Research On Recognition And Ranging Of Traffic Objects In Front Of Intelligent Vehicles

Since its invention in 1886,the vehicle has always been an irreplaceable means of transportation in human society.In the current era,it is also moving from simple transportation to more diversified functions.With the development of automated driving technology research,the improvement of its environmental perception function is increasingly valued by the industry.Usually,this function of an intelligent vehicle consists of radar,camera,GPS positioning,etc.From the perspective of reliability and economy,relying on the camera to obtain target information is more advantageous than other methods.The obtained traffic scene information can be used as input for planning decision-making and motion control modules to jointly accomplish the ultimate goal of automated driving.In important scenes such as intersections,the identification and judgment of the type of traffic lights and the position of the stop line by the vehicle is directly related to the safety and legality of its driving.Based on the driverless taxi project,this thesis studies the detection and ranging of traffic light stop lines.The first work of this thesis is proposed due to the unsatisfactory traffic light datasets that exist today.The quality of the data set will directly determine the quality of the network training results.Through the method of video collection and manual annotation,we constructed a traffic light-stop line dataset suitable for this task,and its working conditions include a variety of road and weather environments.In addition,data enhancements such as Gaussian noise are introduced in the dataset to enhance its generalization ability.This dataset is finely divided into 16 types of traffic lights,including stop lines and crosswalks,with a total of about 50,000 images and 80,000 labels.Compared with the existing traffic scene datasets,the dataset constructed in this paper has certain advantages.The detection algorithm based on this is also more suitable for joint debugging with other modules of automated driving.The second work is based on the Yolov3 convolutional neural network,which optimizes the network structure and other aspects.Through the statistics and classification of the k-means algorithm on the size of the bounding box of the target object in the above data set,combined with the pyramid output structure in the original Yolov3 network,a detection output layer with a smaller target was added,and the original large size was deleted.The target output layer,the structure of Yolov3-4layers and Yolov3-new3 is proposed.At the same time,the update optimization is also made from the perspective of the intersection ratio and activation function in the loss function.Under the same experimental conditions,the new network can improve the detection accuracy by 5.2% and has more advantages in memory usage.On this basis,the third work of this thesis is to deduce the ranging algorithm of the target object by studying the imaging principle of the monocular camera.Because the spatial positions of various targets are different,different ranging schemes are proposed,and their accuracy and feasibility are compared through experiments.Next,import the selected algorithm into the target detection model,and carry out the road experiment of target detection and ranging algorithm based on the hardware equipment required for the actual vehicle platform assembly experiment of Dongfeng E70.It makes up for the deficiency that related problems are mostly used for simulation and lack of real vehicle experiments.Finally,by analyzing the preliminary ranging results,the error amount is fitted with a regression equation to obtain the correction amount.The correction equation is fed back to the ranging algorithm to improve the accuracy of ranging,and the ranging error of the stop line can be controlled within 10%.