Dense pedestrian detection algorithm in complex scenes based on improved Yolov8

In response to the problems of environmental occlusion, diverse pedestrian target sizes causing false positives, missed detections, and poor real-time and adaptability of models in current dense pedestrian detection scenarios, This article proposes an improved model YOLOv8n-LDC based on YOLOv8n. Firstly, this paper proposes and uses the LSKPF module to improve the neck network, effectively solving the problem of missing spatial pyramid feature information in traditional methods. Secondly, the CGA content-guided attention mechanism is integrated, and the DySample upsampling operator is introduced to realize dynamic upsampling of feature maps. The experimental findings demonstrate that the proposed algorithm enhances the P, R, AP₅₀ and AP_50:95 scores by 1.0%, 5.1%, 4.1%, and 3.1% respectively on the Crowdhuman dataset compared to the primeval YOLOv8n . Compared with other algorithms, YOLOv8n-LDC also shows higher accuracy, significantly improving the performance of pedestrian detection.