AN IMPROVED TRAFFIC SIGN DETECTION ALGORITHM BASED ON YOLOV4

It is necessary to detect traffic signs as early as possible and make driving decisions in a timely manner in the automatic driving scene, traffic signs at this point are small targets. The detection accuracy of small traffic signs is low. In order to solve this problem, an improved traffic sign detection algorithm based on YOLOv4 is proposed. The improvement of the algorithm mainly included the following parts: the integrated attention module was embedded into the backbone network to strengthen the attention of channel and spatial information; the binary cross entropy loss function was changed to focal loss to solve the imbalance problem of positive and negative samples; multi-scale information of picture was used for feature extraction and dilated convolution increase receptive field. The proposed methods were trained and tested on TT00K dataset respectively. The experimental result shows that the total mAP of the improved network model is improved by 14.16% compared with the original YOLOV4, and its overall performance outperforms other detection methods.