Research On Road Target Detection Method Based On YOLOv3

Road target detection is a key technology in many research fields such as intelligent assisted driving and intelligent transportation.With the development of deep learning,more and more researchers use target detection methods based on deep learning to carry out road target detection.As one of the target detection algorithms with the best comprehensive performance at present,YOLOv3 can meet the real-time requirements of road target detection and achieve a better detection effect.However,YOLOv3 still has the problems of low detection rate of small targets and the inaccurate location of BBox.In order to improve the detection accuracy and maintain the real-time performance,this paper optimizes and improves YOLOv3.Firstly,in view of the problem that YOLOv3 is not effective in detecting small targets such as pedestrians,the original three feature scales of YOLOv3 are increased to four,thus reducing the rate of missed detection of small targets such as pedestrians.Secondly,aiming at the problem that the accuracy of YOLOv3 bounding boxes needs to be improved,the K-Means++ target box clustering is used to get a new candidate box for road target detection,so as to improve the detection accuracy.Finally,by merging the BN(Batch Normalization)layer into the convolutional layer,the antecedent reasoning speed of the network is improved to speed up the detection.Based on the above improvements,this paper proposes an improved road target detection network YOLO-RT(YOLO for Road Target)based on YOLOv3.In this paper,YOLO-RT is tested on the self-collected domestic road data sets,and compared with YOLOv3.The experiment shows that in the scene of road target detection,the m AP of YOLO-RT reaches 96.9%,and the AP of pedestrians reaches91.2%,which is 6.1% and 16.5% higher than that of YOLOv3 respectively,and the detection ability of small targets has been significantly improved.Under the GTX1080 Ti graphics card,the FPS is 50.9 frames/second,which is basically the same as that of YOLOv3,and can also meet the real-time requirements.