关键词: 目标检测, 密集人群, 大核卷积, Transformer

Abstract: With the acceleration of urbanization, efficient and accurate monitoring of dense crowds has become a critical need for public safety and smart city management. To address key challenges in dense crowd object detection, including target occlusion, shallow feature degradation, and bounding box overlap, this paper proposes a dense crowd object detection network collaboratively combining Transformer and large -kernel convolutional neural network (LK-CNN). A large-kernel global-aware multi-stage fusion (LK-GAMF) module is designed to com鄄 pensate for the decay of long-range dependencies and suppress shallow positional degradation. In addition, a height-aware repulsion loss (HARL) function is proposed to optimize the bounding box adjustment process and reduce overlapping detection errors. Experimental results on the CrowdHuman dataset show that the pro鄄 posed method achieves an average precision (AP50) of 89.1% and an average log miss rate (MR-2 ) of 47.7%. In ab鄄 lation experiments, compared with the baseline network, the proposed method achieves an AP50 improvement of 0.7%, an AP50-95 improvement of 0.8%, and a recall improvement of 1.0%, confirming the effectiveness of the proposed model. The proposed method enhances the localization capability of occluded targets through largekernel global-aware perception and multi-scale feature optimization. In addition, the HARL function optimizes the position of overlapping boxes in dense rear regions of the image, providing a high-precision solution for dense crowd analysis

Key words: object detection, dense crowd, large-kernel convolution, Transformer