Research On Key Technologies Of Obstacle Detection For Intelligent Vehicle Based On Lidar

Intelligent vehicles are important for improving personal mobility,promoting the upgrading of the automobile industry,and reducing traffic accidents.In order to realize the unmanned driving,Firstly,Intelligent vehicles must be able to accurately detect the surrounding obstacles and be able to predict the trends of dynamic obstacles..Lidar has higher precision than visual sensors and is less affected by light.Therefore,Lidar is indispensable in the detection of obstacles.Obstacles are mainly divided into dynamic and static obstacles.To detect moving and static obstacles,the lidar points must be classified as ground points and obstacle points.To detect dynamic obstacles,the position and posture of the vehicle must be known.Based on the above reasons,the thesis mainly studied from the following three aspects:Aiming at the problem of ground segmentation of 3D points cloud.Because of the invalidation of traditional height threshold algorithms when encountering the ramps and bumpy roads.Thesis applied Markov Random Field Model in image segmentation to point cloud segmentation.Gradient grid map was used to replace the image.Firstly,thesis established a fan-shaped grid map,calculated the gradient of each grid and built a Markov random field based on the gradient grid map.This thesis overcomed the influence of irregular roads such as ramps on point cloud segmentation by Markov's characteristics.Finally the algorithm in thesis achiecved point cloud segmentation under different road conditions.Aiming at the problem of obtaining a precise judgment basis for determining dynamic and static obstacles at low cost.In this thesis,the Generalized-ICP algorithm and the extended Kalman filter combined with gyro and acceleration information were used to estimate the position and posture changes of the vehicle in a short time.Firstly,the Generalized-ICP algorithm was used for point cloud registration,and the registration results were used to calculate vehicle pose changes.The pose changes were processed to update the extended Kalman filter.The gyroscope and accelerometer are used as the input of the extended Kalman filter.The result was the posterior estimate of the kalman filter.Finally this thesis completed accurate estimation of vehicle position changes.Aiming at the problem of detecting and tracking dynamic obstacles,an improved DBSCAN algorithm was used to cluster the obstacle data points after separating the ground points,and the contour and angle features of the obstacles were extracted by the method of least convex hull and fuzzy line segment.The vehicle pose changes was used to distinguish dynamic and static obstacles.Then thesis used a multi-target hypothesis method to perform data associations for all two consecutive dynamic obstacles,and used a Kalman filter for velocity estimation.Finally,thesis completed the detection and tracking of dynamic obstacles.Finally,algorithm in this thesis were tested to verify the feasibility and stability by the lidar,accelerometer,and gyroscope data collected by our Intelligent vehicles experimental platform and the actual vehicle data of the kitti data set.