Joint detection and re-identification for occluded person search.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Person search is crucial for intelligent video surveillance but is hindered by occlusion, leading to missed detections.
• Existing methods struggle with accurate pedestrian identification across multiple cameras.

Purpose of the Study:

• To propose a novel method, DINOPS, for jointly optimizing pedestrian detection and re-identification.
• To improve the robustness of person search systems against occlusion and enhance cross-camera identification.

Main Methods:

• Developed DINOPS based on the DINO detection model, incorporating a part attention mechanism.
• Integrated global and local pedestrian features to reduce missed detections.
• Employed multi-scale feature fusion for richer pedestrian representations in re-identification.

Main Results:

• Achieved a 6.6% increase in mean Average Precision (mAP) on the PRW dataset.
• Improved top-1 re-identification accuracy by 0.6% compared to state-of-the-art methods.
• Demonstrated significant improvements in both detection and re-identification tasks.

Conclusions:

• DINOPS effectively enhances person search capabilities by addressing occlusion challenges.
• The proposed method offers superior performance in pedestrian detection and re-identification.
• DINOPS represents a significant advancement for intelligent video surveillance systems.