Rapid Segmentation And Classification System For 3D Objects In LiDAR Point Cloud

In outdoor-oriented environmental perception technology,using lidar to collect high-precision 3D point cloud data has become an important means for researchers to obtain spatial information on the surface of environmental objects.In the task of point cloud data processing,fast and efficient point cloud segmentation and classification technology is an important technical support for the environment-aware function of intelligent robots and intelligent driving vehicles.However,the 3D point cloud data collected by Li DAR is complex and large,and the point cloud data are formally distributed offline,which makes it difficult to process 3D point cloud data efficiently and accurately.Because the point cloud data is arranged in a different way than 2D point cloud data,the rules of the pixels in the image are arranged differently,which prevents the convolutional neural network from being used to directly acquire the local between 3D point clouds correlation information.Therefore,in order to improve the segmentation speed and classification accuracy of objects in Li DAR point clouds,this paper uses vehicle-mounted Li DAR to collect large-scale 3D point cloud data and completes the following work:1)Installing 32-wire LIDAR equipment on the roof of the unmanned vehicle,collecting raw 3D point cloud data of the environment around the unmanned vehicle outdoors and completing data refinement operations on it.2)The fast,real-time segmentation of 3D point cloud data is accomplished with the graphical processor-accelerated connected area marker-based algorithm proposed in this paper.The CPU-GPU performance comparison of the proposed algorithm was carried out.The proposed 3D object segmentation algorithm can achieve a segmentation speed of more than 30fps(frames per second).This can be achieved by optimizing the operational efficiency of unmanned vehicles in sensing the surrounding scenes and improving the ability of autonomous decision making of unmanned vehicles.3)A multi-resolution packet-sampled PointNet++ network model structure adapted to sparse point clouds is built in a deep learning-based 3D point cloud object classification algorithm.This model uses the ball query method to query the neighborhood of the sample points,and uses the local features of the neighborhood of the center point as the input of the miniature PointNet network for feature learning;In order to normalize the input point cloud and improve the classification accuracy of the model,this paper also adds an input transformation layer to complete the alignment of the input point cloud.It also incorporates the idea of multi-resolution feature fusion to complete the task of efficient and accurate 3D point cloud classification and identification.4)In a comparative experiment,this paper compares the prediction accuracy of some classical deep learning classification models,with the model presented in this paper having a higher the good classification effect with 91.5% classification accuracy compared to PointNet++ network model.The accuracy is improved by 0.8%.In addition,this paper also adopts the traditional manual feature extraction based method to extract the global features of the segmented point cloud objects,and then uses different trained machine learning classifiers to complete the classification task on the 3D object models,and statistics the classification accuracy of different models under the same test data.