Robotic Visual-Inertial Calibration via Deep Deterministic Policy Gradient Learning

Visual-inertial calibration is important in robotic vision navigation systems, and calibration errors will reduce navigation accuracy for the longtime autonomous operation. Aiming at the problems of the complicated offline calibration process and the high calculation cost of self-calibration, a novel visual-inertial calibration method using deep deterministic policy gradient learning is proposed. Firstly, the error model of visual-inertial calibration is established considering the intrinsic and extrinsic parameters of the camera and IMU simultaneously. Secondly, the nonlinear observable analysis of the visual-inertial system is carried out. The rank decomposition of the Fisher information matrix is used to establish the relationship between the parameters to be calibrated and the nonlinear observability. Then, the visual-inertial self-calibration process is modeled as a partially observable Markov decision process to facilitate the design and optimization of subsequent reinforcement learning policies. Finally, a reinforcement learning network model is established for visual-inertial calibration using deep deterministic policy gradient, which is used to determine unobservable motion sequences. Meanwhile, the experience playback and target network are adopted in visual-inertial calibration algorithm to solve the problem of hyperparameter training and the instability of the network. Experiments in two different environments show that the proposed method achieves comparable performance comparing with the informative segment approach and the batch calibration approach. Moreover, the proposed method has the shortest trajectory length selected for calibration.