An Improved Dense Crowd Detection Model Based on YOLOv8

With the increase of crowd monitoring and management needs, dense crowd detection has become an important research direction in the field of computer vision, mainly using target detection methods to detect the position and number of pedestrians in pictures and videos. Aiming at the problem of missed detection and misdetection due to the occlusion between pedestrians or the target pedestrians being too small in dense crowd scenes, this paper proposes an improved detection algorithm CscSCA-YOLOv8 based on YOLOv8. Firstly, the Conv-SPD-Conv module is used instead of each convolutional step and each pooling layer in the trunk and the neck, to prevent the loss of fine-grained information and the less efficient feature representation Learning. Then the SCAMix hybrid attention mechanism is introduced into the YOLOv8 network architecture to utilize the global perception capability of self-attention and capture local features via CoordAttention to improve detection in dense scenes. Experimental results show that this paper's method achieves 74.1% and 46.6% for mAP50 and mAP50-95 on the WiderPerson dataset, respectively, and improves 3.6% and 2.8% compared with the original YOLOv8n.