基于YOLOv3的车辆检测算法

Vehicle Detection Algorithm Based on YOLOv3

针对城市道路场景下车辆检测精度低的问题,提出了一种基于YOLOv3的车辆检测方法。借鉴了可变形卷积网络(Deformable Convolutional Networks,DCN)的思想,对YOLOv3的Darknet-53主干网络结构进行优化,在残差模块中加入了可变形卷积增加网络的特征提取能力,提高了模型的检测精度。实验分析表明,改进的YOLOv3模型在KITTI数据集的车辆类别上mAP达到了92.33%,对比YOLOv3模型精度提高了1.74%。

Aimed at the problem of low vehicle detection accuracy in urban traffic scene,a vehicle detection method based on YOLOv3 was proposed.In the research of object,detection model,the darknet-53 backbone network structure of YOLOv3 was optimized,and the idea of Deformable Convolutional Networks(DCN)was used for reference.The Deformable Convolutional Networks(DCN)were added into the residual module to increase the feature extraction capability of the network,which improved the detection accuracy of the model.Through the experimental comparison,the improved YOLOv3 model reached 92.33%mAP in the vehicle category of KITTI dataset.,the accuracy 1.74%better than the original YOLOv3 model.