基于轻量级YOLO的行人和车辆检测

Pedestrian and Vehicle Detection Based on Lightweight YOLO

针对道路上的行人和车辆的目标检测问题,提出了一种改进型YOLOv4的轻量级检测网络YOLOSpring,通过在YOLOv4骨干网络中加入深度可分离卷积结构来简化网络模型,并减少网络模型中的残差块的数量,通过使用K-means聚类方法对车辆和行人锚点框进行重新聚类,改进后的网络模型具有网络参数少,检测精度高的优点。算法在实时交通场景下进行模型训练,检测目标包括行人、小汽车、大巴、摩托车和自行车。实验数据抽取了PASCAL VOC 2007数据集中的Person、Car、Bus、Motorcycle、Bicycle共5个种类的图片样本进行训练和测试,实验结果显示较原算法在检测精度上提高了5%。

For pedestrians and vehicles on the road of the target detection problem,this paper proposes a modified YOLOv4 lightweight detection network,YOLO-Spring,through add depth in backbone networks YOLOv4 separable convolution model to simplify the network structure,and reduce the amount of residual block in the network model,by using the K-means clustering method of vehicles and pedestrians to clustering anchor box,the improved network model has the advantages of low network parameters,the advantages of high detection precision.the algorithm carries out model training under real-time traffic scenes,and the detection targets include:pedestrians,cars,buses,motorcycles and bicycles.Experimental data were extracted from 5 photo samples of person,Car,Bus,Motorcycle and bicycle in the PASCAL VOC data set in 2007 for training and testing.The experimental results showed that the detection accuracy was improved by 5%points compared with the original algorithm.