GaitSFF: improving gait recognition performance based on selective feature fusion in video surveillance

Gait recognition is a vital biometric technique with broad applications across various fields, including video surveillance and infrared-based identification. It aims to identify an individual's unique walking pattern based on the periodic movements of different body parts. However, most existing methods extract features directly from video frames captured by optical sensors and assign equal importance to each body part, neglecting the critical role that specific body parts play in recognition. Therefore, these methods may fail to generate high-discriminative features. This paper proposes a feature selection and fusion network encompassing three main elements to address this issue. Firstly, this paper designs a selective part feature fusion module to extract discriminative motion representations. Secondly, this paper adopts a multi-scaled local time aggregation strategy to mine the temporal information and obtain high-quality time features. Finally, this paper uses a spatial attention mechanism to help the model focus on the most meaningful regions, which can further improve the accuracy performance of the proposed model. Experimental results on two public datasets, i.e., CASIA-B and OUMVLP, show the effectiveness of the proposed method and verify its superiority over several state-of-the-art approaches. Specifically, the Rank-1 accuracy of the proposed method for the normal walking, backpack walking, and walking with coat scenarios on the CASIA-B dataset is 97.8%, 95.4%, and 87.0%, respectively. Besides, the Rank-1 accuracy rate of the proposed model is 90.6% on OUMVLP, demonstrating its robustness across different optical sensing conditions.