Feature Trajectory Retrieval with Application to Accurate Structure and Motion Recovery

Common techniques in structure from motion do not explicitly handle foreground occlusions and disocclusions, leading to several trajectories of a single 3D point. Hence, different discontinued trajectories induce a set of (more inaccurate) 3D points instead of a single 3D point, so that it is highly desirable to enforce long continuous trajectories which automatically bridge occlusions after a re-identification step. The solution proposed in this paper is to connect features in the current image to trajectories which discontinued earlier during the tracking. This is done using a correspondence analysis which is designed for wide baselines and an outlier elimination strategy using the epipolar geometry. The reference to the 3D object points can be used as a new constraint in the bundle adjustment. The feature localization is done using the SIFT detector extended by a Gaussian approximation of the gradient image signal. This technique provides the robustness of SIFT coupled with increased localization accuracy. Our results show that the reconstruction can be drastically improved and the drift is reduced, especially in sequences with occlusions resulting from foreground objects. In scenarios with large occlusions, the new approach leads to reliable and accurate results while a standard reference method fails.