High-precision universal three-dimensional point cloud identification method

Abstract

The invention discloses a high-precision universal three-dimensional point cloud recognition method. The method comprises the following steps: performing nonlinear mapping on feature information of a point cloud to obtain representations in different transformation spaces; extracting local coding features from the obtained representations by using a graph attention network, obtaining global coding features through global space fusion, and obtaining fused point cloud features from the local coding features and the global coding features through a gated feature aggregation mechanism; sending the fused point cloud features into a classifier to obtain a prediction result; and constructing an overall prediction loss function, and adding independence discrimination loss for network training and optimization, and storing appropriate model parameters at the same time. According to the method, point cloud features are effectively extracted by using a graph attention convolutional neural network in each transformation, efficient fusion is performed through a gating mechanism, a Hilbert-Schmidt independence index is introduced to measure the similarity of the point cloud features, and information redundancy is reduced by minimizing the feature similarity, so that richer feature representation can be obtained.

Description

technical field [0001] The invention relates to the field of computer vision technology, and belongs to a three-dimensional recognition technology based on deep learning. Specifically, it is a high-precision and general-purpose three-dimensional point cloud recognition method, which completes the classification of point clouds in scenes by establishing a unified network framework , partial segmentation and semantic segmentation tasks to achieve high-precision point cloud recognition. Background technique [0002] In recent years, the recognition of two-dimensional images has achieved certain success in applications, but since objects in actual scenes exist in three-dimensional form, effective understanding of three-dimensional scenes can promote intelligent systems (such as robots, autonomous driving, virtual / The rapid development of technologies such as augmented reality). As a data form that can intuitively represent 3D scene space and geometric shape information, point ...

AI Technical Summary

Problems solved by technology

These methods all use a single transformation to capture the representation of the point cloud context. However, there are two problems: First, due to the variety of objects in the actual scene and the short spatial distance between different objects, feature aliasing is prone to occur. It affects the result of point cloud recognition; secondly, these methods use a single transformation to extract the features of the point cloud, without considering multiple angles, so the feature information obtained is not rich enough to meet the needs of the application; in addition, in other Multi-transform methods in the field have been proven to have good performance to help networks learn richer feature representations

Images