Dense crowd detection algorithm based on improved YOLOv5

WU Wen; DAI Wantao; CHEN Wenzhuo

doi:10.19956/j.cnki.ncist.2024.06.011

Abstract：While emergences occur,such as earthquake,fire,terrorist attacks and so on,it is necessary to assist in evacuation and supervision in real time through target detection algorithms.However,the precision of algo-rithm is not high enough due to the complex crowd postures which lead to occlusion and dense distribution.Therefore,this paper proposes an improved dense crowd target detection algorithm.Based on YOLOv5,the backbone network and Neck are improved respectively.CBAM(Convolutional Block Attention Module)at-tention mechanism and Bidirectional Feature Pyramid Network(BiFPN)are introduced to improve the algo-rithm's ability to extract feature information and improve the accuracy of target detection.Finally,compared with the mainstream algorithm,it is proved that the improved algorithm has only 7.12 M parameters and calcu-lation speed of 16.1GFLOPs.While comparing with the other mainstream algorithms,our proposed model has the highest mAP@50 of 41.7%.And the ablation experiments showed the precision and mAP@50 for all classes have increased by 17.4%and 2.4%,respectively.At the same time,the precision for the class of crowds has increased up to 41.2%.These provided theoretical basis and data support for emergency rescue and management.