Computer Vision-Based GAIT Recognition System Using Deep Learning

Abstract

Gait identification is a process that seeks to recognize persons based on their walking patterns. It has an extensive variety of uses in investigation, forensics, and healthcare. Conventional gait identification systems depend on manually designed features that do not have good generalization ability. Recent advancements in deep learning have resulted in data-driven techniques for learning features, which have achieved the highest level of performance. This research presents a new computer vision-based gait recognition system that utilizes deep convolutional neural networks (CNNs). The system utilizes a two-stream CNN architecture to directly extract temporal and spatial characteristics from gait sequences. The spatial stream processes each frame individually, while the temporal stream collects the motion dynamics across several frames. Explicit attention modules direct the network's focus towards specific joint areas that are important for discrimination, and this focus is maintained throughout time. Additionally, a technique called temporally-weighted feature pooling is implemented to combine individual frame-level information into a condensed gait signature. Our methodology has been widely verified on four benchmark gait datasets, and the outcomes display that it executes much better than previous model-based and deep learning methods. Ablation experiments confirm the effectiveness of the different elements of our system. The suggested system offers a precise and effective framework for gait recognition based on vision, utilizing deep learning techniques.