Research On Point Cloud Classification Technology Of Obstacle For Unmanned Vehicle Based On Lidar

The identification of type labels for unknown point cloud data is an important part of unmanned vehicle environment perception at present.Point cloud classification refers to the classification and recognition of multi-type objects based on the environmental point cloud data collected by lidar.In this paper,the obstacle point cloud data collected by the laser radar of unmanned vehicle is taken as the research object,and two methods of traditional machine learning,classifier and deep learning,are studied to complete the obstacle point cloud classification task.Aiming at the disadvantage of low accuracy caused by the low utilization of local features in the classification process of Point Net network in deep learning,an OOB-Point Net with high accuracy and robustness is proposed by introducing the idea of out-of-bag error of random forest Out of Bag Point Net(OOB-Point Net)network.The point cloud sample data sets of four kinds of obstacles,such as cars,pedestrians,trees and bicycles,are constructed.On the basis of KITTI original data set collected by lidar,the outlier removal operation of original point cloud data is carried out by using Statistics Outlier Removal(SOR)algorithm,so as to reduce the complexity of noise points for subsequent feature extraction and obstacle segmentation;The point cloud data of four kinds of obstacles,such as car,pedestrian,tree and bicycle,belong to non-ground point cloud.The existence of ground point cloud will lead to the difficulty of obstacle classification or the wrong classification.RANSAC(Random Sample Consensus)algorithm is used to segment the non-ground points such as obstacles.In order to further segment the obstacle point cloud from the non-ground point cloud,DBSCAN(Density-Based Spatial Clustering of Applications with Noise)algorithm is used to segment the object.The global features of the segmented object point cloud cluster are extracted and the Naive Bayes(NB)classifier algorithm is used to filter out the obstacle point cloud cluster and construct the obstacle point cloud sample data set.The experiments verify that the classifier in traditional machine learning is suitable for multiclass obstacle point cloud classification.in that traditional machine learn,the classifier need to extract the feature attributes of the obstacles manually when performing the multi-class obstacle point cloud classification task.in this pap,the global features containing the whole contour information of the obstacles and the local features and their mathematical statistics such as surface normal,surface curvature and fast point feature histogram(FPFH)in the neighborhood are extracted;By comparing and analyzing the effect of SVM(Support Vector Machine),Decision Tree(DT)and Random Forest(RF)classifiers in completing the obstacle point cloud classification task,the experiment shows that the RF classifier model performs better in the overall performance of the four categories of obstacle point cloud classification,the SVM model performs better,and the DT effect is the worst.The classification accuracy of the RF classifier in the two categories of obstacles such as trees and bicycles is much higher than that of SVM and DT.However,it is slightly lower than SVM in vehicle and pedestrian classification,and DT performs generally in four categories of obstacle classification.Finally,the Point Net network classification task for the weak ability to use the shortcomings of local features to improve,OOB-Point Net using the original model input channel scalability attributes,the introduction of random forest out-of-bag error ideas,the characteristics of the obstacles to calculate the importance of local features and obstacles in the top three features of the three-dimensional coordinate information to form a new input channel data,build a new 6 * 6transformation matrix to deal with the new input channel data,So as to strengthen the influence of local features in the classification task,and build Point Net,Point Net++ and Point MLP network for comparative experiments,the experiments show that the improved OOB-Point Net model is better than Point Net network in overall accuracy,average accuracy,robustness and classification of obstacles in their respective categories,and performs better in tree and bicycle obstacle classification,and is better than Point Net++ and Point MLP network in classification speed.