Research On RGB-D Visual Odometry For Embedded System

In recent years,with the development of image sensing technology and computer technology,the ego-motion estimation method based on machine vision(visual odometry)has become a research hotspot in the field of robotics;this technology can be used for the mobile robots working in indoor environments,providing stable and reliable state estimation without the needs of GPS signal.It is the basis for the subsequent decision-making and motion control of the robot system,and is of great significance to the research of mobile robots.However,the real-time image analysis performed by the visual odometry system has high computational complexity,which put extremely high demands on the computing power of the equipment,thus it is difficult to deploy in a low-power-consumption embedded system.In response to this problem,an efficient visual odometry that can stably work in embedded system with real time performance is proposed,which is benefits from the parallel computing capabilities of the embedded GPU hardware,plenty of processing of feature-based visual odometry are adapted according to the sampling characteristics of the RGB-D camera.Firstly,the mathematical description of rigid body transformation in threedimensional space and the imaging principle of RGB-D camera are briefly introduced,the observation model of RGB-D camera is constructed;for the state estimation problem of the 6-Do F camera pose,the 3D-2D Pn P algorithms and 3D-3D point cloud registration methods are derived respectively;Then,the basic model of our visual odometry is constructed according to the classic framework of feature-based visual odometry.The works above are essential for the subsequent algorithms research and improvement.In the front-end of the RGB-D visual odometry,we make full use of the parallelism of embedded GPU hardware and algorithms,improving the efficiency of image feature extraction,3D feature matching and camera pose estimation based on the GPGPU technology,which eventually improves the real-time performance of the system;In feature matching processing,a novel matching filtering method based on histogram statistics is proposed,which is based on the consistency check of the Euclidean distance,filtering out the set of interior points(correct matches)of the input matches,so the follow-up can directly use the closed-form point cloud registration algorithm estimates the relative transformation of the camera.Compared with the traditional RANSAC-based method,this method significantly improves the pose estimation efficiency of the input frame.In the back-end research,we derived the optimization problem of system based on the maximum a posteriori estimation,the sparse structure of the back-end optimization problem is analyzed,and an information integration method for constructing a graph model based on the feature matching relationship between keyframes is studied,which accurately establish the nodes and edges in graph model.In the optimization process,the optimization effect of the system and the solution rate are weighed,and the optimization strategy of local bundle adjustment is adopted to effectively suppress the increase of the cumulative error of the system.Finally,the basic model of RGB-D visual odometry and the improved front-end and back-end modules are integrated,the above algorithms are deployed to the NVIDIA Jetson TX2 embedded computing platform,and experimental testing and analysis are carried out through the public RGB-D datasets,verifying the real-time performance and reliability of the system's work.The results show that our system achieves realtime data processing,robust relative pose estimation,the estimated trajectory of camera maintains good consistency for a long time.