Method, device and apparatus for estimating point cloud object attitude based on deep learning

Abstract

The embodiment of the invention provides a method, device and apparatus for estimating the point cloud object attitude based on deep learning. The method comprises the steps of obtaining data requiredto be learnt; designing a network model; and performing model training and prediction. The design of the network model comprises the steps of modeling a point cloud object attitude estimation probleminto a non-distinctive multi-classification problem; designing a residual block structure to extract features; obtaining global feature through a maximum pooling layer according to extracted features; respectively sending the global features to three parallel multi-layer perceptrons to perform predicted category scoring on coordinate axes; performing final category prediction on features after predicted category scoring on the coordinate axes by using a classifier; carrying out equal-weight summation on loss values obtained after processing of the classifier, and enabling the sum to serve asan overall multi-classification loss function; and optimizing the multi-classification loss function by using adaptive moment estimation. The method provided by the embodiment of the invention can accurate estimate the point cloud object attitude so as to improve the accuracy of object attitude positioning and prediction.

Description

Technical field [0001] The present invention relates to the field of computer vision, in particular to a method, device and equipment for estimating the posture of a point cloud object based on deep learning. Background technique [0002] At present, in the research of deep learning methods for 3D point clouds, 3D point clouds are often projected onto 2D planes or converted into 3D voxels to adapt to the highly regular input data format of convolutional neural networks. Man-made noise is often introduced in the process of conversion operation. In order to remove man-made noise, the amount of calculation is often increased. The process is more complicated, and the introduction of man-made noise will seriously affect the accuracy of the object attitude evaluation, resulting in the existing evaluation of the object attitude The accuracy rate is not high, which seriously affects the effect of the technology in the application field. Summary of the invention [0003] In view of this, ...

AI Technical Summary

Problems solved by technology

[0002] At present, in the research of deep learning methods for 3D point clouds, 3D point clouds are often projected onto 2D planes or converted into 3D voxels to adapt to the highly regularized input data format of convolutional neural networks, but in the data Artificial noise is often introduced during the conversion operation, and the calculation amount is often increased in order to remove the artificial noise. The process is more complicated, and the introduction of artificial noise will seriously affect the accuracy of the object pose assessment, which leads to the existing evaluation of the object pose. The accuracy rate is not high, which seriously affects the effect of the technology in the application field

Images