Three-dimensional fully convolutional network realizing equipment

Abstract

An embodiment of the invention provides three-dimensional fully convolutional network realizing equipment The equipment comprises the components of a receiving unit which is used for receiving initialpoint cloud information that has a tree-shaped storage structure; and a convolutional unit which is used for performing multilayer convolutional and fully convolutional operation on the initial pointcloud information through the three-dimensional fully convolutional network, thereby obtaining reverse volume pixel prediction information which comprises a plurality of kinds of information. Becausea 3D environment comprises abundant set characteristic information, a traditional vision-based settlement method cannot effectively perform real-time analysis on a 3D scene, the embodiment of the invention provides a 3D convolutional solution based on the tree-shaped structure, semantic classification to the 3D point cloud environment can be finished, thereby improving understanding of the 3D environment.

Description

technical field [0001] The invention relates to the field of computers, in particular to a device for realizing a three-dimensional fully convolutional network. Background technique [0002] In recent years, with the continuous development of deep learning, great breakthroughs have been made in the field of image object detection, object recognition, scene segmentation, scene understanding and other issues, but the current deep learning framework is basically based on 2D There is still relatively little research work on 3D data (3D data based on depth information or 3D information based on laser point cloud), and one of the key problems is that the computational complexity of the convolution operation for 3D data needs to be Much larger convolution operations on 2D data. [0003] Taking ImageNet as an example, if the input image size of the first layer of convolutional network is 244×244×3 (3-channel RGB), the output is 96×7×7 (96 filters, 7x7 convolution kernel), Then the...

AI Technical Summary

Problems solved by technology

However, for the 3D point cloud data of unmanned vehicle environment perception, if the three-dimensional space of 50m x 50m x 10m is taken, and the grid resolution is 0.1m, the corresponding data size is 500×500×100 (regardless of color information), The computational complexity brought by this is far from meeting the real-time requirements

Images