Research On Traffic Sign Detection Based On Improved YOLOv5

Cars gradually become one of the indispensable choices of travel mode,however,a large number of traffic accidents followed.People urgently require a smarter and safer way of driving,for this sake,assisted driving and autonomous vehicles have increasingly become the hottest topics and research hotspots.A full-fledged autonomous driving system will be equipped with multiple sensors.Among these sensors,camera is widely used due to its low price and strong applicability,which has become the standard configuration of the automatic driving system.Tesla’s Pure Vision Autopilot system is a classic example.Camera sensors are mainly responsible for sensing and predicting the complicated situation mixed with pedestrian,traffic signs,vehicles and lane lines,etc.Among above detection tasks,in the process of traffic-sign-detection task,researchers meet many problems such as small targets,weather effect ion,surface contamination,and various types of traffic signs.Due to its powerful fitting and generalization ability,deep learning has pushed the development of image tasks to a new degree.And traditional image detection methods have gradually replaced by deep learning which has occupied a dominant position in the autonomous driving solutions.Furthermore,convolutional neural network is the main topic in this paper.The significated contributions of this paper are as follow ing:（1）This paper proposes a basic component with stronger feature-extraction capability,CACSP,which embeds a coordinate attention in the residual block.CACSP,utilizing long-distance information dependencies,can alleviate the degradation of target features caused by convolution stacking and pooling.It is worth mentioning that this component is vital to keep small target features.In addition,CACSP is also very helpful for small target-based traffic sign detection tasks.Compared with the YOLOv5s baseline,the mAP_0.5:0.95^val indicators on the COCO2017 and TT100k validation se ts are improved by 1.1%and 2.7%,respectively.（2）Combining the advantages of SPP and Corner Pool,this paper proposes the Corner SPP module.This module can extract rich semantic features at multiple scales,and also has the ability to fuse long-distance important features.Moreover,the Corner SPP module can make up for the lack of convolution to capture long-distance features.At the same time,the influence of noise introduced in the process of expanding the receptive field can be alleviated to a certain extent with the help of this module.Compared with the YOLOv5s baseline,the mAP_0.5:0.95^val indicators on the COCO2017 and TT100k validation sets are improved by 1.2%and 0.5%,respectively.（3）In view of the disadvantages of the original shared detection head of YOLOv5,this paper proposes a decoupled detection head with stronger representation ability.The decoupled detection head can effectively solve the problem of contradictory characteristics of different tasks.It can improve the per formance of a multi-class tasks such as traffic sign detection.Compared with the YOLOv5s baseline,the mAP_0.5:0.95^val indicators on the COCO2017 and TT100k validation sets are improved by 1.6%and1.9%,respectively.（4）This paper proposes a set of data augmentation strategies that are more suitable for traffic scenarios,which can improve the performance of the model under the condition with limited amount of data and without increasing the time-consuming of inference.Hence,the model will be provided with stronger generalization ability with the support of these strategies.Compared with the YOLOv5s baseline,the mAP_0.5:0.95^val indicators on the TT100k validation set increased by 0.5%,and the mAP_0.5:0.95^val indicators increased by 0.8%with this data enhancement strategy in this paper.In conclusion,the scheme in this paper is based on the solution proposed by the existing excellent model YOLOv5s,and the mAP_0.5:0.95^val indicators of the validation set on the public data sets COCO2017 and TT100k are respectively 3%and 4%higher than that of the YOLOv5s baseline.