Research On 3D LIDAR-based Environment Perception For ALV

Autonomous Land Vehicle is a kind of mobile robot, which can drive both in urban and countryside environment. It has great research value and broad application prospect both in national defense and civilian traffic field. This thesis mainly aimed at the autonomous driving in both structure and non-structure environment, it carried out the research on three-dimension LIDAR-based environment perception for Autonomous Land Vehicle theoretically and practically.The main results and innovations of this thesis are as follows:Firstly, because fitting the line in every sector cannot describe the ground in countryside, and the computation efficiency of the two-dimension Gaussian Process regression in the elevation map is very low, a novel ground segmentation algorithm was proposed, which use the one-dimension Gaussian Process regression and Incremental Sample Consensus for every sector in the polar grid map. This method can segment all kinds of grounds in daily life and reduce the cost of computation to meet the real-time demand of Autonomous Land Vehicle.Secondly, considering upon the data distribution of three-dimension LIDAR, an improved Toe-Finding algorithm for road intersection detection was presented; this method first analyzes the admissible space in front of the vehicle, then uses the beam model of the laser range finder and Toe-Finding algorithm with constraint of vehicle width to detect the road intersection, it can recognize all kinds of intersections twenty meters away from the Autonomous Land Vehicle. Experiment results show the reliability of the algorithm. Because of the poor SNR of the curb, a new method based on the max-min elevation map was proposed, it uses the Hough Transform algorithm to extract the candidate lines, and to select the best result with the constraint of the vehicle-curb relationship. This method can extract the curb in urban environment accurately.Thirdly, an improved vehicle tracking method which is based on the vehicle measurement model and particle filter was proposed. This method directly uses the box model to represent the results computed by differencing two consecutive frames, and eliminates the need for data association. According to the model, a particle filter based vehicle tracking algorithm was implemented with great efficiency.We have used these methods in our Autonomous Land Vehicle which won the champion in the third Chinese Future Challenge. The robot ran ten kilometers autonomously in urban environment of the city Ordos in twenty-one minutes.