Point Cloud Data Classification Method Based on Deep Learning

Abstract

The invention discloses a point cloud data classification method based on deep learning. The method proposes a multi-scale point cloud classification network. First, based on the completeness, adaptability, overlap and multi-scale characteristic requirements of local area division, A multi-scale local area division algorithm is proposed, and the point cloud and different levels of features are used as input to obtain multi-scale local areas. Then a multi-scale point cloud classification network including single-scale feature extraction, low-level feature aggregation and multi-scale feature fusion modules is constructed. The network fully simulates the working principle of the convolutional neural network, and has the basic characteristics that as the network scale and depth increase, the local receptive field becomes larger and the feature abstraction degree becomes higher and higher. The present invention achieves classification accuracy rates of 94.71% and 91.73% respectively on the standard public data sets ModelNet10 and ModelNet40, which is at a leading or comparable level in similar work, and verifies the feasibility and effectiveness of this method .

Description

technical field [0001] The present invention relates to the technical fields of computer graphics, computer vision and intelligent recognition, in particular to a method for classifying point cloud data based on deep learning. Background technique [0002] With the wide application of 3D sensors such as lidar and RGBD cameras in the field of robots and unmanned driving, the acquisition of point cloud data is becoming more and more convenient and fast, and its position in 3D data is becoming more and more important. However, in order to make good use of point cloud data, how to quickly and efficiently identify point cloud data is a basic problem that needs to be solved urgently. [0003] Point cloud data is a form of representation of three-dimensional data. It is a collection of points scanned by various devices that record the shape of the outer surface of an object in the form of points. Usually, each point contains X, Y, and Z three-dimensional coordinate information. Wi...

AI Technical Summary

Problems solved by technology

Wu et al. (Wu Z, Song S, Khosla A, et al. 3D shapenets: A deep representation for volumetric shapes[C] / / Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Washington DC: IEEE Computer Society Press. 2015: 1912-1920) first proposed the voxel-based 3D deep belief network 3DShapenets in 2015, and achieved 83.54% and 77.32% classification accuracy on the ModelNet10 and ModelNet40 standard data sets respectively, 3-5 points higher than traditional methods Percentage points. However, the nature of the voxel data itself makes the three-dimensional convolution calculation very redundant, and the performance is largely limited by the voxel resolution and the computational cost of exponential growth.

Due to the above shortcomings of three-dimensional convolution, Su et al. (Su H, Maji S, Kalogerakis E, et al. Multi-view convolutional neural networks for 3d shape recognition[C] / / Proceedings of the IEEE international conference on computervision.Washington D C:IEEE The MVCNN proposed by Computer Society Press, 2015:945-953) takes a set of two-dimensional views as input, constructs the initial features of the two-dimensional views through CNN, and fuses the features of each two-dimensional views through the view pooling layer to obtain three-dimensional data The characteristics of the classification are completed. The classification accuracy of this method on ModelNet40 is 89.9%, which is higher than the voxel-based deep learning classification method proposed in the same period. However, due to the existence of point cloud data (1) the point cloud is disordered; (2) The point cloud is sparse; (3) The amount of point cloud information is limited

Extending the above methods directly to point cloud data is not ideal

[0006] Qi et al. (C.R.Qi, H.Su, K.Mo, and L.J. Guibas. Pointnet: deep learning on pointsets for 3d classification and segmentation[C] / / Proceedings of the IEEEConference on Computer Vision and Pattern Recognition, pages 652–660, 2017.) The PointNet network proposed for the characteristics of point cloud data first applies deep learning to point cloud classification tasks. PointNet uses T-Net to achieve effective alignment of data and features, and uses Maxpooling symmetric function to extract order-independent global features. ModelNet40 A classification accuracy of 89.20% has been achieved on the above. However, PointNet cannot extract the local features of point cloud data, and a series of work has been designed on how to effectively extract the local features of point cloud data.

Images