基于改进YOLOv3-tiny的道路车辆检测算法  

Research on Improved YOLOv3-tiny Road Vehicle Detection Algorithm

作　　者：邢镇委 伋淼 张梦龙 XING Zhenwei;JI Miao;ZHANG Menglong(Anhui University of Science and Technology, Huainan 232001, China)

机构地区：[1]安徽理工大学电气与信息工程学院,安徽淮南232001

出　　处：《洛阳理工学院学报（自然科学版）》2021年第4期58-63,共6页Journal of Luoyang Institute of Science and Technology：Natural Science Edition

基　　金：安徽理工大学研究生创新基金项目(2019CX2049).

摘　　要：道路车辆实时检测是智能驾驶技术中的一项关键技术。针对现有车辆检测算法准确性和检测速率不能同时兼容的问题,提出一种改进YOLOv3-tiny的道路车辆检测算法。将空间金字塔池化(SPP)整合到网络的特征提取部分,通过不同规模的池化层对图像中车辆特征进行融合,增强网络性能。实验表明:改进后的算法检测速率能够稳定在29.7帧/s,mAP相比于原始的YOLOv3-tiny算法提高6.32%,在复杂道路环境下的车辆检测也能取得很好的效果。Real-time detection of road vehicles is a key technology in intelligent driving technology.Aiming to solve the problem that the detection algorithm accuracy is not compatible with its efficiency,an improved YOLOV3-tiny detection algorithm is proposed in this paper.Spatial pyramid pooling(SPP)has different size pooling layers,and SPP can be used as YOLOV3-tiny network feature extraction to enhance network performance.The results show that the detection rate of the improved algorithm is 29.7 f/s,and mAP is increased by 6.32%.Vehicle detection even in the complex road environment can also achieve good effects.

关 键 词：车辆检测 改进YOLOv3-tiny SPP k-means++ 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]