Research On Detection And Recognition Of Taillights Of Vehicles Ahead Based On Deep Learnin

In recent years,due to traffic jams and frequent vehicle collision accidents on urban roads,intelligent driving system has become a research focus in academia and industry.The task of identifying the taillight signals of the vehicle ahead can not only assist the driver in keeping a safe distance from the vehicle ahead,but also provide basic data for the intelligent driving system,which plays an important role in path planning.However,there are not many researches on taillight detection and signal recognition of front vehicle.Therefore,this paper studies the algorithm of taillight detection and signal recognition of front vehicle on road driving.The main work is as follows:(1)As taillights are small and dense,a taillight detection algorithm fused semantic information is proposed.Firstly,the CSPRes Ne Xt-40 network is designed to expand the number of convolutions on large-scale feature maps and reduce the number of convolutions on small-scale feature maps,which enhances the feature extraction ability of small targets without increasing the amount of calculation.Secondly,the feature fusion module is optimized,and multi-scale fusion is adopted,so that deep and shallow feature information can be transmitted and fused more fully.Finally,the algorithm combines the high-level semantic Mask branch module to extract the semantic information of the deep feature,which enhances the algorithm’s ability to recognize the semantics of the taillights.Experimental results show that the algorithm we proposed has better detection performance.Our algorithm achieves90.55% m AP on the vehicle taillight test set,which is 14.42% m AP higher than the YOLOv4 algorithm,and speed reaches 34 frames per second.(2)On the basis of vehicle taillight detection,a vehicle signal recognition algorithm based on attention mechanism is proposed.Firstly,the feature extraction process of the DLA network is improved,and a new feature depth aggregation module is designed on this basis,so that deep and shallow features can be better combined.Secondly,a global-local network structure is used to collect texture feature matrices from the shallow layers and retain the semantic information from the deep layers.Finally,the local attention area is mapped on the fused feature layer and the double-line attention pooling module is added.By effectively using multiple attention maps,the efficiency of model training is further improved,and the detection accuracy of the model is improved without decreasing the detection speed.The experimental results show that the accuracy rate is 91.91% on the vehicle signal test set,and the accuracy rate is 89.20% on the Rear Signal data set,indicating that the algorithm has better detection performance and generalization.(3)A software system is built to detect the taillights and recognize the signals of the vehicle ahead.The system integrates a variety of deep learning target detection algorithms,which can realize the complete process from the taillight detection to the signal recognition of the vehicle ahead in the image.In view of the domestic application environment,the image data of the urban roads around the school is collected by the driverless vehicle platform and annotated.The training is based on transfer learning,and the k-means plus plus clustering algorithm is used to generate new anchors to match with the taillight targets.In the real data,the m AP of taillight detection reaches 93.86%,the m AP of vehicle signal recognition reaches87.15%,and speed reaches 32 frames per second,which proves that the algorithm proposed has better detection performance.