Real-Time Texture-based 3D Object Tracking for Advanced Robotic Manipulation

Kurzfassung

In many applications, the 6 DoF pose of an object is required. This includes roboticapplications, such as in-hand manipulation by an anthropomorphic robot and ap-plications in augmented reality to link the real-world with augmented objects. Forthese purposes, optical tracking has been used due to its low financial cost and flex-ible algorithms. However, it is still a challenge to have a versatile tracker that canprovide accurate pose estimates for objects with different properties. Specifically,applications of in-hand manipulation pose constraints on the tracker, like the com-putational efficiency and the ability to handle possible occlusions. These challengeshave motivated methods that rely on different information, such as edges or objectsilhouette. While these approaches work well for complex objects, they strugglewith objects that have non-distinct shapes and result in pose ambiguities.To minimize this problem, we develop a texture-based tracking modality that con-siders feature points on the object surface to estimate the pose. Each new poseusually depends on the pose estimation from a previous frame, leading to possibleerror accumulation. While this does not greatly impact the performance in manyscenarios, the drift error can be minimized by including a fixed reference of the objectcharacteristics. For this purpose, we additionally develop a framework to generate asparse feature map of a tracked object, which acts as a fixed reference during furthertracking instances. The complete texture-based tracking modality is then combinedwith an existing region-based tracker in a multi-modality framework. This exploitsthe advantages of both methods and results in an accurate, robust tracker. Finally,we perform a threefold evaluation on the RBOT dataset, YCB video dataset, anda self-recorded sequence to assess the performance of the tracker and compare it tothe current state of the art. The results show that the combined tracker deliverssuperior tracking performance to existing methods.