Research On Indoor Scene Localization Algorithm Based On RGB-D

Over the past years,with the fast development in computer vision,mobile robots are getting more and more applied.Simultaneous Localization and Mapping(SLAM)is the core technology for intelligent robots to achieve autonomous motion.Since people spend most of their time indoors,and the traditional localization based on Global Positioning System(GPS)and inertial navigation have the drawbacks of signal loss in indoor scenes,this paper focuses on the indoor localization algorithms based on RGB-D camera.The main contributions are summarized as follows:(1)This article proposes a novel algorithm for camera pose estimation of an indoor mobile robot under planar motion with an RGB-D camera.Through the introduction of the Manhattan World assumption,the six-degree-of-freedom motion estimation is reduced to a 3-degree-offreedom motion.The algorithm deals mainly with the corridor-like structured scenarios and different strategies are taken according to the prior knowledge about the 3D environment.When at least one vertical plane is detected with the depth data,camera pose estimation is realized based on one normal of the vertical plane and one point correspondence.When there are no vertical planes,a 2-point homography-based algorithm using only point correspondences is presented for the motion estimation.Then the proposed method is applied to a frame-to-frame visual odometry framework.We evaluate our algorithm on the synthetic data and show the application on the real-world data.The experiments show that the approach proposed in this paper is very efficient and robust for estimation of the camera pose in the Manhattan-like environments compared with the state-of-the-art methods.(2)In order to reduce the cumulative error of the camera pose estimation by long operation,and existing loop closure detection algorithms are not robust in environments with inconsistent path and large viewpoint variation,this article proposes an improved algorithm for loop closure detection.We extract feature points of keyframes and optical flow is applied to track features.The operational efficiency of optical flow is higher than traditional feature-point matching algorithms.We autonomously compute visual attention map of keyframes through an improved frequency domain-based method to guide the feature sampling within the range of salient regions.And a geometric matching probability is calculated to improve our algorithm's performance of distinguishing perceptual aliasing scenes.Finally our experimental results demonstrate that the algorithm proposed in this paper is very efficient and meets the real-time application for loop detection problems.