Human Gait Recognition System Using Enhanced Deep Learning Architectures Under Covariate Conditions

Abstract

Gait recognition, a promising biometric modality, offers a robust and unobtrusive method for person identification but faces significant challenges that require innovative newlinesolutions. This research addresses these challenges by exploring advanced techniques in silhouette image pre-processing, gait cycle detection, and the application of deep learning approaches. Traditional image enhancement methods often struggle with issues such as low resolution and uneven illumination in silhouette images. To overcome these limitations, Modified Contrast Limited Adaptive Histogram Equalization (MCLAHE) is introduced, significantly enhancing the clarity and quality of silhouette images, thereby improving gait feature recognition accuracy. The Hamming Distance Correlated Gait Cycle Detection (HDCGC) algorithm is proposed to enhance the precision of gait cycle segmentation from noisy silhouette sequences, directly contributing to improved gait recognition performance. Various deep learning newlinemodels, including CNN, VGG19, Xception, and MobileNet, are evaluated, demonstrating the superior performance of the Xception model, particularly when combined with the HDCGC approach. To address covariate factors such as viewpoint variations, attire changes, and carrying bags, the Customized Visual Transformer (CViT) model exhibits remarkable robustness and adaptability, outperforming newlinetraditional CNNs like VGG-19 and Xception. Furthermore, recognizing the computational constraints of edge devices and smartphones, a Lightweight Convolutional Neural Network (LWCNN) is developed, which maintains high newlineaccuracy with reduced computational complexity, achieving a recognition rate of 97.95% on enhanced silhouette images. This study contributes significantly to the field of gait recognition by introducing innovative methods that enhance image quality, improve gait cycle detection accuracy, and optimize deep learning models for resource- newlineconstrained environments. Future research focuses on further optimizing lightweight architectures for deployment on ed