Abstract
Gait recognition is a developing biometric technique capable of identifying individuals from a distance, with wide-ranging applications such as video surveillance. A primary challenge is the extraction of discriminative gait features from silhouettes that are robust to variations in apparel, carried objects, and camera viewpoints. To address these limitations, this study introduces GaitSTR — a novel framework that harnesses pyramid mapping for enhanced temporal and spatial feature extraction, integrated with a deep neural network comprising dense layers. Pyramid mapping decomposes gait sequences into multi-scale spatial features, enabling GaitSTR to capture fine-to-coarse motion patterns and improve recognition under varying conditions. The method focuses on extracting distinctive feature representations at different frame levels, effectively utilizing spatial and temporal variations within video sequences. The proposed model utilizes a memory-augmented recurrent neural network (RNN) enriched with temporal attention to capture sequential motion cues, while spatial features are extracted through a densely connected attention-guided network By employing the pyramid-based hierarchical feature extraction, along with attention mechanisms in both spatial and temporal component, the network can prioritize the most significant video segments, improving its efficiency and learning capacity for processing intricate gait data. The results are evaluated on four widely used benchmark datasets: GREW, OU-ISIR, OU-MVLP, and CASIA-B—achieving 92.4% on GREW, 95.2% on OU-ISIR, and 0.96 mean accuracy on OU-MVLP, and 98.4% (normal) on CASIA-B, surpassing state-of-the-art methods. These results underscore the robustness of our approach under diverse conditions, establishing a new benchmark for performance in gait recognition.

This publication has 107 references indexed in Scilit: