Localization Theories And Methods Without Relying On GPS

An intelligent vehicle, or manned autonomous vehicle, is one of the top issues in both scientific research and product development. To be truly autonomous, autonomous vehicles should have the capability to estimate their own locations and learn about the surrounding environment in unknown dynamic circumstances. This paper mainly focuses on the localization algorithm without GPS and tries to get accurate positioning information for autonomous vehicles in this circumstance.In most cases, building a motion model or equipping odometer for vehicle is not so practical, because traditional algorithms with vehicle model cannot be used into any vehicle localization. To locate the vehicle without odometers, a proposed model of estimating robot motion state is introduced. The proposed algorithm estimates the robot position, its pose and its motion state (such as speed) during localization. The proposed algorithm uses estimating motion state instead of odometer and then predicts the vehicle loacation. To verify the performance of the algorithm, the paper will compare the proposed algorithm to algorithm with odometry information by simulation and Victoria database of Sydney University. Experimental results show that the proposed algorithm can be achieved same accuracy of the algorithm with odometer.Then, this paper has a research on visual odometry and presents a new visual odometry algorithm. By decoupling rotation-translation estimation, the proposed algorithm enhances visual odometry performance in a dynamic environment. Ideal visual odometry has the ability to deal with the motion estimation by observing static environment feature, but there are a large number of dynamic features in physical environment. So eliminating dynamic features and reducing its impact on the visual odometry is the effective way to improve the performance of visual odometry. Based on stereo vision system, the proposed algorithm divides feature points into "far point" and "near point" and handles them separately:"far points" will be used to estimate the posture of the visual system; with the rotation constraints, "near points" will be used to calculation camera translation. Here, rotation constraints reduce the impacts of attitude moving objects up close visual odometer. Experimental results show that on the road, the proposed algorithm with decoupling the rotation-translation estimation can more effectively eliminate the dynamic features than the algorithm with simultaneously rotation-translation estimation. Then, the proposed algorithm improves the performance of visual odometer.To achieve better localization result, simultaneous localization and mapping (SLAM), as an important technique for autonomous mobile vehicle navigation, is researched.In the context of SLAM, the current pose of the vehicle is updated by the measurement of the surrounding environment and an incremental map is simultaneously built on the basis of the estimated vehicle path. Although most existing SLAM methods can effectively reduce the computational cost, they suffer from the fatal drawback of accumulating higher-order nonlinear errors. This paper presents new algorithms, named Square-root cubature quadrature Kalman based SLAM algorithm (SCQKF-SLAM) and cubature quadrature based FastSLAM (CQFastSLAM) by combining the cubature rule with the Gauss-Laguerre quadrature rule to achieve higher estimate accuracy. The proposed method is scalable and applicable to systems with higher-order nonlinearities. Comprehensive examination on the computational cost and accuracy indicates that the proposed SCQKF-SLAM algorithm and the proposed CQFastSLAM algorithm are effective and flexible tools for enhancing the SLAM performance.A series of simulation and experimental results also validate that under large-scale environment, the proposed algorithms improve the performance of SLAM significantly.