Research On Dynamic Obstacle Detection And Tracking Based On 3D Lidar

With the development of science and technology,unmanned technology and mobile robots are closely integrated to form intelligent robots that can move autonomously.They have been widely used in many fields such as commerce,industry,and social public services,bringing great convenience to people’s life and production.Intelligent robots move autonomously in unknown and complex environments.First,they need to observe the unknown environment through the "eyes" to find a way forward,and always pay attention to the obstacles in front to avoid collisions.Here,the eyes correspond to the environment perception system.Therefore,this ability plays a key role in intelligent mobile robots.This topic uses three-dimensional lidar as the core sensor of intelligent mobile robot environment perception.It mainly studies the registration of three-dimensional laser point clouds,obstacle detection,dynamic obstacle detection and tracking.The specific research work is as follows:Firstly,the ground segmentation method based on RANSAC was studied.Accurate segmentation of ground point clouds is the basis for obstacle detection.Ground segmentation of the original point clouds can remove a large number of useless ground points while retaining effective obstacle information.Aiming at the phenomenon of false detection of multi-plane point clouds by traditional RANSAC ground segmentation algorithm,this thesis uses the height difference in each grid to obtain the pre-screening area of ground points,and performs RANSAC ground point screening in this restricted area to solve the false detection problem of RANSAC algorithm.Finally,the public data set and self-collected data set are used to prove that the proposed algorithm has better segmentation effect in multi-plane environment,and the segmentation efficiency is better than the traditional RANSAC algorithm.Secondly,the point clouds registration method based on ICP was studied.Point clouds registration can effectively obtain the rotation changes of adjacent frame point clouds,and play an important role in the subsequent loss of obstacles and dynamic obstacle recognition.Aiming at the problem that the accuracy of traditional ICP point clouds registration depends on the selection of initial values and the registration efficiency of large-scale point clouds data,an improved ICP algorithm based on feature points is proposed,and a fast NDT algorithm is first introduced to shorten the distance of adjacent frame point clouds.Secondly,in order to reduce the number of point clouds and greatly ensure the basic characteristics of the original point clouds,the feature points extracted by combining curvature and normal phasor angle are used as input point clouds to complete the registration.Finally,the algorithm is verified by using public small and large datasets and self-mining experimental data,and the results show that the improved ICP algorithm is superior to the traditional algorithm in terms of accuracy and computational efficiency.Thirdly,the obstacle detection algorithm based on DBSCAN was studied.The obstacle detection algorithm can cluster and segment the obstacle set into independent obstacle targets.Aiming at the problem that the traditional DBSCAN algorithm is prone to error detection and missed detection due to the density of the obstacle point clouds set along the horizontal direction of the lidar,an improved DBSCAN algorithm based on kernel density is proposed.The kernel density is introduced to estimate the density distribution of the obstacle set,and the point clouds is segmented regularly and clustered in turn,so as to improve the influence of the traditional DBSACN clustering algorithm on the uneven distribution of point clouds density.Finally,experimental verification is carried out on the KITTI dataset and the self-collected dataset.The results show that the positive detection rate is better than the traditional method,and the improved method can effectively segment adjacent obstacles.Finally,dynamic obstacle tracking based on unscented Kalman filter was studied.The problem of tracking failure and trajectory fracture caused by occlusion of dynamic obstacles is analyzed emphatically,and an improved method is proposed.Aiming at the problem of local occlusion of obstacles,the density and Euclidean distance features between targets with good anti-occlusion ability are selected for multi-feature data association to reduce the influence of local occlusion on the accuracy of association.Aiming at the problem of target loss caused by complete occlusion of obstacles,the tracking target pre-screening mechanism is added to the unscented Kalman tracking algorithm,and a method of identifying the lost target based on the change of the number characteristics of the overlapping point clouds of each target after point clouds registration is proposed.The selected lost target is combined with the unscented Kalman filter for continuous state prediction until the target reappears,thereby solving the problem of tracking trajectory fracture caused by occlusion of the target.Finally,the feasibility of the algorithm is verified by self-collected data sets.The experimental results show that the algorithm can achieve accurate and continuous association tracking of pedestrian motion,and has good tracking accuracy for occluded targets.