Comparison of multiple fault detection methods for monocular visual navigation with 3D maps

Abstract

Within the newly defined 3D maps, many extracted visual keypoints have been assigned with real-world coordinates. Such geospatial information can make monocular visual navigation feasible as a camera on the user platform that can capture the common keypoints within the 3D maps, and then, the coordinates and attitude of the user's platform can be determined. However, multiple faults within visual measurements produced through the keypoint matching process often exist with a high possibility due to various reasons, such as illumination changes, image noise, mismatches and calibration biases. Besides, the corresponding world frame coordinates of these keypoints may also contain faults. Moreover, these faults usually do not appear individually, which means that multiple faults are frequently encountered in vision-based navigation. All these factors will lead to failures in navigation. Therefore, multiple fault detection methods are necessary for indoor monocular vision based navigation. In this paper, six multiple fault detection methods, which include forward search (FS), least median squares (LMS), least trimmed squares (LTS), M estimator, S estimator and MM estimator, are tested and analyzed. The experimental results reveal their feasibility and potentials for use in indoor monocular vision based navigation. At the same time, with detection capability and false alarm rate acting as two performance indicators, the Monte Carlo simulation in the three indoor scenarios demonstrates that MM estimator and LTS estimator have the best performance with high detection capability and low false alarm rate.