Mobile Robot Simultaneous Localization And Mapping In Unknown Environments

With the environments getting more and more complex and unstructured, the mobile robot Simultaneous Localization and Mapping (SLAM) in unknown environments has been considered to be an important and fundamental problem in the mobile robotics research domain. As the base of intelligent navigation and environments exploring, mobile robot SLAM problem has attracted significant attention within the mobile robotics research communities. Based on sensor information processing, particle filter localization and hybrid topological-metric mapping aspects, this dissertation does some in-depth studies on the above three aspects, the main contributions are as follows.1. Data fusion algorithm research based on visual and laser sensorsAs an autonomous robot, possessing visual and range acquisition capability is very crucial to explore unknown environments reliably, it is very important to obtain the local environments features by analyzing the sensor information effectively in robot SLAM researches. The image SIFT (Scale Invariant Feature Transform) features are invariant to image scaling and rotation, and partially invariant to changes in illumination and 3D camera viewpoint. It is a powerful tool to recognize local environments based on its highly distinctive aspects. Due to the large number of key points generated for each image and high dimensions of each vector with up to 128, SIFT feature extraction is a very time-consuming task. For resolving the above limitations, we propose a Maximum Gradient Rotated Normalized (MGRN) method based on each sub-area. Compute the gradients within 2×2 sub-areas around the key point, after rotated and normalized operations; we take only four gradients which including the maximum gradient as the key point vector for further image matching. This method can significantly reduce the key point vector's dimension. In addition, we select some key points with bigger magnitude proportionally and need not compute all points. Decrease the matching threshold to enhance the SIFT matching reliability. Optimize the line approximate algorithm based on laser data sets, set weight to each line extracted from local environments. Propose a data fusion algorithm based on Weighted Line Vector (WLV) and image SIFT key point. These strategies can effectively memory request and algorithm complicacy, enhance the real time computation capability.2. The SLAM problem research based on particle filter algorithm Nowadays most methods for SLAM are focused on probabilistic Bayesian estimation, such as the Kalman Filter (KF), Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), Maximized likelihood Estimation (MLE), Particle Filter (PF), Rao-Blackwellized Particle Filter (RBPF) and Markov localization algorithms. Based on comprehensive analysis of the above-mentioned algorithms, we propose an UKF-based Assistant-Proposal Distribution (UKF-APD) particle algorithm, propose the concept of Euclidean distance of particle approximate distribution to the UKF assistant proposal distribution, and take it as an adaptive particle-resampling criterion. Partial particles are sampled from the UKF-APD during the particle-resampling process; the method can avoid particles'impoverishment and deviation to the real posterior distribution. Assign important weights to each Topological Node (TN), we adopt the strategy to resample particles within topological node and optimized the particles'diversity by utilizing genetic algorithm. We employ the UKF algorithm instead of the EKF algorithm to estimate landmarks to avoiding the derivation of complicated Jacobian Matrix and reducing the error generated by linearizing the nonlinear system. These strategies can reduce the complexity and enhance the algorithm's real time speed and reliability.3. The research of Hybrid Topological-Metric (HTM) mapThe robot's main task is to build a map to represent the unknown environments abstractly by utilizing exterior sensors data, with the help of the map, robot can implement self-localization and finish more complicated tasks. We propose some new ideas and algorithms on Topological Node (TN) constructing, TN recognizing, and closed loop TN detecting problems. The local TN is composed of SIFT key points and laser scan data. The TN is constructed based on equidistant or new landmark method. The closed loop TN delay-estimation method enhances the accuracy of TN recognizing, the sub-linked matrix maintaining on line can improve real-time computational capability. The TN constructing and recognizing is completed by using the data fusion of SIFT features and Weighted Line Vectors.Experiment shows the validity of the algorithm.上海大学...