Research On3D Simultaneous Localization And Mapping Fusing Color And Depth Information

Abstract:Simultaneous Localization and Mapping (SLAM) is the basis for mobile robot autonomous navigating in unknown environment, and also one of the prerequisites for realization of autonomous and intelligent. In recent years, the theories and methods of two-dimensional map building have been comprehensively studied and have achieved fruitful results. With the advances in sensor technology and the continuous development of SLAM computation theory, three-dimensional map building for the6-DOF robots has attracted researchers'attention. In June2010, Microsoft Corp launched a cheap RGB-D sensor named Kinect, which provides new possibilities for creating environment map with rich3D spatial information and color texture information.This paper conducted simultaneous localization and map building based on RGB-D color information and depth information for indoor unknown environment. Without any prior knowledge, a Kinect does6-DOF motion in indoor scenes and perceive the surrounding environment information, extracting stable feature points of the environment to represent the actual physical point in3D space which is used as landmarks to create feature-based geometry map of the environment. The research work includes the following five parts:(1) Reviewed the applications of the typical depth camera "Kinect" in computer vision processing. To resolve the significant depth distortion inherent in Kinect, which degree aggravated with increasing distance, an unsupervised learning algorithm without human intervention for depth multiplier image is proposed to achieve the purpose of depth correction. It first builds the environment map using a common visual odometry+pose graph optimization RGB-D SLAM algorithm from the relatively high accurate measurement data of short distance measurement (during the process loop-closing is needed). Then, the depth multiplier image is studied driven by the errors between the map and the depth measurement data, using the maximum likelihood estimation method gradually to optimize. Differ from the methods that require human intervention, this method can complete the learning of depth correction automatically during SLAM process, which makes it easier to use.(2) In order to reduce the complexity of SLAM and increase the credibility of data association, interest points detection algorithm is deeply investigated. By analyzing the two parameters, threshold t and octave parameter o, impact on the performance of the BRISK-AGAST detection algorithm in OpenCV, an adjustable adaptive feature detection algorithm is proposed:adjustable-BRISK-AGAST detector. The new detector has the advantages that enhancing the stability of the extracted feature points, increasing the probability and reliability of data association in SLAM process, and avoiding excessive environmental features indicated in the map so as to reducing the complexity of SLAM.(3) In order to take full advantage of RGB-D image depth information to more effectively distinguish between points of interest, the RGB-D image descriptors fusing appearance and geometric shape information are studied, focusing on the analysis of the mechanism of BRAND descriptor. Experimental results show that BRAND descriptor is superior to EDVD, SURF, SIFT, CSHOT, SPIN in processing time, memory consumption, matching performance.(4) The current graph optimization based RGB-D SLAM algorithm is not suitable for online applications because in many cases the error accumulation will be very large due to absence of loop-closing. In order to overcome this defect, a new RGB-D SLAM method based on visual odometry and extended information filter, referred to as VO-EIF SLAM, is proposed. Using the pinhole camera model and the depth uncertainty measure model based on Gaussian mixture, a RGB-D features'three dimensional uncertainty measure model is established, which can be seen as the observation model of EIF SLAM. A visual dead reckoning algorithm based on visual residuals is devised, which is used to estimate motion control input. In addition, our observation model considers observations as sets of landmarks determined by their3D positions and their BRAND descriptors. We avoid explicit data association by marginalizing out the observation likelihood over all the possible associations, thus overcoming the problems derived from establishing incorrect correspondences between the observed landmarks and those in the map.(5) Deeply studied a fast feature point association algorithm based on binary descriptors, and applied it to address the problem of loop-closing for RGB-D SLAM. Designed and implemented two fast binary features searching algorithms to solve the problem of fast data association for single feature point:the locality-sensitive hashing searching algorithm and hierarchical clustering based searching algorithm. Based on these, put forward a kind of multi feature point matching algorithm fusing local geometric constraints, thus achieve the purpose of quick close-loop detection for RGB-D SLAM. These algorithms use Hamming distance to compare matching degree and effectively improve the speed and accuracy of loop closure detection.The conclusions and directions for future research work are discussed in the last chapter. There are53figures,3tables and189references.