Road scene small target detection method based on modified YOLOv5
Target detection in autonomous driving has received much attention in recent years as one of the basic tasks of computer vision,and target detection and recognition in road scenes is one of the most core modules in autonomous driving.The YOLOv5(YOLO-You Only Look Once)target detection algorithm is commonly used to detect and recognize the targets in road scenes during vehicle driving,and the algorithm usually achieves excellent detection speed but poor detection accuracy.The algorithm usually obtains excellent detection speed,but the detection accuracy is not satisfactory.Based on the YOLOv5 model,the loss function CIoU is changed from CIoU to EIoU,which improves the detection accuracy,and the multi-scale feature fusion method is used to add the ultra-small detection head module,which changes the original network model from 3 to 4 detection heads,and the three most suitable candidate box sizes for the dataset are selected by K-means clustering method.suitable candidate frame sizes for the dataset,solving the problem of inaccurate detection of tiny targets by the original model.Trained on the VisDrone2019 dataset,the mAP-50 of the final model network can reach 42.8％,which is improved by 7.6％compared with the original model,and the experimental results show that the proposed method is more suitable for the task of target detection and recognition in the road scene compared with the original model of YOLOv5.
deep learningtarget detectionYOLOv5road sceneautopilot
NETL
NSTL
万方数据
