Research On 3D Point Cloud Recognition Based On Deep Learning

Point cloud data contains rich semantic information,with the characteristics of high density,high precision,it has become one of the main data of 3D target recognition research,and realizes the understanding and analysis of complex 3D scenes.At the same time,because 3D target recognition has important research value,it has quickly become a research hotspot in the field of computer vision,and is widely used in aerospace navigation,remote sensing mapping,military investigation and intelligent transportation systems.It is of great practical significance to propose an efficient and intelligent three-dimensional target recognition method.Therefore,this paper proposes a point cloud deep learning model named MSS-PointNet,uses deep learning technology to identify 3D point cloud data,and further expands the application field of point cloud data.Since PointNet only performs feature learning on a single point and does not lack local information of points,this paper improves the PointNet by the neighborhood sampling method,using the farthest point sampling algorithm to select the sample points and the radius-based ball query method to sample neighboring points for these sample points.Feature learning is performed by using the neighborhood of the point instead of a single point as the input to the network.Then based on the idea of scale space,the radius of the ball query is taken as the scale parameter.Different scales are obtained by changing the scale parameter.The neighborhood sampling is used to construct the point cloud scale space based on different scales,and then the point cloud scale space is selected.The feature learning is performed on multiple scales,and these features are combined into multi-scale features which are input into the classifier to realize the classification and segmentation of the point cloud.The performance of the model is also tested by object classification,object part segmentation and indoor scene segmentation.Higher classification accuracy and segmentation IoU results show that MSS-PointNet models are used for both small object classification and large scene recognition.A better recognition effect is achieved,which proves the effectiveness of the neighborhood sampling method and multi-scale structure.In addition,this paper also extends the application of the improved model MSS-PointNet,applies it to the recognition research of outdoor complex scenes.This paper uses the Semantic3 D dataset as a training set to train MSS-PointNet,then uses the 3D terrestrial laser scanner for field measurement and the supporting software ScanMaster for data pre-processing to obtain the scene point cloud data of the laboratory building of the School of Geography and Tourism before the renovation.Finally,this paper uses the obtained scene point cloud data to test the trained model.The overall classification accuracy is up to 0.972,and the average segmentation IoU is 0.537.The recognition result is also visualized.The experimental results show that the MSS-PointNet model is well suited for the recognition of outdoor complex scenes.