基于改进的YOLOv3多目标小尺度车辆检测算法研究  

MULTI-TARGET SMALL-SCALE VEHICLE DETECTION ALGORITHM BASED ON IMPROVED YOLOV3

作　　者：田智慧[1,2] 杨奇文 魏海涛[2] Tian Zhihui;Yang Qiwen;Wei Haitao(School of Information Engineering,Zhengzhou University,Zhengzhou 450001,Henan,China;School of Earth Science and Technology,Zhengzhou University,Zhengzhou 450052,Henan,China)

机构地区：[1]郑州大学信息工程学院,河南郑州450001 [2]郑州大学地球科学与技术学院,河南郑州450052

出　　处：《计算机应用与软件》2023年第12期169-175,共7页Computer Applications and Software

基　　金：国家重点研发计划项目(2018YFB0505004-03)。

摘　　要：针对传统车辆检测算法效率低、漏检率高、对小目标车辆检测效果不好等问题,提出一种改进的YOLOv3车辆检测算法。使用K-means++对训练标签进行聚类,确定车辆检测的anchor box;将特征提取能力更强的EfficientNet作为特征网络,并采用4个特征尺度融合深层的语义信息和浅层的位置信息,提升小尺度车辆的检测效率;引入CIoU和Focal loss函数,提高了网络收敛速度和检测精度。实验结果表明,在UA-DETRAC和自建的数据集上,所提算法的MAP、Recall和FPS分别达到90.9%、88.3%和30帧每秒,提升了小目标车辆的检测精度。Aimed at the problems of low efficiency of traditional vehicle detection algorithms,high missed detection rate,and poor detection of small target vehicles,an improved YOLOv3 vehicle detection algorithm is proposed.K-means++was used to cluster the training tags to determine the Anchor box for vehicle detection.EfficientNet was used with stronger feature extraction capabilities as the feature network,and 4 feature scales were used to fuse deep semantic information and shallow position information thus improving the detection efficiency of small-scale vehicles.CIoU and Focal loss functions were introduced to improve the network convergence speed and detection accuracy.Experimental results show that on the UA-DETRAC and self-built data sets,the MAP,Recall and FPS of the proposed algorithm reach 90.9%,88.3% and 30 frames per second respectively,which improves the detection accuracy of small target vehicle.

关 键 词：车辆检测 YOLOv3 深度学习 EfficientNet 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]