Research On Image Feature-Based Mobile Robot Stereo Visual SLAM

Simultaneous Localization and Mapping (SLAM) is one of the fundamental challenges of autonomous navigation, and also one of essential conditions to fulfill the intelligence for the mobile robot. SLAM is to combine localization and mapping into one procedure, that is to say, mobile robot incrementally builds a consistent map of the environment by its pose estimation and sensing, while simultaneously determining its location within this map.The paper focuses on the mobile robot visual SLAM, which is based on the image features of the indoor unknown environment. Without any prior knowledge, the robot obtains image of environment based on the binocular stereo vision, and extracts the image features as the natural landmarks. Then the feature-based geometry map is built, and simultaneously robot localization is fulfilled by the current pose estimate based on the matches between the image features and the landmarks in the prebuilt database. This dissertation does some in-depth studies on the following sub-problems in the visual SLAM: image local invariant features extracting algorithm, global localization and environment mapping based on the image features.In order to represent the distinctive information in the scene, the image local invariant features extracting algorithm is studied. As an important local feature, interest point preserves typical information and reduces the computation efficiently. Moreover it is invariant to the rotation transformation and almost not affected by illumination changes. The paper proposes a novel method for detecting scale, rotation, illumination, and affine invariant interest points, coined PLOT (Polynomial Local Orientation Tensor). PLOT is based on the local orientation tensor, which is constructed from the second order polynomial expansion of the image signal. The corresponding coordinates and scales features are obtained after extracting the PLOT interest points. The paper also uses the SIFT descriptor to describe its feature. PLOT shows strong performance based on the evaluation criterion using the recall vs. 1-precision graphs.Global localization and mapping are taken as two-in-one procedure in the robot SLAM. They interact and support each other. Accurate global localization relies on accurate map. The procedure of global localization is to obtain mobile robot current 3D coordinate and direction in the global coordinate system. It is based on the PLOT features to obtain a current best pose estimate by matching the PLOT features in the current frame with the PLOT landmarks in the built database under some constraints. Treating the global localization as model parameters estimation problem, the paper proposes a global localization algorithm by using extended RANSAC to obtain accurate position and orientation for mobile robot in its movement plane.The geometry map is built based on the 3D coordinates and other information of the PLOT features, such as orientations, scales, and so on. Once PLOT features are matched, their features are also obtained, such as 3D coordinates in the camera system, scales, orientations, descriptors and so on. They can be taken as the natural landmarks in the scene, and used to build the landmark database. As the robot moves in the environment, the database is updated every frame, including adding a new landmark, incrementing miss count or accumulative count, pruning a invalidate landmarks, etc.Experiments of mobile robot stereo visual SLAM based on the image PLOT features are implemented to verify the validity and accuracy of the proposed algorithm. Then the error analysis in mobile robot SLAM based on the PLOT features is discussed. The paper analyze the errors of the robot pose estimate and landmark 3D coordinate in the map using Kalman filter and error propagation formulae. The covariance error matrix is used to express their uncertainty. In the end the error sources and their reducing methods are discussed.