Mobile Robot 3D Environment Modeling Based On Variational Model

A fast and efficient mobile robot map building is one of the important reflections of robot’s intelligent perception for fulfilling some complex tasks, such as its self-localization, action planning and dynamic decision by using a variety of sensors. In an unknown environment, simultaneous localization and map(SLAM) is one of hot research in mobile robot. In this paper, some researches have been done for the vision based mobile robot pose estimation and environment modeling. Firstly, an improved probability-guided random sample consensus(IPGSAC) is proposed to address the robust pose estimation in the vision system. Secondly, two map building algorithms are proposed for mobile robot by using the sparse image feature matching and the hybrid model with the advantages of active vision sensor. On this basis, the monocular vision based mobile robot odometry estimation algorithms are proposed by employing the ground features. Furthermore, an improved PTAM(Parallel Tracking and Mapping) algorithm is proposed to achieve a image features map. Finally, a visual feedback based variational depth estimation algorithm is employed for the mobile robot dense 3D map building with the sparse features map. The achievements of the thesis can be summarized as follows:(1) In some improve RANSAC(Random Sample Consensus) algorithm, the evaluations of test points are estimated by using partial data during the iterations. In this paper, the IPGSAC algorithm is proposed to deal with the low reliability of those algorithms. A hybrid probability distribution is employed to initialize of probabilities of test points. By using the total probability formula and statics of the size of maximum inlier set contains the test point, a multi-source evaluation strategy is employed for all test points. According to DS(Dempster-Shafer) evidence theory, a more reliable evaluation of test point is achieved by fusing the multi-source evaluations. Furthermore, the forget theory is adopted to reduce the influence of incorrect judgment of test points and improve the reliability of the proposed algorithm. Based on this updating probability evaluation, the inverse mapping sampling principle is utilized to improve the sample possibility of test points with high probability and speed up the convergence rate of the proposed algorithm. The experimental results demonstrate that the proposed algorithm has more advantages in convergence rate and precision than the standard RANSACA, MSAC, NAPSAC and MLESAC.(2) As the active vision sensors are easily to achieve the depth of the scene, two SLAM algorithms are proposed based on a depth camera. Firstly, an image feature matching procedure is used to establish the data association during the map building. Based on the proposed IPGSAC algorithm, the pose estimation model is solved to achieve the accurate localization of mobile robot. Secondly, a hybrid pose estimation-based robot mapping algorithm is proposed to overcome the low geometry constraint during the 3D map building. Combining with the local texture association, the hybrid method is established to ensure the accuracy of pose estimation and decrease the failure rate with weighted the point cloud and image texture model. Finally, the keyframe selection mechanisms are employed according to the proposed estimation algorithms and a vision based loop detection and optimization algorithm is adopted to achieve a global consistency map. The indoor experimental results demonstrate the efficiency and feasibility of the proposed algorithm.(3) Inspired by the active vision based mobile robot pose estimation approach, the monocular visual odometry algorithm is proposed with ground features for mobile robot. The data association is established by using the ground features, and a homography-based pose estimation model is used for localizing the robot. Based on the framework of RANSAC algorithm, the pose estimation algorithm is solved to roughly remove the non-ground feature and achieve an initial pose estimation of mobile robot. To improve the accuracy of localization, the non-ground features are excluded by using difference of the normal estimation under the triangulation structure of features or the spatial structure of the features. Furthermore, a dense non-ground segment and localization algorithm is proposed by employing the dense optical flow estimation. Finally, the visual odometry algorithm is applied to improve the PTAM algorithm. A sparse 3D map is built with the homography constraint during the epipolar searching for the mobile robot. The experimental results demonstrate that the high accuracy of the proposed visual odometry algorithm and the improved PTAM algorithm is able to build a high quality 3D sparse feature map.(4) To overcome the lack of geometry constraint in the traditional variational model, the smoothing of the optical flow achieved in the observation view is employed to establish a feedback mechanism. And a vision feedback based variational depth estimation algorithm is proposed for the mobile robot dense 3D map building. The discrete depth space sampling and triangulation structure of features are employed to estimate the initialization of the depth. The proposed model is optimized to achieve the dense depth estimation of the environment by using a dual-primary algorithm. Finally, a variational model based monocular camera dense 3D reconstruction algorithm is achieved for mobile robot with the sparse feature map. The experimental results demonstrate that the proposed algorithm is able to achieve a dense 3D modeling for mobile robot in a certain degree and verify the feasibility and effectiveness of the proposed algorithm.