| With the continuous advancement of technology,people’s lives have also begun to enter the era of intelligence.In terms of transportation,the most closely related to us is Driver Assistance and Autonomous Driving technology.These technologies can not only improve our travel efficiency,but also ensure our travel safety.Therefore,the research on intelligent transportation technology has very important significance and application value.In the road traffic system,traffic signs are an important part.Whether it is Driver Assistance or Autonomous Driving,the problem of traffic sign detection must be solved.However,in the process of practical application,because the traffic signs mainly exist outdoors,the environment is more complicated,and the quality of the images collected by the camera is quite different due to the influence of factors such as light,rain and fog.In addition,in the process of image acquisition,due to the shooting angle,shooting distance,and focus problems,the acquired images may be blocked,distorted,blurred,etc.All of these have brought huge challenges to the detection of traffic signs.Based on these problems encountered in traffic sign detection in practical applications,this article aims to solve the problem of traffic sign detection in complex environments.The main work content is as follows:(1)The traffic sign data preprocessing module TSP is proposed.Through the analysis of the data preprocessing methods of the existing object detection algorithms,it is found that they have problems in the task of traffic sign detection.One is that it is easy to cause small-sized traffic signs to be indistinguishable after the image scaling process,and two It is that the direction of some directional traffic signs changes when the image is flipped,and the third is that the annotation information is not handled well in the process of cropping the image.Therefore,in response to these problems,a TSP preprocessing module is designed to preprocess the input image without reducing the image quality and the meaning of traffic signs.And through the comparative analysis of experiments,the advantages of this module in the detection effect of traffic signs are illustrated.(2)Proposed SA-YOLOv4-tiny network based on attention mechanism.The attention mechanism is to imitate the way humans observe things.Introducing the attention mechanism into the deep learning network can help the model quickly focus on important parts and ignore redundant background information.In this paper,by adding the SA attention module to the YOLOv4-tiny network,under the condition of only a small amount of calculation,the detection ability of the model is effectively improved.(3)Design an improved YOLOv4-tiny network.Because the traffic sign itself is small and mostly captured by long-distance shooting,it often occupies a small size in the image.Therefore,the design of traffic sign network should focus on the detection of small targets.The YOLOv4-tiny network is aimed at general target detection,and it is not designed to focus on the detection of small targets.Therefore,this article analyzes the difference in information description of feature maps of different depths,and combines the idea of feature fusion to improve the network structure of YOLOv4-tiny,thereby improving the effect of the network on traffic sign detection.(4)The TS-YOLO model based on traffic sign detection is proposed.In this paper,the YOLOv4-tiny model is improved in three different aspects,and the networks before and after the improvement are trained and tested.The comparative analysis shows the effectiveness of the improved method proposed in this paper in traffic sign detection.Finally,the three methods proposed in this paper are merged,and compared this model with YOLOv4-tiny,YOLOv3,Faster R-CNN object detection algorithms on the TT100K traffic sign dataset.Experiments show that the algorithm in this paper is significantly better than the other three object detection algorithms in the task of traffic sign detection... |
