Research On Traffic Sign Detection Technology For Unmanned Driving Vehicles

In recent years,unmanned driving has received more and more global attention,and safety performance is the first criterion for unmanned vehicles.Unmanned vehicles use cameras to accurately detect road traffic signs used to convey instructions,prohibitions and warnings in real time,which can guide vehicles to abide by traffic rules and avoid traffic accidents.The farther the unmanned vehicle recognizes the traffic sign,the more it can reduce the response time of the vehicle and avoid the collision of the vehicle.Usually long-distance traffic signs appear as small targets on the image,and their small size and unobvious features make it difficult to detect.How to detect traffic signs in real time and with high accuracy is a difficult point,and the existing algorithms have low accuracy and real-time detection of traffic signs.In response to these problems,this paper proposes a series of improvements based on YOLOv3 algorithm to achieve accurate detection of traffic signs in real-time.The research topic of this article comes from the Guangxi innovation-driven development major special project "Research on Intelligent Driving Technology of Electric Sightseeing Vehicles".The main research contents of this paper are as follows:(1)Aiming at the problem of low accuracy of long-distance traffic sign detection under monocular camera conditions,a traffic sign detection algorithm of YOLOv3 based on deconvolution feature fusion(DFF-YOLOv3)is proposed.Deep features with rich semantic information and shallow features with high resolution are fused to form a deconvolution feature fusion(DFF)module and then input to the detection layer for detection.The fusion method with the highest accuracy is selected through experiments and the predicted feature map after fusion enhances the ability to express small targets.The breadth and depth of the unmanned vehicle perspective system model and formula derivation infer the distance of the unmanned vehicle to the farthest traffic sign.DFF-YOLOv3 achieved an accuracy of 86.8% and an FPS(Frame Per Second)of 41.6 on the TT100 K traffic sign dataset,which improved the accuracy of traffic sign detection.(2)Aiming at the problem of low real-time traffic sign detection,a traffic sign detection algorithm based on Inception structure of YOLOv3(Inc-YOLOv3)is proposed.The network model of the YOLOv3 is pruned to reduce network parameters and network complexity.Further an Inception-redefined module structure is designed.The receptive field information of different scales in the image is obtained through convolution kernels of different scales.Then the information obtained at different scales is merged.Finally,the Inception-redefined module is connected to the shallow network of YOLOv3 to enhance the ability of the shallow network to predict traffic signs.The channel ratio and distribution method with the highest accuracy of Inception-redefined module are selected through experiments.Experiments on the TT100 K data set show that the detection accuracy and FPS of the Inc-YOLOv3 algorithm have been improved to 85.0% and 49.6 respectively,achieving a balance between detection accuracy and real-time performance.(3)In order to improve the accuracy of coordinate regression when training the network model,GIOU is introduced as the regression loss function.GIOU can not only reflect the various overlaps between the predicted bounding box and the real bounding box,but also reflect the distance between the two boxes through difference values.In order to make the anchor box more compatible with the target,the K-means clustering algorithm is improved by calculating the aspect ratio of the object coordinates to eliminate the invalid data in the dataset,which not only eliminates the influence of invalid annotation data on the cluster center,but also improves the matching degree between the anchor box and the traffic sign.DFF-YOLOv3,Inc-YOLOv3 and YOLOv3 are tested comparatively on the TT100 K data set.The experimental results show that the detection accuracy is improved after the improvement of GIOU and the clustering algorithm.Specifically,the accuracy and FPS of DFF-YOLOv3 are 92.0% and 41.6 respectively,and the accuracy and FPS of Inc-YOLOv3 are 90.2% and 49.6 respectively.