基于改进YOLO v3模型与Deep-SORT算法的道路车辆检测方法  

Road vehicle detection method based on improved YOLO v3 model and deep-SORT algorithm

作　　者：马永杰[1] 马芸婷 程时升 马义德[2] MA Yong-jie;MA Yun-ting;CHENG Shi-sheng;MA Yi-de(College of Physics and Electronic Engineering,Northwest Normal University,Lanzhou 730070,Gansu,China;School of Information Science and Engineering,Lanzhou University,Lanzhou 730030,Gansu,China)

机构地区：[1]西北师范大学物理与电子工程学院,甘肃兰州730070 [2]兰州大学信息科学与工程学院,甘肃兰州730030

出　　处：《交通运输工程学报》2021年第2期222-231,共10页Journal of Traffic and Transportation Engineering

基　　金：国家自然科学基金项目(62066041)。

摘　　要：针对道路车辆实时检测遮挡严重与小目标车辆漏检率高的问题,提出了基于改进YOLO v3模型和Deep-SORT算法的车辆检测方法;为提高模型对道路车辆的检测能力,采用K-meansSymbolk@pSymbolk@p聚类算法对目标候选框进行聚类分析,选择合适的Anchor box数量,并在网络浅层增加了特征提取层,可提取到更精细的车辆特征;为加强网络对远近不同目标的鲁棒性,在保留原YOLO v3模型输出层的同时,增加了一层输出层,将52像素×52像素输出特征图经过上采样后得到104像素×104像素特征图,并将其与浅层同尺寸特征图进行拼接,实现车辆目标的检测;为了降低目标遮挡对检测效果的影响,提高对视频上下帧之间关联信息的关注度,将改进YOLO v3模型和Deep-SORT算法相结合,以此来弥补两者之间的不足。试验结果表明:改进YOLO v3模型有效地提高了车辆检测的性能,与在网络浅层增加特征提取层的模型相比,平均精度提高了1.4%,与增加一层输出层的模型相比,平均精确度提高了0.8%,说明改进YOLO v3模型提取的特征表达能力更强,增强了网络对小目标的检测能力;改进YOLO v3模型在引入Deep-SORT算法后,查准率和召回率分别达到90.16%和91.34%,相比改进YOLO v3模型,查准率和召回率分别提高了1.48%和4.20%,同时保证了检测速度,对于不同大小目标的检测具有良好的鲁棒性。A vehicle detection method based on the improved YOLO v3 model and deep-SORT algorithm was proposed to address the problems of serious occlusion and high misdetection rate of small target vehicles in the real-time detection of road vehicles. To improve the detection ability of the model for road vehicle, the K-meansSymbolk@pSymbolk@p clustering algorithm was used to cluster the target candidate boxes, the appropriate number of anchor boxes was selected, and a feature extraction layer to the shallow layer of the network was added to extract more refined vehicle features. The robustness of the network for different distant targets was enhanced by retaining the original YOLO v3 model’s output layer but adding another layer to it. After the 52 pixel×52 pixel output feature map was upsampled, a 104 pixel×104 pixel feature map was obtained, which was spliced with a shallow layer feature map of the same size to achieve the vehicle target detection. To reduce the influence of target occlusion on the detection and improve the attention to the association information between the upper and lower frames of the video, the YOLO v3 model was improved and combined with the deep-SORT algorithm to compensate for their shortcomings. Experimental results show that the improved YOLO v3 model can enhance the vehicle detection performance. Compared with the model adding feature extraction layer in the shallow layer of the network, the average accuracy improves by 1.4%, and compared with the model adding one output layer, the average accuracy improves by 0.8%. It indicates that the improved YOLO v3 model has a stronger feature expression ability and enhances the network’s ability to detect small targets. After the deep-SORT algorithm is introduced into the improved YOLO v3 model, the precision and recall are 90.16% and 91.34%, respectively. Compared with the improved YOLO v3 model, the precision and recall increase by 1.48% and 4.20%, respectively. At the same time, the detection speed is maintained, and the detection of dif

关 键 词：交通图像识别 卷积神经网络 车辆检测 YOLO v3模型 Deep-SORT算法 K-means++聚类算法 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]