A large FOV camera calibration method based on predictive contribution multi-agent proximal policy optimization

Abstract High-precision calibration for stereo vision systems with a large field of view, which is several times larger than the target size, is a challenging task. The positions and poses of the targets are critical factors for calibration precision. In this paper, a target position and pose optimization method based on predictive contribution multi-agent proximal policy optimization is proposed. This method predicts the contribution of each target agent to assess the impact of their actions on the overall decision, and allocates corresponding weights to the reward of each agent. Then the action strategies are adjusted according to the rewards, thereby optimizing the positions and poses of the targets. Further, the optimized positions and poses are adopted to camera calibration process. The experimental results show that the back-projection errors are all less than 0.3 mm within a measurement range of 4.6 m × 2.3 m. In particular, compared with other reinforcement learning based calibration methods, the proposed method demonstrates significant advantages in terms of accuracy and stability.