Research On Traffic Light And Digital Detection And Recognition Based On Embedded GPU

Object detection is a challenging computer vision problem,with a wide range of application prospects.Vision-based perception of the road environment is an important means of obtaining information for assisted driving systems,and of the numerous road information,the traffic information provided by traffic lights is particularly important.It is the baton for vehicle traffic and the key to smooth roads.Traffic lights contain important and accurate information and have a very high detection value.This article will study and implement a traffic light and digital detection system based on embedded GPU.In view of the characteristics of traffic light detection tasks,that is,high accuracy and fast detection speed,this article selects the YOLOv3 model in the deep learning regression method for in-depth research.First,make improvements to YOLOv3's feature extraction network.Design a down-sampling block including 3 × 3 and 5 × 5 convolution kernels and maximum pooling layer,which is used to replace the original convolutional layer with reduced feature spectrum size of the network to reduce the loss of features in the feature extraction network.Secondly,on the basis of the YOLOv3 detection network,up-sampling with a step size of 4 is added to merge the depth features with different shallow features to make the feature spectrum output features more comprehensive and improve the depth features in the multi-scale detection network.Use and improve the detection rate.Next,this article introduces the lightweight network MobileNet to reduce the amount of parameters and calculation of the improved YOLOv3 model,thereby increasing the speed of inference.Finally,the number and size of anchor frames are selected by K-means clustering,thereby improving the detection accuracy and recall rate.In this paper,a data set of 15 classification traffic lights including digital lights,turn signals and traffic lights are produced.The design model is tested on it,and the results show that the model has good detection accuracy.In addition,the NVIDIA Jetson Tegra X2 embedded platform used in this article comes with GPU and TensorRT inference acceleration library to meet the model's demand for computing power.TensorRT makes GPU utilization higher by reducing computational accuracy and streamlining the network structure.After accelerating with TensorRT,on the premise of reducing the accuracy,it greatly improves the inference speed of the model on the embedded platform.In this paper,multi-threaded detection is adopted,so that videoreading and model inference do not interfere with each other,so as to achieve the purpose of real-time detection.The traffic lights and digital detection system designed in this paper are based on advanced assisted driving system.The system transmits the detected traffic light information to the main control unit of the auxiliary driving system to assist in decisionmaking.Tests under homologous and non-homologous scenes show that the system has good detection effect on traffic light targets in the middle and close range,and the average frame rate of detection is above 20 FPS,which can adapt to the detection task under lowspeed driving.