Development of Human Gait Recognition Techniques under Covariate Conditions for Surveillance Systems

Abstract

Human gait recognition draws significant attention in the field of artificial intelligence, and computer vision over the last four decades due to its unobtrusiveness, non-contactable, non-intrusive, and non-invasive characteristics. However, the recognition process faces challenges due to various covariates such as different observation views-angles, carrying objects, clothing variations, walking speeds, and health conditions. This thesis addresses the challenges posed by covariate variations in gait recognition using both vision-based and sensor-based data. The work initially focuses on mitigating the impact of covariates in vision-based gait recognition through an innovative feature selection process. A novel approach is introduced, integrating outlier detection with structural relationship analysis. This method enhances feature selection by identifying features that deviate from the core data distribution and employs a linear predictive classifier within a dictionary framework to improve individual identification accuracy. To further improve the efficiency of gait-based feature selection, the Gait Covariate Invariant Generative Adversarial Network (GCI-GAN) is designed to process gait data in the frequency-domain instead of relying on direct silhouettes, which are particularly Invariant Generative Adversarial Network (GCI-GAN) is designed to process gait data in the frequency-domain instead of relying on direct silhouettes, which are particularly susceptible to covariate influences such as view variations and carrying conditions. Moreover, the integration of an attention mechanism enhances the process by focusing on covariate-invariant regions, improving robustness under diverse conditions. In both identification and authentication, the method assesses the quality of each gait sample and employs an adaptive technique to handle covariates. This is achieved by incorporating the quality measures of individual gait samples into the authentication process and dynamically adjusting the identify selection threshold for optimal performance. To extend the exploration of covariate-invariant feature selection, the proposed Domain-Invariant Graph Convolutional Network (DiGCN) is designed to capture the complex dynamics of gait patterns. This approach models the interdependencies between motion vectors of different body parts by employing multi-hop transformations. These transformations aim to preserve information not only in the feature space but also in the structural space through a locality-preserving technique within a graph-based convolution framework. An attention-based mechanism is introduced to dynamically update node selection across both temporal, and spatial-domains. This mechanism enables the model to focus on relevant regions of the graph, enhancing its adaptability to complex motion dynamics. The propagation process jointly optimizes the features and structure of the nodes using predefined loss functions, ensuring both discriminative feature learning and domain-invariant identification. This unified strategy effectively addresses covariate effects while retaining the distinctive characteristics required for accurate recognition. This doctoral research has also explored different approaches to manage these covariates in gait recognition systems using sensor signals. The report has addressed challenges in smartphone-based gait sensor signals, such as capturing long-term sequences and the computational demands of large kernels. A weighted multi-scale deep ensemble CNN network (WMsCNN) has been developed to address covariate variations by employing independent subnetworks, each designed to process information at distinct temporal resolutions. During training, an innovative strategy is used to identify and retain only independent subnetworks, each designed to process information at distinct temporal resolution. During training, an innovative strategy is used to identify and retain only the most essential components of the network, ensuring efficient learning while discarding less critical elements. This approach significantly reduces computational overhead while maintaining high performance. Additionally, a specialized alignment technique minimizes differences across data distributions. Further, to solve overfitting problems commonly faced by sensor-based gait recognition systems due to variations, a multi-kernel temporal CNN network has been designed, incorporating a regularized channel attention technique termed as (ReChAtt), thereby further enhancing the performance and generalization of the model. All the techniques have been validated using publicly available different gait datasets, demonstrating a substantial increase in recognition rates compared to other recent approaches available in literature.