| With the rapid development of the transportation industry,the convenience of people’s travel has improved a lot.In the meantime,many problems and hidden troubles have emerged,for example,several traffic jams,frequent traffic accidents,and so on.As an alleviating solution,in recent years,the traffic signal detection and recognition technology in intelligent driving has entered a practical stage.Two detection and recognition algorithms are in common use for traffic signal: one is based on radar sensors and the other is based on vehicle vision.However,due to the complex requirements and high cost of radar equipment,the traffic signal detection and recognition technology based on vehicle vision has received more and more attention from researchers.In the actual driving scene,the traffic signal captured by the vehicle camera accounts for a small proportion and the background of the location is complex,so the detection and recognition algorithm needs to spend a lot of computing power to improve the accuracy,which brings some difficulties to the real-time detection of Advanced Driving Assistance System.From the perspective of detection accuracy and recognition accuracy,this paper improves the existing algorithm model to meet the need of raising the practicability of detection and recognition algorithm,and studies the whole design scheme of traffic signal detection and recognition based on vehicle vision.The main contents can be stated as follows:In order to improve the detection performance of the algorithm based on YOLOv5l on traffic signal data sets,an optimization algorithm based on YOLOv5l detection is proposed.K-Means++ algorithm is used to calculate the anchor box,and the cluster centers far away from each other are selected adaptively to optimize the clustering results.Then the Bidirectional Feature PyramidNetwork is introduced into the feature fusion network,and the spanning connection is added to enrich the feature information on the basis of the bidirectional feature fusion.Finally,the Generalized Intersection Over Union is replaced by Distance Intersection Over Union in the prediction network to improve the accuracy of the prediction frame.In order to solve the problem that CNN relies too much on convolution operation,a traffic sign recognition optimization algorithm based on Vision Transformer is proposed.The ViT model is used as the traffic sign recognizer.In order to reduce the risk of information loss in the process of image slicing,the DenseNet network is introduced before the ViT encoder to extract features,which realizes the dense connection between the original features and the convoluted features,and convoluted features and strengthens the transitivity of the features.Finally,importing the encoded features into the ViT encoder,and making the attention mechanism focus on the features,and then the recognition results are output.This paper’s overall design performance can be validated experimentally.The results show that the occurrence of false detection and missed detection can be avoided effectively by the use of YOLOv5l detection optimization algorithm,and the detection accuracy is improved to a certain extent,and the Mean Average Precision value is increased by 5.1%.The ViT recognition optimization algorithm can recognize the traffic signs more accurately,and the recognition accuracy is improved by 3.57%.Finally,in a variety of scenarios,the detection and recognition ability of the overall design scheme of this paper on traffic signals is verified,which proves the effectiveness of the overall design scheme of this paper. |
