Unsupervised estimation method for synthesizing to real LiDAR point cloud scene flow

Abstract

The invention relates to the field of computer vision, and discloses an unsupervised estimation method for synthesizing to a real LiDAR point cloud scene stream, which comprises the following steps of: 1, compiling and generating a. Asi format dynamic link library file based on Scrip Hook V by utilizing a GTA-V game engine, copying the. Asi format dynamic link library file to a game path, and starting the GTA-V; according to the unsupervised estimation method for synthesizing the real LiDAR point cloud scene flow, data and labels required for training a network model can be directly generated by a game engine, manual labeling does not need to be consumed, practicability is achieved, meanwhile, a synthetic data set generation and domain adaptation method involved in the method has expandability, and the method is suitable for large-scale popularization and application. Attributes and scales of the generated data can be adjusted according to actual conditions, and the domain adaptation method can be conveniently deployed in various existing mainstream scene flow estimation networks.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to an unsupervised synthesis-to-real LiDAR point cloud scene flow estimation method. Background technique [0002] Scene flow is defined as a three-dimensional motion field (motionfield) between two consecutive frames of input (such as RGB-D images, three-dimensional point clouds), which represents a point-by-point motion vector in three-dimensional space. Usually, scene flow is used to describe the motion state of three-dimensional objects within a certain time interval, and can be used to estimate the future motion trend of objects, which has important practical significance for robot navigation, automatic driving and other fields. Early work usually estimated scene flow from RGB-D images, but with the popularization of 3D sensing technology and the development of deep learning field, the use of deep learning methods to estimate point-by-point scene flow from continuously ...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the deficiencies of the prior art, the present invention provides an unsupervised synthesis-to-real LiDAR point cloud scene flow estimation method, which solves the problem that the data and labels required for the traditional training network model need to be manually marked, and the traditional synthetic data set generation And domain adaptation methods do not have scalability, the quality of traditional synthetic data is poor and the generalization performance of domain adaptation methods is weak

Images