Clustering And Tracking Method Research Of Lidar-based Obstacle In Urban Environment

The Autonomous Vehicle is an outdoor mobile wheeled robot with autonomous driving capability.In addition to the driving functions of traditional cars,driverless cars also have functions such as environmental awareness,behavioral decision making,path planning,motion control and active obstacle avoidance.Unmanned vehicles can not only reduce traffic accidents,reduce energy consumption and exhaust emissions,but also have broad application prospects in tasks such as search and rescue in extreme environments such as danger and harshness.As the "eye" of the driverless car,the real-time and stability of the sensing system directly determines the safety and reliability of the unmanned vehicle and the adaptability to the changing traffic environment.At present,there are still some problems in the research method of perceptual layer theory based on lidar:data registration between sensors and calibration methods of sensors and vehicles are unscientific,clustering of point cloud data is prone to over-segmentation,and obstacle matching is time-consuming.Causes problems such as un-stable tracking.This paper takes the unmanned vehicle developed by the Hefei Institute of Material Science of the Chinese Academy of Sciences as the research platform,and makes the following three researches on these three problems:1.Aiming at the problem of inaccurate search for the corresponding point pairs in the coarse registration phase of the dual laser radar point cloud data registration process,this paper uses the sphere as the auxiliary object,and calculates the spherical center as the corresponding point pair for coarse registration,which improves the correspond-ing point pair.Coincidence rate;afterwards,the method of calculating the center of the sphere is also used for calibration.The experimental results show that the method effec-tively improves the accuracy of the initial conversion matrix in the coarse registration stage and has good calibration results.2.Aiming at the problem of over-segmentation of LiDAR point cloud clustering at present stage,a two-stage clustering algorithm is proposed in this paper.In the first stage,the window neighborhood is used to find the adjacent points,and the obstacles are clustered into subgroups based on the European clustering algorithm,and the second stage uses the DBSCAN algorithm of adaptive determination parameters to gather the clusters into the final obstacles.Experiments show that the two-stage clustering algo-rithm has good robustness and effectively avoids the problem of over-segmentation of obstacles.3.Aiming at the problem of unstable obstacle detection,a similarity measure method is proposed to optimize and correct the confirmation matrix associated with joint probability data.By extracting the NDT(Normal Distributions Transform)char-acteristics of the obstacle and other rigid body characteristics,the similarity measure is performed on the target and the measurement to form a similarity matrix.By optimiz-ing and correcting the confirmation matrix,the probability of the combined explosion is reduced,thereby improving The efficiency of target matching is used;then the Kalman filter algorithm is used to predict the motion state of dynamic obstacles.Experiments show that the method effectively reduces the probability of occurrence of combination matrix explosion,improves the efficiency of data association and state prediction,and has better real-time and stability.Finally,the feasibility and effectiveness of the method described in this paper is verified by a large number of real vehicle experiments in urban road environment.