Research On Traffic Sign Detection And Recognition Algorithm In Complex Environment Based On Deep Learning

Traffic signs are crucial for maintaining road safety and efficiency.According to road status,density,and visibility,setting corresponding traffic signs along the road can provide drivers or intelligent vehicles with information and restrictions on road conditions to achieve efficient traffic regulation.Environmental perception is a powerful source of information for intelligent transportation systems(ITS)and intelligent vehicle vision systems.Traffic signs are an intuitive reflection of traffic information,and their rapid and accurate detection is the key to achieving safe and stable road operation.At present,domestic and foreign research has made some progress in traffic sign detection and recognition,but there are many challenges to achieving accurate and realtime detection of traffic signs due to their rich types,fast update,and small size,especially their complex application environment,which is easily affected by uncontrollable factors such as light,weather,occlusion,and complex condition.In view of the above problems,the existing domestic and international research is summarized and analyzed,and the deep learning-based object detection algorithm and adaptive image enhancement algorithm are combined to design a traffic sign detection and recognition network for complex environments.The main research and contributions of this thesis are as follows:The adaptive image enhancement network(AIENet)for complex environments is designed for the diverse application environments of traffic sign detection and recognition tasks.Based on an adaptive joint filtering strategy,AIENet achieves real-time optimization of images through six image filters under six more common working conditions: rain,haze,exposure,dimness,motion blur,and fading.In the testing phase,we arrange and combine images from six environments to produce 720 synthetic videos with 30 frames,reflecting the dynamic changes of complex scenes.The average confidence in the detection of videos before and after using AIENet is improved by 9%,and the average confidence of the network is higher than 90%,with strong robustness;A new traffic sign detector is proposed to face the challenge that existing detectors have difficulty in the trade-off between speed and accuracy.The proposed detector optimizes the feature extraction and information retention capability of the network to achieve accurate localization and recognition of small-sized targets while controlling the model size and computational cost.The proposed detector is evaluated on the generic datasets and the traffic sign datasets to examine the model performance in various aspects.And the testing accuracies reach 62.0%,65.9%,and 79.6%on three traffic sign datasets,TT100 K,STSD,and DFG,respectively,where the testing accuracy of small-size targets is 46.2%,59.2%,and 32.4%,respectively,which is better than most existing traffic sign detection networks;This paper achieves real-time traffic sign detection on mobile,which is of great significance for the development of ITS and autonomous driving.In this paper,the proposed detector is deployed on the Jetson Xavier NX mobile platform and the TensorRT hardware acceleration is applied to transform the model and generate engine files,which significantly improves the detection speed of the detector on the mobile platform.We installed the model deployed Jetson Xavier NX platform to the in-vehicle side for road tests,achieving traffic sign detection at about 26 frames per second,providing the first traffic sign information to intelligent vehicles or drivers,and improving vehicle driving safety.