Three-dimensional point cloud data instance segmentation method and system in automatic driving scene

Abstract

The invention provides a three-dimensional point cloud data instance segmentation method and system in an automatic driving scene. The method comprises steps of carrying out the preliminary recognition and division of an outdoor street scene through the spatial position information of a target object, and forming a point cloud visual column of an interested region; visual column point clouds containing objects and negative sample visual column background point clouds distributed in the same way are extracted from the point cloud visual columns of the region of interest to form a visual columnpoint cloud data set; and extracting high-dimensional semantic feature information of an object contained in each visual column point cloud in the visual column point cloud data set, and meanwhile, introducing a multi-classification focus loss function with a weight to obtain a category to which each point cloud in the visual column belongs, thereby realizing instance segmentation of the point cloud data. According to the three-dimensional point cloud data instance segmentation method in the automatic driving scene, target detail feature expression can be effectively enhanced so that the prediction capability of point cloud difficult samples can be enhanced and the performance of point cloud instance segmentation in the automatic driving scene can be enhanced.

Description

technical field [0001] The present invention relates to the technical field of computer vision, in particular to a method and system for segmenting three-dimensional point cloud data instances in an automatic driving scene. Background technique [0002] With the application of 3D sensors such as lidar in unmanned driving and robot projects, 3D data has attracted more and more attention from academia and industry. Point cloud is an important form of three-dimensional data representation. It is obtained through direct measurement, which is similar to the way humans observe the world. It can preserve the real three-dimensional structure information of objects to the greatest extent. Computer vision tasks based on point clouds are of great practical significance. How to construct mathematical models to represent, process and analyze point cloud data has become an urgent problem in the field of autonomous driving. [0003] For instance segmentation of 3D point cloud data, how to...

AI Technical Summary

Problems solved by technology

However, the projection from 3D to 2D in this algorithm causes irreparable information loss, resulting in limited learning ability of convolutional neural network.

In addition, complex background noise seriously affects the recognition of target objects, making the overall segmentation accuracy very low. Therefore, the performance of the algorithm needs to be improved. Even if the calculation rate is increased, it is still very far away from unmanned driving applications.

[0006] In addition, after investigation, it was also found that the current point cloud segmentation algorithm shows a great imbalance between categories. The model is overfitting for simple samples and underfitting for difficult samples, resulting in low average recognition accuracy of objects. An Efficient Algorithm for Improving Instance Segmentation of Difficult Samples in Point Clouds

Images