Pedestrian detection in dense scenes faces significant challenges due to the mutual occlusion between pedestrians and the small proportion of distant pedestrians in the image. In this paper, a pedestrian detection algorithm named YOLO-CPEE is proposed to address these challenges for dense scenes. First, the cascaded group attention is integrated into the C2PSA module resulting in a new module (C2CGA) to achieve a good balance between detection efficiency and accuracy. Second, a high-resolution feature layer (P2) is introduced in YOLO-CPEE to better capture shallow pedestrian feature details and location information of small objects. Third, to enhance the ability of multi-scale feature extraction, the cross-scale feature interaction and adaptive weight allocation are achieved through efficient multi-scale attention (EMA). Moreover, a new lightweight dynamic detection head (Detect-Efficient) is designed, which significantly improves the object detection ability while reducing the model parameters. To test the proposed YOLO-CPEE algorithm, a self-built dataset DensePed-LR and the public dataset CityPersons are taken as reliable evaluation benchmarks, which both are high-density crowd dedicated datasets containing challenging scenes for testing. On DensePed-LR, YOLO-CPEE achieves improvements of 4.1% in recall and 3.2% in mAP@0.5:0.95, along with a faster inference speed of 135.3 FPS relative to the baseline. On CityPersons, YOLO-CPEE further achieves improvements of 3.9% in recall and 3.2% in mAP@0.5:0.95, while reducing the model size by 2.6%.
Chen et al. (Wed,) studied this question.