Research and Performance Analysis of Person Re-Identification Methods in Occluded Scenarios

Pedestrian re-identification (ReID) has gradually become a core supporting technology in intelligent security systems. However, its practical deployment is still significantly hampered by occlusion issues. To tackle this key challenge of occluded person ReID, this paper reviews and categorizes existing research methods into three major types: data driven, model driven, and auxiliary information-driven approaches. Data-driven methods simulate real occlusion scenes and implement diverse data augmentation strategies to effectively compensate for the shortage of occluded samples in training data. Model driven methods focus on optimizing the feature extraction process removing interference from occluded regions via feature dropping, and enhancing effective feature representations via feature strengthening to improve the model’s occlusion resistance. Auxiliary information-driven methods leverage external techniques such as pose estimation and semantic parsing to accurately locate visible pedestrian regions, thereby facilitating effective feature matching. To objectively evaluate the performance of typical algorithms, this paper summarizes and compares their experimental results on four mainstream benchmarks based on published literature. It is shown that feature-level enhancement methods achieve outstanding performance in occluded scenarios and good generalization in non-occluded scenarios, yielding the best overall effectiveness.