Proposing Camera Calibration Method Using PPO (Proximal Policy Optimization) for Improving Camera Pose Estimations

Abstract

This paper highlights camera orientation estimation accuracy and precision, as well as proposing a new camera calibration technique using a reinforcement learning method named PPO (Proximal Policy Optimization) in offline mode. The offline mode is used just for extracting the camera geometry parameters that are used for improving accuracy in real-time camera pose estimation techniques. We experiment and compare two popular techniques using 2D vision feedbacks and evaluate their accuracy beside other considerations related to real applications such as disturbance cases from surrounding environment and pose data stability. First, we use feature points detection ORB (Oriented FAST and Rotated BRIEF) and BF (Brute-Force) matcher to detect and match points in different frames, respectively. Second, we use FAST (Features from Accelerated Segment Test) corners and LK (Lucas-Kanade) optical flow methods to detect corners and track their flow in different frames. Those points and corners are then used for the pose estimation through optimization process with the: (a) calibration method of Zhang using chessboard pattern and (b) our proposed method using PPO. The results using our proposed calibration method show significant accuracy improvements and easier deployment for end-user compared to the pre-used methods.