基于改进的YOLOv3的实时车辆检测  

Real Time Vehicle Detection Based on Improved YOLOv3

作　　者：刘婷娜 郑敏华[2] 朱仲杰[2] LIU Tingna;ZHENG Minhua;ZHU Zhongjie(Zhejiang Wanli University,Ningbo Zhejiang 315100)

机构地区：[1]浙江万里学院信息与智能工程学院,浙江宁波315100 [2]浙江万里学院,浙江宁波315100

出　　处：《浙江万里学院学报》2022年第1期81-87,共7页Journal of Zhejiang Wanli University

基　　金：国家自然科学基金(61671412);浙江省自然科学基金(LY19F010002,LY21F010014);宁波市自然科学基金(202003N4323);宁波市领军和拔尖人才培养工程择优资助科研项目(NBLJ201801006);浙江省教育厅一般科研项目(Y201941122)。

摘　　要：车辆检测是智能交通系统的关键技术之一,对实时性和准确性有较高的要求。对此,文章提出了一种基于You Only Look Once(YOLO)v3改进的车辆检测算法,该算法能够确保实时检测的前提下,大幅度提高检测准确率。首先,改进了YOLOv3的特征提取网络,使用跨阶段残差模块替换原有残差模块。该结构的特征重用特性可以有效提高提取特征的效率;其次,设计了一种新的特征融合网络,通过融合不同深度网络层的特征信息,进一步提高了算法的检测准确性。实验结果表明,与原YOLOv3相比,该算法既满足检测实时性,平均精确率(mean Average Precision,mAP)又提高了8.7%。Vehicle detection is one of the key technologies of intelligent transportation system,which requires high real-time and accuracy.Therefore,this paper proposes an improved vehicle detection algorithm based on You Only Look Once(YOLO)v3,which can ensure the real-time detection and greatly improve the detection accuracy.Firstly,the feature extraction network of YOLOv3 is improved,and the cross-stage residual module is used to replace the original residual module.The feature reuse feature of this structure can effectively improve the efficiency of feature extraction;Then,a new feature fusion network is designed.By fusing the feature information of different depth network layers,the detection accuracy of the algorithm is further improved.Experimental results show that,compared with the original YOLOv3,the proposed algorithm can meet the real-time requirements and improve the mean average precision(mAP)by 8.7%.

关 键 词：YOLOv3 目标检测 FPN 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]