Robust Kalman filtering with long short-term memory for image-based visual servo control

Abstract

Visual servo control systems based on Kalman filter (KF) is susceptible to noise interference, the initialization of the Jacobi matrix is difficult, and the observation value of the Jacobian matrix is not accurate. In order to address these problems, we proposed a robust KF algorithm with long short-term memory (LSTM) for an image-based visual servo control system and applied the system to an uncalibrated image-based visual servo (IBVS) control system to estimate the filtering gain error, state estimation error, and the observation error, which were then used for online training in LSTM. The visual servo control system controls the motion of the manipulator, and simultaneously updates the LSTM network. Therefore, the Jacobian matrix obtained using LSTM was employed to estimate the state volume of the robust KF, which constitutes a circulatory system, and the complementary effect was realized. The method was applied to a six-degrees-of-freedom manipulator of the eye-in-hand model to conduct experiments. The simulation results indicate that the proposed visual servo control system has strong anti-noise interference capability. Furthermore, it facilitates Jacobian matrix initialization and has high observation accuracy for the Jacobian matrix.