Person re-identification based on an improved DeepLabv3+ semantic segmentation network

Pedestrian re-identification is the process of recognizing and matching the same pedestrian in surveillance videos across different cameras or different time periods, which has been widely used in the field of intelligent security and security tracking. However, if the background is complex or persons are occluded, it often results in low accuracy. To solve these problems, we propose a person re-identification method based on improved DeepLabv3+ semantic segmentation, which mainly consists of two parts: semantic segmentation network and person re-identification network. Firstly, we propose an U-DeepLabv3+ semantic segmentation network by combing the advantages of the U-shaped Network(U-Net) and DeepLab version plus network(DeepLabv3+) to generate precise pedestrian masks, which can effectively eliminate background interference and provide high-quality inputs for subsequent person re-identification. Secondly, we propose an improved part-based convolutional baseline person re-identification network (IPCBReID), which leverages these masks to strengthen feature extraction and integrates an attention mechanism along with a "One-vs. -Rest" relational module to achieve deeper feature integration. Extensive experiments conducted on the Market-1501, Chinese University of Hong Kong (CUHK03), and Occluded-Duke datasets demonstrate the effectiveness of our approach. Compared with state-of-the-art methods, our IPCBReID achieves improvements in mAP by 0. 3 %, 0. 5 %, and 0. 3 % on above three datasets respectively, and increases Rank-1 accuracy by 1. 5 % and 1. 6 % on Market-1501 and Occluded-Duke. These results validate the effectiveness of our method in enhancing person re-identification accuracy.

Authors

• Fuzhen Zhu
• Donglin Li
• Huiling Li Hong Kong Baptist University
• Yong Liu
• Bing Zhu

Keywords

• Person re-identification
• Semantic segmentation
• DeepLabv3+
• Attention mechanism
• Intelligent security