Optimized and Improved YOLOv8 Dense Pedestrian Detection Algorithm
Jingkun Mao, Haoyu Wang, Deheng Li · 2024
In large, densely populated scenarios, abnormal human gatherings are common, and due to the presence of many small-scale occluded pedestrians, missed detections are likely. To tackle this challenge, an improved multi-scale occluded pedestrian detection algorithm based on YOLOv8n is proposed. Firstly, to address the large variation in the scale of occluded pedestrians, the algorithm introduces the lightweight network RepViT to enhance multi-scale features. Next, the GAM attention module is added to improve the representation of multi-scale images and positional information. Finally, to improve the recognition precision for minor targets that are often missed, a small object detection head is introduced, using Inner-WIoUv3 as the loss function to enhance the model's convergence capability. Experimental results show that the improved model achieves enhancements on the public dense pedestrian datasets WiderPerson and CrowdHuman, with mAP50 increasing by 1.8% and 5.1%, and recall rates increasing by 2.2% and 5.5%, respectively. The improved algorithm shows a higher level of detection precision when compared with the baseline model.