YOLO-Human: A Simple and Effective Anchor-free Network for Dense Pedestrian Detection

Pedestrian detection is a crucial aspect of human centered services. However, detecting dense pedestrian scenes presents several challenges, such as pedestrians obstructing each other. Although CNN-based cascaded networks with anchor boxes and Transformer-based networks have achieved excellent results in dense pedestrian detection, their structures are complex and require substantial computational resources. To solve this problem, we propose a simple and effective CNN-based single-stage anchor-free network for dense pedestrian detection called YOLO-Human. First, we design a new anchor-free detection head that can predict the full-body boxes and visible boxes of pedestrians. Additionally, we consider pedestrians' visible box and full-body box information when using the SimOTA positive and negative sample assignment and computation loss methods. Finally, we propose a new weighted IOU loss to allow the network to focus more on pedestrians that overlap within classes and are difficult to detect. The experimental results showed that the proposed YOLO-Human method achieved high performance on the CrowdHuman val dataset. It obtained 93.44% AP, 34.96% MR^-2, and 86.28% JI.