Research On Target Perception And Traking Control Of Driverless Formula Car Based On Vision

With the continuous progress of the society and the rapid development of science and technology,the research related to intelligent vehicles is closely concerned by scientific research institutions,enterprises and universities.It involves the cross-integration of multi-disciplinary knowledge,and all parties are increasing the investment in research and development.As a high-level goal of the development of intelligent vehicle,unmanned vehicle has made a breakthrough at present.Based on the ROS platform,this paper takes the driverless formula cars in low-speed environment as the research object,and conducts in-depth research on the environment perception and tracking control technology based on the actual track scene.In terms of environmental perception,32-line lidar and camera are used to perceive traffic cones on both sides of the track.First of all,the vision-based target detection algorithm is used to detect the traffic cone target.The purpose is to ensure the correctness of the recognition category and the accuracy of the positioning frame under the highest detection frame rate,contrast analysis of the target detection algorithm based on artificial design features and target detection algorithm based on the depth of the convolution network,respectively for Faster R-CNN algorithm based on region and SSD based on regression algorithm was improved,so as to determine the optimal detection scheme can meet the test requirements on the existing platform.Then,on the basis of the completion of the lidar and camera calibration,a data fusion method based on the detection of traffic cone targets in the image frame is proposed.The point cloud data provided by the lidar is fused with the image pixel information to determine the position of the cone-bucket targets in the three-dimensional space and provide a basis for the subsequent planning path.Combined inertial navigation can realize high precision positioning on the scale of centimeters,and transform the longitude and latitude information into the global coordinate system of twodimensional vehicles to obtain the pose information of unmanned vehicles.In the aspect of tracking control,the "presight-following" theory widely used in the field of path tracking control technology is selected to carry out research.In the preset expectations path reference point under the premise of vehicles-path error model is established to describe the relative position relations,design algorithm according to the reference point and distance path to calculate the appropriate path tracking target,in the aspect of control strategy using feedback structure,in view of the traditional pure tracking algorithm fixed way to improve at distance,make its preliminary aim distance can adaptive adjustment.The control speed is filtered to ensure the steady change of speed and Angle,and improve the accuracy and robustness of path tracking.Finally,modularized independent experiment and simulation analysis are carried out respectively.In terms of cone barrels target detection,the improved SSD model has better detection effect than the improved Faster R-CNN model,with a frame rate of about 25 FPS,which is more sensitive to the small tar get traffic cones under the simulated track scene.After the sensor data calibration and fusion,the position of the cone barrrels in the three-dimensional space can be located.The improved pure pursuit control algorithm reduces the lateral deviation by 58% on average in the middle and low expected control speeds,improving the effect on path tracking control.It provides some reference for the research of driverless formula racing cars.