Key point-aware occlusion suppression and semantic alignment for occluded person re-identification

Abstract

Partial occlusion is a key factor affecting the performance of person re-identification (re-ID). Although some solutions have been specially designed for occluded person re-ID, the ambiguity of pedestrian appearances and complex backgrounds still pose great challenges. Therefore, a key-point-aware occlusion suppression and semantic alignment (POS) method is proposed in this study to alleviate the existing challenges in occluded person re-ID. This method consists of three main modules: key point-aware semantic alignment (KPA-SA), self-similarity guided feature discriminability enhancement (SGFDE), and global feature extraction under occlusion suppression (GFE-UOS). In particular, the proposed KPA-SA semantically aligns the activated areas corresponding to specific pedestrian key points (e.g., head, shoulders, legs, feet) in multiple channels of different images. In addition, according to the paired left and right pedestrian key points, a cross-fusion mechanism can be applied to information compensation to alleviate information loss in occluded areas. The proposed SGFDE utilizes the self-similarity of non-occluded information of the same pedestrian captured from different views to enhance the discriminability of pedestrian identity features and suppress interference information unrelated to pedestrian identity. The proposed GFE-UOS fuses the heat maps of different key points to suppress the negative impact of occlusion on global feature extraction. The comparative experimental results verify the effectiveness of the proposed POS and its superiority over state-of-the-art methods. The related source codes were released on https://github.com/huangdaichui/occ_reid.