A Study On Simultaneous Localization And Map Building For Intelligent Mobile Robots

When the society comes into an age of information and intelligence, demand for the intelligent mobile robot grows day by day in every walk of life. Mars Rovers 'Spirit' and 'Opportunity' landed on the Mars successfully. Army mobile robots were applied on the battlefield of Iraq. These events indicate that the intelligent mobile robot has come into common people's view. Autonomous navigation is an important researching domain in the intelligent mobile robot community. And Simultaneous Localization and Map Building (SLAM) is the key problem to be solved before a robot navigates autonomously. The SLAM problem means that a mobile robot is placed in an unknown environment and the robot incrementally builds a consistent map of this environment while simultaneously determining its location within this map. The SLAM problem has gradually become the core of the mobile robotics community in the past decade as a solution to the SLAM problem would provide the means to make a robot truly autonomous.The solutions and difficulties of the SLAM problem are studied in this paper. And somewhat in-depth studies are carried on the following aspects including Kalman filters based SLAM solution, particle filters based SLAM solution, data association in SLAM, and map model in SLAM. The detailed contents of the paper can be described as follows.When Extended Kalman Filter (EKF) is used to solve the SLAM problem of a nonlinear system, the error brought by linear prediction and update will lead to severe estimation error or even make the method to be divergent. Two improved approaches, 'Stepwise Increment' and 'Mean Slope', are suggested to solve the linearization error problem of EKF. The approach 'Stepwise Increment' helps to decrease the linearization error by gradually absorbing the innovation between the predicted observation and the actual observation, while the approach 'Mean Slope' works by revising the Kalman gain. When the approach 'Mean Slope' is combined with EKF, a new Kalman filter named 'Mean Extended Kalman Filter' (MEKF) is put forward. Simulation results indicate that the estimation accuracy of MEKF is much better than the one of other existing Kalman filters in the full state SLAM applications. In addition, MEKF performs very well on computational efficiency.Rao-Blackwellised particle filter is used to estimate the path and map recursively in the FastSLAM algorithm. However, the computational requirement of particle filter is hard. To further improve the computational efficiency of the FastSLAM algorithm, we put forward a SLAM solution named 'Fast Kalman SLAM'. It inherits the factorial estimation framework, which is the advantage of FastSLAM. The robot pose is estimated recursively with MEKF or unscented Kalman filter (UKF). And the estimates of observed features in the map are updated with EKF. The presented algorithm, named 'Fast Kalman SLAM', reduces computational complexity greatly and ensures estimation accuracy at the same time. Thus, an effective solution is provided to instantly solve the SLAM problem in complex environments.Particle filter is a feasible kind of tool to solve the SLAM problem of a non-linear and non-Gaussian system. And the FastSLAM 2.0 algorithm is a representative SLAM solution based on particle filter. However, it possesses problems such as "sample degeneracy" and "sample impoverishment". In order to improve the samples' performance and to improve the solution's integrative performance, three improved solutions are introduced. The first approach is to combine FastSLAM 2.0 with genetic algorithm, and a solution named "Genetic FastSLAM" is presented for the SLAM problem. Based on the specialty of FastSLAM 2.0, an improved genetic algorithm is designed with attention to both each particle's weight and the samples' diversity. The second approach is to combine Auxiliary Marginal Particle Filter (AMPF) with the factorial SLAM framework, and a solution named "Auxiliary Marginal FastSLAM" (AMF-SLAM) is put forward. AMF-SLAM estimates the robot pose with AMPF. Appropriate sampling strategy and particle data structure are designed for the compatibility with both AMPF and the factorial SLAM framework. Then, map is estimated with EKF in the factorial SLAM framework. The third approach is to still combine AMPF with the factorial SLAM framework. However, a single map is maintained in the entire SLAM process. And the samples are used merely to represent the robot pose at a time instance. Thus, the logical complexity of the solution is simplified. Simulation results and the 'Victoria Park Dataset' results indicate that Auxiliary Marginal FastSLAM performs the best in estimation accuracy and stability among all the three improved SLAM solutions based on particle filter. In addition, AMF-SLAM meets the demand for on-line use, given the precondition that the quantity of the samples is properly set.Two amelioration approaches, local association approach and dynamic association approach, are introduced for the data association process in SLAM. And the local association approach is used to improve computational efficiency of the data association process, while the dynamic association approach is used to eliminate the impact caused by observation noises and dynamic features and laid on the data association process. With a view to the weak points of Joint Compatibility Branch and Bound (JCBB) and Nearest Neighbor (NN), the local association approach and dynamic association approach are used to improve JCBB and NN. As a result, Optimized JCBB (OJCBB) and Dynamic Joint Nearest Neighbor (DJNN) are introduced. And the realization of DJNN is simple and convenient. What's more, DJNN performs well on both accuracy and computational complexity. Therefore, Dynamic Joint Nearest Neighbor, presented in the paper, is of important application value.Sparse feature map is a widely used environmental model in the SLAM community. However, it can be used merely to improve localization accuracy. It fails to provide sufficient environmental information to a robot. With a view to the physical meaning of the observations, Global Observation Map Model (GOMM) is put forward. Necessary observations are chosen with displacement rule, feature rule, and limit rule. Then, filtration and transformation are applied on the selected observations. And global dense map about the environment is obtained according to physical meaning of the observations and uncertainty of the robotic location estimate. Experiments are carried out to evaluate the performance of the Global Observation Map Model based EKF-SLAM (GOE-SLAM). Experimental results indicate that the global dense map built with GOE-SLAM is creditable. And the computational efficiency of GOE-SLAM is nearly equal to that of EKF-SLAM. Therefore, Global Observation Map Model is of great practical value in the SLAM community.In the end, the whole paper is summarized. And the prospective researching fields are discussed.