Traffic Sign Detection And Recognition Under Highway

In recent years,with the rapid development of technologies such as 5G,big data and the Internet of Things,intelligent transportation and driverless systems have become research hotspots for major automobile companies and Internet enterprises,and traffic sign detection is its core technology,while the detection and recognition of traffic signs under highway conditions is one of the difficult points of research.although the YOLOv5 s target detection model has good results in recognition accuracy,its backbone The feature extraction network and the enhanced feature extraction network have a large number of convolutional operation layers,resulting in large training model weight files,slow detection speed,and the inability to port embedded devices.At the same time,in the environment of expressway,due to high real-time requirements,it is necessary to detect and recognize traffic signs in the scene of small targets,but the general target detection model often has problems such as false detection,missing detection and insufficient detection accuracy of small targets.To address the above two issues,this paper first investigates network model lightweighting and proposes a lightweight traffic sign detection and recognition model by optimizing and improving YOLOv5 s to make the model smaller in size and faster in detection without reducing accuracy.Then,from the perspective of reducing the rate of missed and false detection of small-target traffic signs and increasing their detection accuracy,a more targeted improvement strategy is designed for YOLOv5 s to improve the applicability and robustness of its model when performing traffic sign detection and recognition in small-target scenarios.The main work is as follows:(1)The ban signs with excessive data volume in the CCTSDB and TT100 K datasets were randomly deleted,and Mosaic data enhancement was performed on the warning sign and indication sign datasets.After designing the Robust CP data enhancement strategy,the data volume of all datasets was expanded and balanced again to solve the problem of uneven distribution of various types of traffic signs,and finally a homemade traffic sign dataset with large data volume,balanced categories and highway as the background was established.(2)A lightweight L＿YOLOv5s(Lightweight YOLOv5s)model is proposed to address the problems of large structure of YOLOv5 s training model,slow detection speed and ineffective portability to embedded devices.The YOLOv5 s backbone feature extraction network is replaced with the Inverted resblock inverted residual structure,and the SE attention mechanism is introduced.Based on Ghost module,the general convolution and CSP2＿X components in enhanced feature extraction network are improved,and finally the lightweight model L＿YOLOv5s is obtained.The experimental results show that L＿YOLOv5s is better than the YOLOv5 s model in terms of detection accuracy and recognition speed because the number of parameters and computation decreased by nearly 70%,the training time decreased by 50%,and the goal of model lightweighting was achieved,meanwhile the m AP0.5 increased by 3.45% and the FPS increased by 10 frames per second.(3)For the problems of false detection,missed detection and insufficient detection accuracy of small targets of highway traffic signs,an improvement of YOLOv5 s model for small target detection based on Bi FPN of bi-directional feature pyramid network is presented.The YOLOv5 s enhanced feature extraction network is replaced with Bi FPN,and in addition,the Sim AM attention module is used,and the Poly-1 Loss function framework is put forward to replace the original cross-entropy loss and focus loss,and finally the improved small target detection model based on Bi FPN-YOLOv5 s is constructed.The experimental results show that the training visualization results of the improved Bi FPN-YOLOv5 s model are smoothly converged,and its m AP0.5 is as high as 91.84%,which is much better than all other comparison models in terms of detection accuracy,moreover the FPS reaches 30 frames without affecting the detection speed.The visualization experiments further show that the proposed improved Bi FPN-YOLOv5 s model can effectively solve the problem of missed and false detection of small target traffic signs detection and recognition in highway scenarios.