基于改进YOLOv3模型的道路车辆多目标检测方法

Multi-object detection method for vehicles based on improved YOLOv3 model

针对YOLOv3模型对真实道路环境下近、远端目标车辆检测率低、鲁棒性差的问题,给出了一种基于改进YOLOv3模型的车辆多目标检测模型——YOLOv3-Y模型。模型基于Darknet-53特征提取网络,将网络输出的8倍降采样特征图与4倍降采样特征图进行拼接,建立104×104尺度的检测层;在包含4个类别的车辆数据集中,利用K均值(K-means)聚类算法选取出12个先验框并分别匹配到4个不同尺度的检测层中;同时引入了GIOU损失函数对交并比(intersection-over-union,IOU)损失函数进行优化。采用实验室实际道路车辆数据集,将YOLOv3-Y与YOLOv2、YOLOv2-voc、YOLOv2-tiny、YOLOv3及YOLOv3-tiny模型进行对比,结果表明:YOLOv3-Y模型的平均精度均值与召回率明显优于上述算法,提升最小值分别为11.05%和5.20%。

To solve the problems of low detection rate and poor robustness of near and far object on real road environment,YOLOv3-Y based on the Darknet-53 feature extraction network model was proposed.The 8×down-sampling feature map and the 4×down-sampling feature map output by the network were spliced to create a new detection layer of 104×104 scale.The K-means algorithm was used to cluster the vehicle dataset including four categories,and 12 anchors were selected and matched them to four detection layers with different scales respectively.Meanwhile,GIOU loss function was introduced to optimize the intersection-over-union(IOU)loss function.The actual vehicle dataset in the laboratory was used to compare YOLOv3-Y with YOLOv2,YOLOv2-voc,YOLOv2-tiny,YOLOv3 and YOLOv3-tiny models.The results show that the average precision and recall rate of the YOLOv3-Y model are significantly better than the above algorithms,and the minimum improvement values are 11.05%and 5.20%,respectively.