination of redundant type of information is achieved by data decorrelation in each dimension. Thus, the main aim of PCA is to search an approach for transformation which is more compact and also more representative. Most of the times, explanation of data variance is done by decreased component numbers and also, the possibilities of discarding the other components is done without actually losing information relevant to the system and process. But, to decrease the amount of lost information during the process of PCA, the eigen vectors that are created are analysed during data processing so as to effectively maintain the overall performance of system.

As every sequence of frame defines a respective class of target subject walking, the variance in the intra – class is quite small but whereas, the variance in the inter – class is large. Hence, the technique known as PCA (Principal Component Analysis) is used in extracting the associated features for gait recognition. Variations that exists between the various samples is identified by examining the co – variance matrix of all samples in the group. The applicability of PCA technique is done for all the frames of each class that belongs to the above mentioned four models (SSW, SBW, SEW and SGW). Also, the dimensionality of data is decreased with respect to the original variables, without affecting the relevant information. The important extracted features include the feature vector which will be applied for the classification of silhouette in their own classes.

Every class that is generated from the sequence of frame (SSW, SBW, SEW and SGW) is placed into the sub –space PCA. In order to acquire a particular dimension, certain tests must be conducted that starts with a 32 – dimension and is hence increased in steps which lead to effective classification results. Eigen Gaits are the prototypes or average images of every class and are specifically used in the evaluation of “search sequence” similarity using the specific Eigen Gait in order to store the class features. As soon as the Eigen Face acquires the variance comparing to the faces of all classes, the temporal features, also known as temporal differences are effectively captured, which belongs to human gait among the each class’s frames and thus, these features are projected in a vector of prototype. This process is slightly different from the approach called as Eigen Gait proposed by the researcher named Ben Abdelkader who made use of SP (Similarity Points) among the image pairs, which are applied to the PCA in later stage.

Following are the important steps that must be carried out to obtain a set of EigenGait

• Formation of the training set – the images captured for the purpose of training set basically are the ones that come under the sets of models of SSW, SBW, SEW and SGW. Every image that is taken from the video sequence is considered to be a vector. For this, the original image is considered and their pixel lines are concatenated. For example – consider a matrix that is denoted by M having n rows and m number of columns (n x m), is considered as a vector denoted by S = (n .m) x 1. All the images that belong to the set are saved in the matrix denoted by T. Every column in the matrix denotes an image
• The training set T is considered and the averages of all the images in the set are calculated
• Then, the next step is to subtract the average from every original image in the matrix T and in this way, a new matrix called T1 is obtained. Using the matrix T1, B is calculated, where B stands for co – variance matrix.
• Then the next step is to calculate the eigen values and eigen vectors of the matrix B which is the co – variance matrix
• Considering the feature vector that is denoted by P and their components, the data dimensionality is reduced. Basically the eigen vectors of higher significance are chosen in order to develop the feature vector
• The average of all images belonging to all classes of P matrix is calculated. Then, the prototypes or images of each class are the Eigen Gait (denominated)

The similarity that exists between every class is calculated in this research work by making use of Euclidean metric.

Fusion

Researchers’ such as Jain, Hong and Pankanti have provided a proof and said that the collaboration of various technologies based on biometrics can provide effective results on the significant enhancement on the overall efficiency of the gait recognition system. Hence, in this way, false rejection rate (FRR) and false acceptance rate (FAR) are decreased. Information integration that comes from several biometric indicators can be effectively carried out during the process of feature extraction or decision steps.

Though a single technique of biometric is used, various motion representation options carry different data about the silhouettes and movement of the human body or the target subject. Also, in certain situations, the system is significantly vulnerable such as dress change, change in lighting conditions, shadow presence etc. but the researchers assumed that the fusion model proposed in the research can have the capability of adding few static features about the target subject silhouette, which is done in the context of four different models that are proposed SSW, SBW, SEW and SGW. Using the SSW model, the dynamic features of the movement of human body is analyzed.

Another assumption in the research paper is on the proposed fusion method, in which it is assumed that each model (SSW, SBW, SEW and SGW) output results in similarity score among the classes and each frame that is to be classified. The NN Classifier (Nearest Neighbour) is used to obtain the similarity score. Thus, each model is considered individually and the correct answers percentage is obtained. The representation of model of the single gait which results in better performance and efficiency will possess a higher weight in the classification decision of the frame.

Following are the steps that define the algorithm –

• The similarity is calculated that exists between the every j – th frame that belongs to test set and thus, the formula for Eigen Gait of each class (denoted by c) of model i is given by –

S = Min (||frame_j – EigenGait_c||)

Where, S is the smallest Euclidean distance that exists between the Eigen Gait of each class c of i model and the j – th frame. Model i have varying values from 1 to 4 for SSW, SBW, SEW and SGW. Division of frame is done in the class such that there is decrease in the mean distance

• The average precision is calculated for the correct answers for each i model and is given by –

?i = TPi / TGi

where – Tpi is the quantity of correctly classified frames of the i model and TGi is the overall number of test set samples of i model

• Next, the fusion score (? j,c) is calculated between the class c and the j – th frame and is given by the formula as shown below –

? j,c = summation (i = 1 to 4) ?i * S_j.c ⁱ/ summation (i = 1 to 4) ?i

where – ?i is the i - th model weight which is provided by the number of correct answers irrespective of all models and S_j.c ⁱis the similarity score of i – th model and for the j – th frame which is from the class c

The above equation is known as Fusion as it defines the similarity of weighted average between the model scores.

Materials

In order to implement the algorithms proposed, programs on Microsoft Windows were developed, by making use of Visual Microsoft C++ software of version – 6.0 having the library OpenCV (OpenCV stands for Open Source Computer Vision Library). Also, it involved use of 3 independent image databases so as to do validation on the proposed methodology.

• Database – A: In the database – A, recording of videos was done of surrounding with controlled illumination condition of people walking in a particular area. The resolution of camera chosen was 320 x 240 pixels and the captured video was kept at 15 FPS. In order to develop this database, images of about 10 target subjects were included which means, there were total 10 classes. Each class consisted of about three sequences of video that were recorded on the same day. All the sequences were effectively combined and a single sequence was formed that was normalised to about 100 frames. In every video, the target subjects were asked to walk in the same direction and in a uniform surface and the camera was fixed in such a way that it is perpendicular to the optic axis. Each video consisted of only one target subject.
• Database – B: the images for the Database – B were taken from the “Gait Database” of the NLPR institute (National Laboratory of Feature Recognition). The chosen images were in the jpg format. The images from gait database were developed in an open environment having natural light. Total views generated for the images were three in number: front, side and oblique (90 degree, 0 degree and 45 degree respectively). Every class consisted of three different views and per view there were total four sequences. The sequences were numbered as follows: sequence – 1, sequence – 2, sequence – 3 and sequence – 4 and the directions for these sequences were: right – left, left – right, right – left and left – right. Figure – 17 shows the angle variation in all the four sequences. In order to carry out the research effectively, the video sequences that were taken from the available images were gathered. Overall, twenty different classes were obtained which consisted of about 240 videos and about 8, 400 frames.
• Database – C: Database – C consisted of about 10 classes, which were also taken from the database of NLPR and were in the format of AVI. Further, every subject had three sequence types. They are – nm, bl and cg. The sequence of bg characterised the target subject who was walking on a plain surface and carried a bag on in his / her hand or his / her shoulder. Whereas in the cl sequence, the target subject was walking on the plain surface but did not carry any bag but wore an overcoat. The last sequence of nm characterized the target subject walking on the plain surface but did not carry any bag and the clothes worn by target subject were same as in the bg sequence. The angle variation for sequence of bg is shown in Figure – 17. Total number of cameras used to capture the videos was 11. Angle ? that lies between the target subject and the camera direction can take the values from 0 degree to 180 degree. Every target subject was asked to walk nearly 10 times in the same line and thus, total 110 sequences were generated.

Initially the video sequences of first base were used in the evaluation of proposed algorithms. The second database, which is the database – B, was used in the evaluation of method that was proposed by the researchers and lastly the third database, which is database – C was used in the evaluation of method proposed that is associated with the variation of angles.

Evaluation methods

For all the three databases that were used in research, every chosen image of each frame was focussed into a PCA type of sub – space and then it was compared to the prototype of type Eigen Gait of every class. For all the experiments carried out, there is generation of confusion matrices and hence factors for each class such as False Rejection (FR) and False Acceptance (FA) were calculated. Formulae to calculate the values of FR and FA are:

FA = NFA / NNV ------------- (a)

FR = NR / NT ---------------- (b)

Where – NFA denotes the number of frames that were falsely accepted

NNV – denotes the number of frames that are not true

NR – stands for the number of frames that are rejected

NT – stands for the total number of frames that are present in the class

The protocol of FERET type having a leave – one – out and cross – validation rule was effectively used in the analysis and evaluation of the results of the gait recognition system. Initially the purpose of developing the protocol was to evaluate the algorithms of face recognition but these protocols can effectively be used in different types of biometric systems. The performance (statistical) of this approach is regarded as Cumulative Match Score, which is abbreviated as CMS. CMS is described as the true class’s cumulative probability of a subject which belongs to the k hits that is closest. CMS provide information on for what division of testing samples there is appearance of correct answers in the k closest queries. Thus, the assessment of calculating the number of images which must be examined in order to achieve a pre – set performance level is allowed by the method.

After the evaluation of similarity differences that exists between the training set and the test sample, there is effective application of NN (Nearest Neighbour) to the entire classification. Also, the evaluation of GBM, which is the model proposed in research, for Database – A, a test of independent type is carried out for each sequence type - SSW, SBW, SEW and SGW. The configurations used to carry out the test are as follows:

Number of classes = 10, number of frames / class = 100, PCA Sub – space = 256.

In order to carry out the evaluation on performance of the proposed model of GBM using the database – B, three different experimentation procedures were carried out for each sequence type that are SSW, SBW, SEW and SGW. In all the sequences, the directions in which the subject walks was restricted to three angles. They are, 0 degree, 45 degree and 90 degree. For every movement direction, total four frame sequences were implemented. They are – 1, 2, 3 and 4 and the number of classes was 20 and the number of subjects was also 20. Following is the sequence combinations that were used:

• Angle 0 degree with sequences 1, 2, 3 and 4
• Angle 45 degree with sequences 1, 2, 3 and 4
• Angle 90 degree with sequences 1, 2, 3 and 4
• Angle 45 degree with sequences 2 and 4

For the range of experiments carried out for research, the elements number present in every class is obtained by summing up the frames of all the sequences (individual). For the 4 sequence combination (1, 2, 3 and 4), the total class elements that were used is 144. The correct answers obtained from each model are basically the weights that were implemented in weighted mean for the process of fusion. Evaluation of the GBM approach by using the database C, two different experimentation processes were carried out with the aim of creating sequences of each type, which are: SSW, SBW, SEW and SGW. The evaluation and experimentation was conducted using all the angles in every sequence, and for every angle in range of 0 degree to 180 degree.

In total, 70 frames were used in each class and the total number of frames were = 700. After carrying out the evaluation of every model individually, it was followed by application of fusion process.

1.Research Approaches and Research Methods

The research methodology chosen to carry out the thesis work on the topic “Gait Recognition System for Human Identification” is Qualitative Research Methodology. Though the research approach and topic involves high – level statistical analysis and calculations, qualitative research methodology was chosen over the quantitative research methodology. Because, the main aim of the thesis work was to conduct an in – depth study on the various research methodologies and also approaches adopted and implemented by other researchers on the topic of Gait Recognition System for Human Identification. Under the qualitative research methodology / approach the type of research chosen was Observation, where in I managed to gather wide range of credible resources from different resources and observed each and every resource carefully and later studied them in – depth to include the readings and findings of other researches in this thesis work. Validation of resources or sources was not required as all the sources chosen for thesis research study were from credible online journal articles and online textbooks. Thus, high quality resources and data were ensured by using the above mentioned resources as a fundamental base for conducting the research for completing the thesis.

The Research approach was kept quite simple in order to avoid any kind of hurdles that occur in the research process. The research approach involved choosing wide range of resources from credible sources such as online journal articles and online text books. In order to ensure high quality information gathering, the type of resources / sources to be used in conducting the research was limited to journal articles and textbooks. Each resource was studied in depth to understand the aim of the paper / research work and hence, comparing all the resources, four most effective approaches that are to be presented in the thesis work were selected.

2.Research / Project Design

Key features of the project titled “Gait Recognition System for Human Identification” are as follows:

• Discussion of four unique and most effective and efficient approaches designed and developed by various researchers on the gait recognition system
• Each approach is discussed and presented effectively in the report
• Results obtained from each research approach that include the output figures are also included in the report

Major project deliverables:

• An in –depth study on the traditional biometric systems and their limitations
• Rise of Gait recognition systems and their highlighting features and advantages
• Discussion of various approaches to achieving highly efficient and effective systems for gait recognition system.

3.Data Collection

As already discussed in the previous section of Research Approach and Research Methodology, the data required to conduct the research on thesis work was collected from various credible resources that are available online and this included extensive use of online journal articles and online text books. Each resource was observed carefully and in – depth study was carried out to ensure that all the important contents that are required to enhance the effectiveness of thesis are included. Validation of data that are included in the research was not conducted as they are individual or group work and also use of credible resources avoided the validation problems.

7. Results and Discussion

7.1 Use of modified type of independent component analysis for gait recognition

The experimentation involved extensive analysis in order to identify and explain the effectiveness of the gait recognition system proposed in the report and this was followed by an in – depth comparative analysis and discussion on the outcomes obtained from the research.

7.1.1 Data Acquisition

The experimentation related with the proposed system for gait recognition was carried out using the images that were available in NLPR (National Laboratory of Pattern Recognition) gait database. Following is the description of the gait database that was taken for research study in brief: The digital camera of the type – Panasonic NV – DX100EN was basically fitted on a tripod and was used for the purpose of capturing the gait sequences in open – air surrounding. The images obtained in the results refer to a single subject in the view – field without the presence of occlusion.

The target subjects who were involved in the research were told to walk in front of a stationary camera in three different directions, they are – obliquely (90 degree), frontally (0 degree) and laterally (45 degree) which is with respect to the plane of image. The obtained outcomes of NLPR gait database consisted of 20 target subjects and per subject totally four different sequences per view were obtained. The obtained image features include – full colour of 24 – bits, 352 x 240 is the original resolution of the image and per second total 25 fames were captured. Overall the gait database consisted of about 240 sequences in total. The each sequence length varied with time which was dependent on the target subject and the time he took to traverse the view field. Few such image samples are shown in the Figure -

7.1.2 Results

This section of thesis consists of the research results of first approach discussed in report. In order to train the MICA, the gait database of NLPR that is publically available was used. The following figures show the intermediate results that were presented in the gait recognition system.

The researchers clearly identified the effectiveness of entire proposed gait recognition system by using a group of gait images that are available in the database of NLPR. Calculations on FAR (False Acceptance Rate) and FRR (False Rejection Rate) associated with the gait recognition system are defined in the Table – 1.

Figure – 24 FAR Graph

7.3 Gait recognition system using the concept of inner angle of triangle

In order to acquire the results, the data sets of CASIA A were used for the experimentation purpose and involved 15 target subjects and the proposed system uniqueness was identified and checked. Results showed that the unique identification using the proposed gait recognition system using the concept of inner angle of triangle was obtained with 90 percent (%) CCR. Figure – shows the wide range of frames of gait cycle.

7.4 Results and the comparison with different methodologies

7.4.1 Results obtained from database A

For a set of 100 frames / class the Table - 2 illustrates the percentage of match between their respective FA and FR values

The percentage of right answers was 91.7 %, 83.6 %, 79.1 %, 69.6 % for models SGW, SEW, SBW and SSW respectively. These percentages were made use of as weights to determine the weighted mean in the fusion process.

Table - 2 Percentage of match with respective FR and FA rate for the database A

Table – 2 Percentage of match with respective FR and FA rate for the database A

The SSW and SBW sequences were linked for lower rate of right classifications. This proves that as the data about the global movement present in SGW model is much bigger than SEW, SSW and SBW, the GBM identifies the person from its gait much easier in the two first instances. But all the classes do not have this applied. On the basis of the features of the subject movement, the SSW outputs this data with higher accuracy than all other models.

There are many other factors that lead to an increase in FR and FA rates, such as:

• The amount of shadow occurrence in the image
• Lack of perfection in the collection of both false contours and contours
• Unfilled areas inside the images (small and tiny holes )

These aspects impact the correct class attribution in a negative manner as shown in Figure - 26. Figure 26a illustrates few frames whose classification is hampered by the lack of continuity of the edge points, and Figure 26b illustrates the impact of the shadow present in the image.

By making the dimension of the PCA sub – space greater, more right answers can be created, primarily for sequences of SEW and SSW images. The best outcomes are gained when the dimension of the PCA sub – space has a dimension of 256, as shown in Figure - 27.

Values that are more than 256 for the PCA sub – space create an increase in cost but that does not help in increased mean precision significantly.

Figue – 26 Frames that are poor – quality and impair right classifications. A Frames from the SEW model; b Frames from the SBW model

The models were also joined in the following ways: SEW – SGW, SGW – SBW, SGW – SSW, SBW – SEW – SGW, SGW – SBW – SSW and SBW – SEW – SSW SBW. But as per observation the best result was obtained by joining four models. Table 2 illustrates this result, showing the percent of match. The best outcome was achieved by joining four models together.

The CMS (Cumulative Match Score) is shown in Fig. 9, ranging from a rank between 1 and 20. In table 3, the outcomes of the right answers are shown, using the Cumulative Match Score, which has ranks varying from 1, 5 and 10 taking into consideration the SBW, SGW, SSW, SEW models and fusion approaches explained above.

By calculation of the statistics for GBM model, using the CMS (Cumulative Match Score), we get an accuracy of 99.7 % for the case that is best. The average right number of answers for every model are made use of as weights for calculation of the weighted mean with respect to the fusion module. In fusion procedure, the precision of right answers that belong to the highest would be 100 % for rank 5.

The CMS (Cumulative Match Score) curve in Fig 10 were achieved by the fusion of SBW, SGW, SSW and SEW models, through a mix and match of all the four walking seq. (Sequences 1, 2, 3, 4), that belong to the database B. Performance achieved by the GBM model had been compared to work made by Hong, Wang, Lee, Kale, Philips – Sarkar, BenAbdelKader and Collins. Also another point to note is that their work to recognize biometrics by use of GAIT uses our database B. The outcomes for the experiments conducted are shown in Table 6, along with the percentage of right classifications for ranks 1, 5 and 10 using the four walking seq.

By joining the models, an increase in performance can be observed for Rank 1, with a right classification rate of 97.1 %. The GSW model alone produces much better performance than all other published models.

The outcomes obtained from the GBM model for 90 degrees and 45 degrees angles were measured against the work of Kale and Hong. Rank from 1 to 5 and the four joined walking seq. (1-2-3-4) were made use of, for the 2 views, as in publication with comparison.

Post fusion process taking into consideration the 90 – degree angle, there was almost 100 % accuracy for rank 1. For angle of 45 degree, with least rate of right results, an increase of 0.7 % for Rank 1 was noticed.

GBM accuracy, taking into consideration 4 models individually (SBW, SGW, SSW, SEW) is more efficient than most of the published techniques in literature as illustrated in Table - 4.

These outcomes are improved on application of fusion module, thus confirming highest performance of our approach. The outcomes received for GBM model on application of side view (0 degree), with joining of 4 sequences (left - right and right - left), resulted in superior global statistical performance for SGW method.

Table - 4 Right answers making use of CMS (Cumulative Match Score), with ranks 1, 5, 10

7.4.3 Results for the database C

Taking into consideration the database C, best outcome was obtained with angles 0 degrees, 90 degrees, 108 degrees and 180 degrees. The aspect that quite strongly determines classification is angle of walking. In table 6, the right classification values were received using CMS (Cumulative Match Score) to produce best result in perspective of angle variation. This includes joining of two seq. with target subjects carrying bag and two seq. with target subjects wearing coats. The 1^stseq. produces best performance over wide range of angles. For 0 degree and 180 degree angle, the number of right answers is close to 100 % for rank 1. With respect to the 2^ndseq. in GBM model, the number of right answers is not more than right answers received from seq. target subjects were carrying bag. Taking this into consideration, we come to conclusion clothes and to great extent the coat, are much more problem causing than using bag for right subject classifications.

8.Conclusion and Future Work

The process of identifying people is known as recognition. Identification of humans using the Gait methodology is a process of finding an individual based on his walking characteristics. Basically, the gait recognition system is a type of biometric system and is purely associated with the behavioural features of the biometric system. Thus, the gait recognition system can be considered as a type of biometric technology which can be effectively used to monitor people without having their co – operation. The traditional types of biometric systems include iris, face recognition system, fingerprint recognition system, palm recognition system. The results of these mentioned biometric systems are quite high and are also reliable but they suffer the significant limitation of acquiring physical contact with the target subjects or individuals. Thus, these biometric systems cannot be used in high security areas. High security and controlled environments such as military installations, banks and also airports must have some technology and process / system to quickly identify and detect the presence of threats and hence provide different levels of access to all types of user groups. Gait recognition system provides a specific manner of moving on foot and hence the process of identifying the individuals based on the way they walk. Compared to other traditional biometric systems, the gait recognition is less obstructive biometric, which provides the possibility of identifying people from a certain distance without having co – operation or interaction from the target subject. This unique feature of gait recognition system makes it extremely attractive. In the research thesis, four different approaches that are already presented and implemented by other researchers have been discussed individually in – depth. The main purpose of this thesis work is to carry out analysis on the various approaches and methodologies developed by the researchers and hence, understand the concept of gait recognition more in – depth. Thus, it can be concluded that the gait recognition system is extremely effective and efficient method in uniquely identifying the target subjects by making use of their psychological and physiological state. For this reason, recognition of humans using the system of gait recognition at a certain permissible distance has successfully gained wider interest among the various researchers and their community. The four different approaches that are discussed in the research thesis have certain limitation of restricted gait variations and are evaluated on data sets that are very limited. Also, certain factors such as clothing, light variations, surroundings etc are the ones that put certain limitation on the accuracy and performance of the gait recognition system. The researches’ have not focused entirely on all the external gait features. The future work is to conduct an in – depth on the new set of approaches and methodologies associated with the gait recognition systems that have addressed most of the external gait features.