基于YOLOv4的复杂交通状况下多目标检测算法  

Multi object detection algorithm under complex traffic conditions based on YOLOv4

作　　者：吕辉[1] 董帆 Lv Hui;Dong Fan(College of Electrical Engineering,Henan Polytechnic University,Jiaozuo 454000,China)

机构地区：[1]河南理工大学电气学院,焦作454000

出　　处：《国外电子测量技术》2022年第12期41-47,共7页Foreign Electronic Measurement Technology

基　　金：河南省科技攻关项目(222102210225);河南省高等学校重点科研项目(22B510007);光电传感与智能测控河南省工程实验室项目(HELPSIMC-2020-007)资助。

摘　　要：当前复杂交通状况下多目标检测存在检测精度低,检测速度慢,模型参数量大等问题。针对上述问题基于YOLOv4提出一种能够快速检测并识别多个不同目标的密集神经网络。首先将高效通道注意力机制(ECA)与跨阶段密集连接网络(CSPDenseNet)结合,组成新的E-CSP主干网络,代替传统的残差网络(ResNet)。新的主干网络加强了有效通道的特征表达,提高了特征提取层提取特征的能力;其次使用改进的空间金字塔池化与柔性非极大值抑制(Soft-NMS),加强对于小目标与被遮挡目标的检测能力。实验结果表明,方法的平均类别精度(mAP)、帧率达到0.92%、50 fps,明显高于其他方法。通过与目前主流模型比较,方法在获得较高识别精度的同时,具有参数规模小识别速度快的特点,可以极大的提高交通行驶的安全性。At present,there are many problems in multi-target detection under complex traffic conditions,such as low detection accuracy,slow detection speed and large model parameters.To solve the above problems,a dense neural network is proposed,which can quickly detect and recognize multiple different targets.First,the efficient channel attention mechanism is combined with CSPDenseNet to form a new backbone network to replace the traditional ResNet network.The new E-CSP strengthens the feature representation of effective channels and improves the ability of feature extraction layer to extract features.Secondly,the improved spatial pyramid pooling and Soft-NMS methods are used to enhance the detection ability for small targets and occluded targets.The experimental results show that the mAP and frame rate of this method reach 0.92%and 50 fps,which are significantly higher than those of other methods.Compared with the current mainstream models,the method in this paper has the characteristics of small parameter scale and fast recognition speed while obtaining high recognition accuracy,which can greatly improve the safety of traffic driving.

关 键 词：深度学习 目标检测 密集连接 注意力机制 交通安全 空间金字塔池化 

分 类 号：TP274[自动化与计算机技术—检测技术与自动化装置]