Dense Pedestrian Detection Algorithm Based on Multiscale Feature Fusion in YOLOv8

The pedestrian target in large-scale crowded places has the characteristics of complex environment, diverse scales and dense arrangement, which makes the current target detection methods appear the phenomenon of false detection and missed detection when dense pedestrians. In order to solve this problem, based on YOLOv8, a dense pedestrian detection algorithm based on multi-scale feature fusion was proposed. Firstly, in view of the limitation of the performance of standard convolutional convolution due to the defect of convolutional kernel parameter sharing, RFCAConv was introduced into the backbone network. The convolution module can effectively improve the performance of convolution operation, so as to improve the feature extraction ability of the backbone network and fully extract the feature information at all levels. Secondly, a $160 \times 160$ scale small target detection head was added to improve the model’s detection ability of small-scale targets. Finally, based on the ASFF design ASFF-4P module, QAFPN is proposed to fully integrate the feature information of four feature layers with different feature scales, effectively avoid the loss and degradation of information, reduce the distance of non-adjacent layer semantic information, and improve the detection accuracy and efficiency of the model for targets at different scales. Experimental results show that compared with YOLOv8n, the proposed method improves AP50% and AP50:95% by $3.7 \%$ and $\mathbf{4 . 2 \%}$ respectively on the CrowdHuman datasets, and is also highly competitive compared with other advanced pedestrian detection models.