Occluded Gait Recognition.

Gait recognition suffers from common occlusions in real-world applications. However, academic research on gait recognition usually assumes access to full-body input data. For bridging the gap to practical applications, we propose to identify people when given an occluded gait sequence, namely occluded gait recognition. Since publicly available datasets do not meet the requirements of the intended research, we design a new framework named OccSilGait to generate realistic occluded gait silhouette sequences based on the principle of perspective transformation. Specifically, OccSilGait considers various occlusion scenarios including non-occlusion, crowd occlusion, static occlusion, and detection occlusion. And we employ OccSilGait to build the occluded gait dataset OccCASIA-B for further research. To address challenges brought by occlusion for gait recognition, we propose a novel SpaAlignTemOccRecover network consisting of 1) a Spatial auto-Align module that transforms silhouettes into spatially aligned ones with well-designed self-supervision; 2) a Spatial-Temporal Backbone that alternatively extracts spatial and temporal features to avoid the diffusion of occlusion; 3) a Temporal Occlusion Recovery module that reconstructs the current frame based on time index and temporal context, exploiting gait periodicity for occlusion recovery. Experiments on the newly built occluded dataset show the superiority of the proposed method. Both the OccSilGait framework and the code are available at https://github.com/YunjiePeng/OccludedGaitRecognition.