A Cross-Perspective Gait Recognition Framework Integrating Breadth-First Search and Multi-Scale Feature Map Interaction

Abstract

Gait recognition is a key biometric technology with broad applications, yet cross-perspective variation severely impairs performance. This study proposes a novel gait recognition model that integrates a breadth-first search-guided feature propagation mechanism with gated recurrent unit-based temporal modeling and multi-scale spatial feature map interaction. The model enhances feature fusion across different layers and perspectives while selectively attending to key temporal cues through global max pooling. Experimental evaluations on the CASIA-B dataset demonstrate that the proposed method achieves an accuracy of 0.97, 0.94, and 0.91 under normal walking, carrying object, and wearing jacket conditions, respectively, significantly surpassing the baseline models in recognition performance. The method also obtains the lowest root mean square error of 0.09 and the fastest recognition time of 1.2 seconds. Compared with conventional convolutional neural networks and recurrent neural network-based architectures, the proposed model shows substantial improvements in accuracy, robustness, and computational efficiency. The key innovation lies in the introduction of a breadth first search-driven feature interaction strategy and a hierarchical temporal-spatial fusion structure, which jointly optimize the feature representation for robust cross-view gait recognition.