一种基于YOLOv4的密集人群小目标检测方法  

A small target detection method for dense crowd based on YOLOv4

作　　者：王翀 王同军 周正一 WANG Chong;WANG Tongjun;ZHOU Zhengyi(Equipment Price Evaluation Center,Naval Equipment Department,Beijing 100071,China)

机构地区：[1]海军装备部装备审价中心,北京100071

出　　处：《应用科技》2024年第2期82-89,共8页Applied Science and Technology

摘　　要：针对密集人群中由于视觉受阻和目标被遮挡导致小目标检测精度不佳的问题,本文基于YOLOv4模型将卷积块–像素块注意力机制模块(convolutional-pixel block attention module,CBAM-PIX)融入主干网络CSPDarknet53,并利用级联思想改进特征融合网络。注意力机制方法和特征融合方法不仅提升了数据的丰富性,而且提高了空间通道像素提取信息的能力和目标检测的准确性。此外,通过减少网络层数降低计算量和减少参数,提高了网络模型在有限计算资源和设备需求下的适应能力。实验结果表明,改进的模型算法在用于密集人群小目标检测时精确度提升了1.96%,且鲁棒性强。该算法为解决复杂背景下密集人群小目标检测提供了有效的解决方案,具有应用价值。Aiming at the problem of poor small target detection accuracy due to visual obstruction and target occlusion in dense crowds,this paper incorporates the convolutional-pixel block attention module(CBAM-PIX)into the backbone network CSPDarknet53 based on the YOLOv4 model,and improves the feature fusion network by making use of cascading ideas.The attention mechanism method and feature fusion method not only enhance the data richness,but also improve the ability to extract information from spatial channel pixels and the accuracy of target detection.In addition,lowering the computational amount and parameters by reducing the number of network layers,so as to improve adaptability of the network model under limited computational resources and equipment requirements.Experimental results show that the improved model algorithm has improved accuracy by 1.96%and increased robustness when it is used for small target detection in a dense crowd.The algorithm provides an effective solution to solve the small target detection of dense crowds in complex backgrounds,having application value.

关 键 词：小目标检测 YOLOv4 特征提取 卷积块–像素块注意力机制模块 密集人群 多尺度特征网络 WiderPerson数据集 特征融合 

分 类 号：TN911.73[电子电信—通信与信息系统]