Improved YOLOX for pedestrian detection in crowded scenes

In recent years, object detection in computer vision has developed rapidly. However, crowded pedestrian detection in object detection remains a challenging problem, especially in one-stage detectors where improved solutions are rare. In this paper, we propose a novel crowded pedestrian detection method called YOLO-CPD which works better than other one-stage models in crowded environments. Our method primarily enhances the ability of the one-stage detector to detect multiple overlapping objects in a single area. The core of our approach is to use boxes difference to adjust the IoU value of the Non-Maximum Suppression (NMS) and to improve the Intersection over Union regression loss (IoU Loss), with an Optimised Score Module (OPSC). Compared to the baseline, YOLO-CPD can improve the Average Precision ( AP ) by a 5.04% increase, Recall by a 2.17% increase and the log-average Miss Rate ( M R - 2 ) by a 5.12% reduction on the CrowdHuman dataset. In addition, YOLO-CPD also achieved good results in the WiderPerson dataset, demonstrating the strong generalisation capability of our proposed method.