The Recognition Of Traffic Signs Based On Point Cloud And Images

Traffic sign is a kind of important transportation infrastructure,which contains many important traffic information,such as speed limit,driving behavior restriction,front road condition change and so on.In recent years,in order to provide location information and control decision information accurately,not only the accurate position information,but also the semantic information its expression were needed in the research of intelligent transportation system and unmanned vehicle.Therefore,it is of great practical significance to study the fine extraction of traffic signs and its semantics.On the basis of summing up the previous research results.In this paper,we propose a method of fusing traffic signs with point cloud and image.First of all,using the current best method based on convolutional neural network in the image of traffic sign detection and location preliminary,and through the rich color and shape information of the image,extracting semantic information of traffic signs.The image and laser point cloud are registered to guide the traffic sign detection and precise positioning in the laser point cloud.Then,using a combination of the strength of the information and the characteristics of the dimension of the method,through the identification of the structural characteristics of the traffic signs and the high reverse characteristics of the exact detection and positioning.The advantages of this method lie in:the combination of images and the advantages of laser point cloud data,at the same time for road traffic signs accurately expressed by the three dimensional space information and its semantic information,to make up for a single data source cannot meet the fine defect extraction and semantic.The main work of this paper includes the following aspects:1.Research on the convolutional neural network applied in traffic sign detection and recognition of image based on a fast convolution neural network to complete the detection of traffic signs,and constructs a fast convolution neural network for two stage structure for traffic sign recognition.The training and testing are completed by using the data set of German traffic sign detection standard and the German traffic identification standard data set.2.This paper presents a method of traffic sign cloud extraction based on the combination of intensity filtering and dimension feature.Completed the registration of the point cloud and the image,and the use of the detected traffic signs in the image mapped to the three-dimensional space to guide the point cloud extraction.Firstly,the rough detection is performed based on the intensity filter,and then the dimension feature of the space point group is calculated.According to the difference of the dimensions of the targets on both sides of the road in the actual scene.Finally,based on the combination of the rough detection method based on the intensity filter and the point cloud segmentation based on the dimension feature,the final result of the traffic sign point cloud extraction is obtained.