Gaitts: indoor gait recognition with multi-scale temporal-spatial information aggregation

This paper addresses the issue of insufficient attention to multi-scale spatiotemporal features in gait recognition algorithms for indoor scenes by designing a GaitTS (Gait Temporal-Spatial) algorithm that fuses multi-scale temporal and spatial dimension features. The proposed GaitTS algorithm processes temporal features by aggregating multiple frames of temporal information. Enhancing attention is paid to spatial features of different human body parts by partitioning feature maps into different scales. The dual focus equips the network with stronger capabilities for recognizing global and multi-scale local features, thereby improving the recognition performance of the model. Experimental results are implemented on CASIA-B, CCPG and occCASIA-B datasets. The results demonstrate that the proposed GaitTS algorithm achieves average Rank-1 accuracies of 97.9%, 95.4%, and 85.7% under normal (NM), walking with a bag (BG), and wearing a coat or jacket (CL) conditions for CASIA-B dataset, respectively, outperforming the state-of-the-art methods.