Robust Filter-Based Visual Navigation Solution with Miscalibrated Bi-Monocular or Stereo Cameras

Abstract

This work addressed the problem of miscalibration or decalibration of mobile stereo/bi-monocular camera setups. We especially focused on the context of autonomous vehicles. In real-world conditions, any optical system is subject to various mechanical stresses, caused by vibration, rough handling, collisions, or even thermal expansion. Such mechanical stresses change the stereo pair geometry, and as a consequence, the pre-calculated epipolar geometry or any geometric-based approach is no longer valid. The standard method, which consists of estimating the calibration online, fails in such harsh conditions. The proposed method was based on a robust linearly constrained state estimation technique able to mitigate the model mismatch without estimating the model parameters. Therefore, our solution was able to mitigate the errors with negligible use of additional computing resources. We propose to use a linearly constrained extended Kalman filter for a stereo-based visual odometry or simultaneous localization and mapping approach. Simulations confirmed that the method kept the system (and objects of the map) localized in real-time even with huge miscalibration errors and parameter variations. The results confirmed that the method was robust to a miscalibration of all the extrinsic calibration parameters even when the standard online calibration procedure failed.