Study On Traffic Environment Perception System For Autonomous Driving

The State Council issued the "Development Planning for a New Generation of Artificial Intelligence" on July 8,2017,which repeatedly mentioned the importance of developing Autonomous Driving technology,and called for making breakthroughs in construction of Intelligent Transportation System.The perception of the traffic environment is extremely important in the Autonomous Driving.There are incomparable advantages that using visual images to detect and recognize the static objects on the roads.A variety of objects information on the roads can be obtained simultaneously by using a single visual image.The traffic environment perception algorithm based on visual images studied in this thesis combines Deep Learning algorithms with traditional object detection algorithms for three objects:roadblocks,traffic signs and lane lines.The main work of this thesis is as follows:Deep Learning has made remarkable achievements in the field of object detection in recent years.It can be roughly divided into two frameworks,Two-stage networks based on Regional Proposal and One-stage networks based on Regression.This thesis integrates two mainstream frameworks,then designs a neural network model for object detection based on actual road traffic images to detect the roadblocks and traffic signs on the roads.In the field of object classification,Deep Neural Networks have strong learning ability for the features of images and high accuracy for object classification.However,the Deep Learning algorithms have a strong dependence on the data size,which has poor performance in the face of lack of data,and yet the traditional object classification algorithms can overcome this disadvantage.Due to the difference between roadblocks and traffic signs,the Deep Learning algorithms and the traditional object classification algorithms are applied separately to complete the classification of 4 kinds of roadblocks and 43 kinds of traffic signs in this thesis.Due to the particularity of lane lines,this thesis studies and compares kinds of state-of-the-art Semantic Segmentation algorithms in the field of Deep Learning,and optimizes speed performance,creating a solution combining the Deep Learning algorithms and traditional algorithms,so as to detect the lane lines on the actual traffic roads.In order to verify the performance of the traffic environment perception algorithm proposed in this thesis on the actual traffic roads in China,we established a domestic road traffic dataset consisting of three subsets,including 7419 pieces of artificially-marked pictures containing roadblocks,11449 artificially-marked pictures containing traffic signs and 37501 pieces of artificially-marked pictures containing lane lines.In summary,this thesis based on the visual perception technology,conducts research on static object perception algorithms for traffic roads,and optimizes the performance of traffic environment perception system by combining Deep Neural Networks with traditional object detection algorithms,making certain contribution to the development of Autonomous Driving technology.