| With the breakthrough of artificial intelligence technology,the development of transportation networks and the increasing number of cars,the demand for traffic sign detection and recognition in the field of autonomous driving through computer vision algorithms is also increasing.As an important part of the automatic driving assistance system,TSR(Traffic Sign Recognition)is closely connected with the safe travel of autonomous vehicles,and plays an important role in guiding vehicles to drive correctly,regulating traffic,improving road capacity,and reducing traffic accidents.The core of TSR is performance,and the real-time and accuracy are the most important aspects of algorithm research.In the real scene,the images captured by the cameras around self-driving car is high-resolution distant images,and the proportion of the target is extremely small.Considering the high-speed driving of the car,the real-time nature of the system is the key to the system’s application in real scenes.If the system cannot quickly recognize the traffic signs on both sides of the road and feedback to the driver,the system has no application value.Traditional detection algorithms are not robust enough to cope with various complex scenarios.In order to solve these shortcomings,and for the problems of current object detection algorithms with low real-time performance and poor detection of small targets,this paper proposes a series of improvements to address various difficulties and challenges.The specific research is as follows:(1)An improved YOLOv3 traffic sign detection algorithm is proposed.There are two main points in the optimization of this algorithm.One is to optimize the loss function to ensure that the large and small targets have the same impact on the entire loss,so as to solve the problem that the loss of the large target dominates the total loss during training.The second is to streamline the backbone network Darknet53 and improve the detection speed.(2)A SSD traffic sign detection algorithm(B-SSD)based on block feature fusion is proposed.This algorithm is aimed at the difficult of small object detection.First,the shallow feature map is divided into blocks and then fused with the backbone network in the feature capture stage to preserve the spatial information of some small targets.Secondly,in the prediction stage,the high-dimensional feature map is deconvolved and then fused with the lowdimensional feature to obtain a feature map with both spatial and semantic information.This article conducts in-depth research and experiments on the current mainstream singlestage target detection algorithm,and improves the detection effect under the premise of ensuring the detection speed.The experimental results are evaluated on CCTSDB(Changsha university of science and technology Chinese Traffic Sign Detection Benchmark)and GTSDB(German Traffic Sign Detection Benchmark).Compared with the original algorithm,both have been improved.The improved YOLOv3 algorithm has the mAP(Mean Average Precision)of 89.84% on the CCTSDB dataset,which is an increase of 0.55% compared to the original YOLOv3;on the GTSDB dataset,the mAP has reached 84.67%,increased by 1.19%,and the FPS(Frames Per Second)increased from 33f/s to 56f/s.The B-SSD traffic sign detection algorithm has the mAP of 91.07% on the CCTSDB dataset,which is an increase of 4.79% compared to the original SSD;on the GTSDB dataset,the mAP has reached 86.08%,increase by 4.59%.The FPS of B-SSD is 24 f/s.The experimental results show that the two detection algorithms proposed in this paper improve the detection accuracy under the real-time performance,and have practical value in real scenes. |
