Research On Vehicle Taillight Detection Method Based On Deep Learning

In recent years,with the development of science and technology and the automobile manufacturing industry,people's living standards have continued to improve,and the number of private cars has increased,bringing convenience to mankind,but also bringing about many social problems such as road congestion and traffic accidents.In this environment,intelligent transportation systems are widely studied,and unmanned driving technology is one of the hotspots of research.However,there are relatively few studies on the perception of vehicle driving intentions in the road environment.Taillight detection and recognition are undoubtedly an important branch of unmanned driving technology,which has high research value and practical application significance.Based on the analysis of the current research status of taillight detection and detection difficulties,this thesis proposes a vehicle taillight detection method based on deep learning,aiming to establish a high-precision,real-time taillight detection model.The experiment is based on the BDD100 K data set,establishes the tail light label and makes a data set for comparison and verification of the methods in this thesis.The main work includes the following parts:(1)Aiming at the difficulty of extracting small target features from taillights,this thesis uses Faster RCNN as the basic target detection framework for research.First,use the h-swish function to optimize the Excitation operation of the SENet network to reduce the computational cost of the model and accelerate the convergence speed;then the improved SENet is applied to the VGG16 feature extraction network to focus on channel features with large amounts of information and suppress important channel features;finally,the network structure of Faster RCNN is improved in combination with PANet's network ideas to strengthen the fusion between feature maps of different sizes,and a taillight detection model with multi-scale feature fusion is established.The comparative experiment results show that the improved network improves the detection performance of the model,and the m AP of the detection result reaches84.36%.(2)In order to further improve the accuracy and speed of taillight detection,this thesis selects the typical YOLOv3 in the One-Stage target detection algorithm as the basic target detection framework for research.First,the k-means++ clustering function is used to obtain the anchor,so that the size of the anchor is more suitable for the tail light target;then the Darknet53 feature extraction network is optimized to change the number of repetitions of the residual unit in different layers,and the optimized Darknet46 feature extraction network reduces the number of layers and computational complexity improve the performance of the model,especially the ability to extract features of small taillight targets.In order to strengthen the YOLOv3 network's full use of shallow feature maps and the effective fusion of deep and shallow feature maps,this thesis designs a YOLOv3 taillight detection model based on the weighted bidirectional feature pyramid Bi FPN.On the one hand,a layer of small-scale feature fusion prediction is added on the basis of the original three-scale feature fusion prediction;on the other hand,a Bi FPN weighting method based on extreme values is designed to enable low-level features to maintain higher weights.The comparative experiment results show that the m AP of the new and improved model reaches 91.45% and the FPS reaches 17.6,which greatly improves the accuracy of taillight detection and basically meets the real-time requirements.