Abstract:
This paper proposes a Kalman-filter-based method for camera trajectory correction in monocular visual odometry. The approach resolves the scale ambiguity not by reconstructing the 3D structure of the scene, but by sequentially using relative displacements between the current and several previous camera positions. This mechanism reduces reliance on long-term tracking of the same keypoints and makes the method applicable under limited parallax, repetitive textures, and partial occlusions. However, this formulation increases the method’s sensitivity to noise. To compensate for this drawback, a Kalman filter is employed, which prevents error accumulation during consecutive estimates, ensures correct operation even in the absence of information about camera motion, and allows measurement quality and reliability to be taken into account. A state model and an observation model are proposed, which allow for refining the camera pose using only data from consecutive observations. The state is represented as a sliding window comprising the current and several preceding camera poses, while the observations are constructed from directions of relative displacements estimated from an image pair or a short set of frames, without reconstructing the three-dimensional structure of the scene. Unlike common Multi-State Constraint Kalman Filters (MSCKF), which typically integrate inertial measurements, the proposed method relies exclusively on information extracted from the image sequence. The effectiveness of the proposed method is confirmed with synthetic data in various motion scenarios.
Keywords:visual odometry, Kalman filter, monocular vision, position estimation, robotics.
Fulltext:
PDF file (2122 kB)