Dense Pedestrian Detection Based on GR-YOLO.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Dense pedestrian detection is crucial for safety in public areas like airports and stations.
• Current deep learning methods struggle with feature extraction, multi-scale variations, and high false positive rates.

Purpose of the Study:

• To propose an improved dense pedestrian detection algorithm, GR-yolo, based on Yolov8.
• To enhance feature extraction, information fusion, and detection accuracy for dense pedestrian scenarios.

Main Methods:

• Implemented GR-yolo, optimizing the backbone with the repc3 module for enhanced feature extraction.
• Reconstructed the Yolov8 neck using an aggregation-distribution mechanism for efficient multi-level information fusion.
• Utilized Giou loss for improved convergence and target localization accuracy.

Main Results:

• GR-yolo demonstrated superior performance compared to Yolov8 across multiple datasets.
• Achieved accuracy improvements of 3.1% on Wider People, 7.2% on CrowdHuman, and 11.7% on People Detection Images.
• Successfully reduced missed detection rates in dense and multi-scale pedestrian environments.

Conclusions:

• GR-yolo is effective for dense, multi-scale, and scene-variable pedestrian detection.
• The algorithm offers a promising approach for real-world dense pedestrian detection challenges.
• The proposed enhancements provide new insights for improving pedestrian detection systems in public safety applications.