Abstract:

In view of the problems of low accuracy and high requirements for equipment performance of existing target detection methods in traffic scenes， a target detection method based on improved YOLOv3 network is proposed.A new IOU（Intersection over Union） function is proposed to improve the CIOU（Complete Intersection over Union）function， as CIOU function can not accurately measure the degree of coincidence between the prediction box and the real box under special circumstances. By increasing the detection scale， the feature mAP is refined and the deep and shallow semantic information fusion is enhanced to enhance the small target detection ability； by adding SE（Squeeze-and-Excitation） attention mechanism， the model pays more attention to learning the feature information of important channels； according to the characteristics of the data set， M-YOLT enhancement method was used to improve the performance of small target detection. Through the structural channel pruning of the network， the equipment performance requirements are reduced， and Matrix NMS is used to replace NMS to improve the detection speed. The experimental results show that the proposed method can improve the detection effect and reduce the detection time. The test on KITTI data set reaches 88.4% mAP， or the average detection time of each image can be reduced to 8.6 ms while maintaining 86.3% mAP. This method can achieve better detection effect with lower hardware requirements.

Key words: intelligent transportation, target detection, deep learning, convolutional neural networks（CNN）, YOLO