Obstacle Detection And Identification Of Unmanned Driving Based On Lidar

Obstacle detection and identification is an important component of the unmanned vehicle's environment perception function and plays an important role in improving the driving reliability and safety of unmanned vehicles.Due to its high measurement accuracy and wide measurement range,lidar is widely used as the main sensor in the unmanned vehicle.This paper studies obstacle detection and identification based on lidar measurement data in road traffic environment.An obstacle detection method based on point cloud clustering segmentation and an obstacle identification method based on point cloud features and machine learning are established.the main research work and results can be summarized as follows:A set of universal and effective processing methods and procedures are proposed and developed for the raw measurement data of the lidar mounted on unmanned vehicles,which have a certain reference value for data processing and analysis,such as point cloud acquisition,filtering,motion distortion correction,spatial distribution characteristics analysis and so on.In terms of obstacle detection,the definition of obstacles and the main detection objects are clarified,and the detection method based on point cloud clustering segmentation is determined.In order to avoid the interference to the detection,a method based on line fitting is proposed to segment and remove the ground point cloud as the background of the obstacle point clouds,and achieve a good ground segmentation effect.In view of the problems of under-segmentation and over-segmentation that may be caused by the current commonly used clustering segmentation methods,an adjacent growth clustering method based on Euclidean distance is proposed to segment obstacle point clouds.Specifically,the k-d tree data structure of the point cloud is established for fast neighborhood search in the point cloud space and by automatically adapting the neighborhood search radius threshold to changes in distance,the problems of under-segmentation and over-segmentation are avoided to a certain extent.At the same time,the cluster size constraint and reflection intensity constraint are introduced to achieve effective segmentation of neighboring obstacles,which further improves the segmentation accuracy.The principal component analysis(PCA)method is proposed to analyze the obstacle point cloud clusters.Then the oriented bounding box(OBB)models are established to describe spatially the obstacles in the driving scene and extract the basic spatial information of obstacles.In terms of obstacle identification,the classification method based on point cloud features and machine learning was specified.The point cloud database of obstacles is established through manual collection and annotation,and the basic outline characteristics,reflection intensity characteristics,point cloud statistical characteristics,inertia tensor characteristics,and axial projection height characteristics of the obstacle are extracted to establish the obstacle's combined feature vector and feature data set.In order to achieve the goal of obstacle classification,the support vector machine(SVM)is selected as a classifier model and trained using feature data set.The model is evaluated by indicators such as accuracy and precision,which proved the effectiveness of the extracted combined feature vector for classification.The obstacle identification experiments on the point cloud of driving scenarios verify the identification effect of the method based on point cloud features and machine learning.