Improved Visual Tracking using Motion Models and Physical Constraints

Abstract

Object tracking is an important prerequisite for the interaction of robotic systems with their environment. Visual tracking frameworks provide a possibility to obtain a 6 Degrees of Freedom (DOF) pose estimate for objects of interest. However, the provided visual information becomes less expressive in the presence of ambiguities, occlusions, or motion blur. This poses a major limitation for tracking approaches, since it can lead to inaccuracies in the pose estimation or in the worst case to complete tracking loss. However, apart from the visual information, we have more knowledge about the scenes at our disposal. In this work, we aim to incorporate scene information into the tracking process without additional sensor data. For this, we introduce two extensions to the multi-body, multi-modality, multi-camera tracking framework M3T. First, we propose incorporating prior knowledge of the object's pose through motion models. With this, the tracker's accuracy improves because of an increased robustness to ambiguities and partial occlusions. Moreover, we include collision constraints into the optimization process in order to consider physical properties of rigid bodies which do not allow object intersection. With this check for plausibility, the pose estimation avoids impossible configurations and influences the tracker's convergence. We demonstrate the changed tracking behavior of our extended framework on the Robot Tracking Benchmark (RTB) dataset and in the development environment of the robotic surgery system MiroSurge at the German Aerospace Center (DLR).