Estimating the position and orientation of a mobile robot using neural network framework based on combined square-root cubature Kalman filter and simultaneous localization and mapping

Abstract

The real-time performance of target tracking, detection, and positioning behaves not well for non-Gaussian and nonlinear model with circumstance uncertainty. The weak observability of the system under large noise causes the algorithm unstable and slow to converge. A new estimation algorithm combining square-root cubature Kalman filter (SRCKF) with simultaneous localization and mapping (SLAM) is proposed. By connecting neural network weights, network input, functional types and ideal output network, the algorithm firstly update iteratively the SRCKF-SLAM state model and observation model, then conduct the cubature point estimate (weights) neural network framework. Thus, a point set better representing the target state and a more accurate state estimation are achieved, which can improve the filtering accuracy. This paper also estimates robot and characteristic states by filtering in groups. The simulation results showed that the proposed algorithm is feasible and effective. Compared with other filtering algorithms such as SRUKF and SRCDKF, it improves the estimation accuracy. Applying the new algorithm to the position filtering estimation of mobile robot can effectively reduce the positioning error, achieve high-precision tracking detection, and improve the accuracy of robot target detection.