Ground Traffic Sign Recognition Method And Application

With the accelerated application of foreword technologies such as artificial intelligence,big data and 5G,the development of autonomous driving is of great significance to ensure road safety and improve traffic efficiency.Autonomous vehicles must first have the ability to detect and recognize the external environment,and then integrate the information into the decisionmaking system to control driving actions.The automatic driving of vehicles relies on the rich road information brought by ground traffic signs,but there are many challenges in completing ground traffic sign recognition under complex background conditions in natural scenes.The use of traditional algorithms is limited by changes in the environment and target diversity,and the current vehicle-mounted computing equipment also restricts the detection speed of the algorithm.In order to solve the problem of multi-class recognition of ground traffic signs in natural scenes,this paper uses a lightweight DAB-Net semantic segmentation network to recognize ground traffic signs.First,DAB-Net combines attention feature association and multilayers information fusion to improve network performance.The Squeeze and Excitation channel attention mechanism,Criss Cross horizontal and vertical spatial attention,Object Contextual Representations context information fusion and Mutilayers Information Fusion have been added to DAB-Net successively,so that DAB-Net has three dimensions from pixel space,channel,layer and layer to re-integrate the correlation between features.In the case of adding a small number of parameters,the segmentation accuracy of the network is effectively improved.In order to enable the segmentation model to effectively identify the ground,the improved DABNet network with the highest segmentation accuracy was selected to automatically label Baidu ground traffic sign data and increase the ground area category.The basic DAB-Net network is used to train the ground traffic sign data containing the ground area categories,and then the segmentation result is subjected to image follow-up processing to complete 14 types of main ground traffic sign recognition.In order to improve the detection speed,this paper applies 8-bit quantization to the model and designs a multi-threaded parallel pipeline structure to optimize the detection speed.Finally,the complete algorithm is deployed in Nvidia Jetson TX2 and Xavier embedded systems.This paper has conducted experiments on the Cityscapes dataset,Baidu ground traffic sign dataset,Nvidia Jetson TX2 and Xavier platforms.Experiments show that the improved DABNet segmentation network has an accuracy of 73.15% m Io U on the Cityscapes dataset,which is3% higher than the original network,and the model parameter is only 1.26 M.On the Baidu ground traffic sign dataset,the segmentation accuracy of 14 main ground traffic signs is 75.3%m Io U.After model quantization and multi-thread deployment optimization,when the input image is 1280×360,the detection speed of the PC(1080Ti)is increased from 42 FPS to 133 FPS.The detection speeds achieved in Nvidia Jetson TX2 and Xavier embedded systems are 12 FPS and 50 FPS respectively,and have real-time detection capabilities on low-power devices.This article has carried out a complete scheme design and realization,which has good reference value for further application.