A Fusion Algorithm for Pedestrian Anomaly Detection and Tracking on Urban Roads Based on Multi-Module Collaboration...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Abnormal human behavior detection in traffic is crucial but challenging due to complex environments and detection errors.
• Existing methods struggle with interference from complex backgrounds and small targets.

Purpose of the Study:

• To propose an integrated AI methodology for robust abnormal human behavior detection and tracking in complex traffic scenes.
• To enhance detection accuracy and tracking stability by combining advanced deep learning models.

Main Methods:

• Utilized YOLOv8 as a baseline, integrating Swin Transformer for global feature modeling.
• Incorporated CBAM and CA attention mechanisms, GSConv modules, and Focal-EIoU loss function.
• Enhanced tracking with BoT-SORT-ReID by incorporating ReID feature distances and behavioral category information.

Main Results:

• Achieved mAP@50% of 92.2%, precision of 90.75%, and recall of 86.57% on a custom dataset, with a speed of 328.49 FPS.
• Demonstrated generalization on the UCSD Ped1 dataset with a 92.7% mAP, outperforming mainstream models.
• Tracking achieved MOTA of 90.8%, MOTP of 92.6%, and a 47.6% reduction in ID switches (IDS).

Conclusions:

• The proposed integrated YOLO-SGCF and BoT-SORT-ReID algorithm significantly improves abnormal human behavior detection and tracking in complex traffic environments.
• The methodology offers superior performance, accuracy, and tracking stability compared to existing state-of-the-art approaches.
• The system demonstrates effectiveness in identifying and tracking specific abnormal behaviors like climbing, falling, fighting, and phone usage.