Visibility-Guided Dual-Branch Network for Occluded Person Re-Identification

Abstract

Person re-identification (Re-ID) is a fundamental task in intelligent video surveillance, with widespread applications in urban security and intelligent transportation. However, in real-world scenarios, occlusions caused by pedestrians or environmental objects often degrade visual features, significantly impacting the robustness and accuracy of Re-ID systems. To address this challenge, we propose a Visibility-Guided Dual-Branch Network (VGDNet) for person occluded Re-ID. The network integrates a visibility-aware mechanism to dynamically guide feature extraction, adaptively balancing the reliance on global and local features under varying occlusion levels. Specifically, a global complementary branch is designed to capture holistic semantic information through explicit–implicit feature mining, while a dual-path local branch enhances fine-grained representations via localized reorganization and concatenation. To optimize feature discrimination, a multi-objective joint loss function is introduced by combining triplet loss, label-smoothed identity loss, and multi-similarity loss. Extensive experiments on three public benchmarks—Occluded-Duke, DukeMTMC-ReID, and Market-1501—demonstrate the effectiveness of the proposed method, achieving mAP/Rank-1 scores of 57.1%/64.3%, 83.1%/91.7%, and 91.0%/96.4%, respectively. Ablation studies and visualization further validate the contribution of each module in mitigating occlusion effects and enhancing feature robustness.