Point cloud object identification method with scale invariance

Abstract

The invention discloses a point cloud object identification method with scale invariance. The method comprises the following steps: step 1, extracting convolution center points of a multilayer convolutional neural network layer by layer according to first point cloud data in a model sample, calculating scale invariant features of a last convolution layer in the multilayer convolutional neural network according to features of the convolution center points, and generating a point cloud recognition neural network; step 2, training a point cloud recognition neural network by utilizing a classification loss function according to second point cloud data in the training sample; and step 3, identifying a to-be-identified three-dimensional point cloud object of any scale by using the trained pointcloud identification neural network, and outputting an identification classification result. Through the technical scheme in the invention, the influence of point cloud data scaling on the output of the point cloud object recognition network is reduced, and the learning efficiency of the point cloud object recognition network on the three-dimensional object is improved.

Description

technical field [0001] The present application relates to the technical field of computer vision, in particular, to a point cloud object recognition method with scale invariance. Background technique [0002] Point cloud is the original data structure that can be collected by many three-dimensional environment sensing devices such as lidar and depth camera. It has attracted much attention because of its high accuracy and precision in representing three-dimensional objects. With the popularization of these acquisition devices in recent years, point clouds are widely used in fields such as autonomous driving, architectural design, and game modeling. In the field of academic research, with the rise of deep learning and the rapid improvement of computing power, point cloud data with a huge amount of information has gradually become one of the most popular data representations in the field of 3D object recognition. [0003] In real applications, due to the uncertain relative pos...

AI Technical Summary

Problems solved by technology

[0004] However, since most deep neural networks do not have scale invariance and are sensitive to the overall size of the input value, the existing point cloud object recognition technology based on deep neural networks cannot well adapt to the scale changes of the collected objects in the real environment.

[0005] In the existing technology, in order to solve this problem, methods such as data enhancement and feature enhancement are often used to process the collected data of the collected object, so as to approximate the adaptation to the scale, but such a solution does not fundamentally solve the problem. And it will bring many problems, such as: reducing the efficiency of network training, bringing additional overhead, etc.

Images