Research On Traffic Sign Detection And Recognition Based On Improved YOLOv4

Traffic sign detection and recognition technology is an important research direction in the field of Computer Vision.This technology can transmit effective sign information to drivers in time,thereby reducing driving safety problems caused by human factors.At present,the comprehensive performance of the YOLOv4 object direction model is good.However,in the tasks applied to traffic sign detection,there are still problems such as little feature information to be checked,the loss of feature information in the fusion stage,and the volume of the model.In view of the above problems,this paper studies based on YOLOv4,and proposes two improvement strategies,aiming to find a more suitable method for traffic sign detection and recognition.The main research contents of this article are as follows:First,a method for improving YOLOv4 for small traffic sign detection based on an attention feature fusion network is proposed.It mainly operates on small traffic sign feature information.On the one hand,the method uses small object data augmentation techniques to enrich the number of objects to be detected.This method can obtain more feature information of small objects on one image.On the other hand,we built an attention feature fusion network to highlight useful small object information and suppress invalid interference information in the background.The network mainly includes an attention space pyramid pooling network and an attention path aggregation network.The former can capture small object information in the larger receptive field with more context.The latter introduces feature semantic information into the shallow layer,highlights the small object information of traffic signs,and reduces the loss of feature information in the process of feature fusion.Finally,experimental results on TT100 K and GTSDB datasets show that our method improves the detection performance of small traffic signs.Compared with YOLOv4,YOLOv3 and Faster R-CNN,it has increased by 12% mAP,17% mAP and 45% mAP respectively.Second,a lightweight traffic sign detection method based on GhostNet is proposed to improve YOLOv4.It mainly lightweights the backbone network of the original model.On the one hand,building a YOLOv4 model with lightweight GhostNet as the backbone network.GhostNet can greatly reduce the number of model parameters and reduce calculate costs when extracting rich feature information.On the other hand,the method improves the classification loss and confidence loss of YOLOv4.Specifically,the method introduces the Focal Loss function to control positive and negative samples and distinguish the weight of difficult and easy samples.This method can improve the recognition accuracy of traffic signs.Finally,a large number of experiments has been carried out on the CCTSDB and Lisa datasets.The results show that our method reduces the model size by 82.3% compared to YOLOv4,the detection speed is increased by 45%,and the calulation amount has also been greatly reduced.It shows that our method realizes the lightweight of the model.