面向智能驾驶的密集行人检测算法  

Dense pedestrian detection algorithm for intelligent driving

作　　者：姜鑫 苏昭宇 薛松 Jiang Xin;Su Zhaoyu;Xue Song(School of Information and Communication,Guilin University of Electronic Science and Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学信息与通信学院,桂林541004

出　　处：《现代计算机》2024年第10期1-10,共10页Modern Computer

基　　金：广西自然科学基金2023(GXNSFAA026294)。

摘　　要：针对智能驾驶场景下密集行人检测泛化能力差和检测精度低的问题,提出一种改进的SF-YOLO算法。该方法首先通过引入无参注意力机制——SimAM注意力机制,挖掘更深层次的特征通道间关系和特征图空间信息,增大神经网络模型的感受野,增强模型在特征提取阶段获得更加丰富的特征信息;然后在Neck部分借鉴BiFPN思想进行多尺度特征融合;其次增加信息融合模块,结合头部关键信息提高遮挡行人的置信度,抑制无效特征,降低网格的训练难度,提高对遮挡行人的识别能力;最后在后处理阶段的目标检测框的选定方法上,设计Soft Confluence替代原始算法,优化anchor的回归预测,改善行人密集时因距离过近而被漏检的情况,提高模型收敛能力。实验结果表明,所提的改进YOLO算法在人员密集区域的行人检测精度高达92.1%,相较于原始模型提高了4.9个百分点,平均精度均值提高了6.1个百分点。Aiming at the problems of poor generalization ability and low detection accuracy of dense pedestrian detection in in-telligent driving scenarios,an improved SF-YOLO algorithm is proposed.This method first introduces the parameter-free attention mechanism-SimAM attention mechanism to mine deeper relationships between feature channels and feature map spatial informa-tion,increase the receptive field of the neural network model,and enhance the model to obtain richer features in the feature extrac-tion stage.information;then draw lessons from the BiFPN idea in the Neck part to perform multi-scale feature fusion;secondly,add an information fusion module to combine key head information to improve the confidence of occluded pedestrians,suppress invalid features,reduce the difficulty of grid training,and improve the recognition of occluded pedestrians.ability.Finally,in terms of the selection method of the target detection frame in the post-processing stage,Soft Confluence was designed to replace the original al-gorithm,optimize the regression prediction of anchor,improve the situation of missed detection due to too close distance when pe-destrians are densely packed,and improve the convergence ability of the model.Experimental results show that the improved YOLO algorithm proposed in this article has a pedestrian detection accuracy of up to 92.1%in densely populated areas,which is 4.9 percentage point higher than the original model,and the average accuracy is increased by 6.1 percentage point.

关 键 词：智能驾驶 行人检测 遮挡问题 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术] U463.6[机械工程—车辆工程]