Model-Based 6D Visual Object Tracking with Impact Collision Models

Abstract

3D Object tracking is an essential technique in computer vision and has many fields of application. In this study, the focus lies on tracking objects impacting the environment. We show that state-of-the-art methods lose track of objects in this context and we investigate how to overcome this problem by adding prior information regarding the object and surface where collision is expected to occur. For illustration purposes and application relevance, we focus on the case of a box impacting a surface, which is, e.g., encountered in robot tossing in logistics applications. We model the effects of impacts and friction in a motion model, and consider the state of the box to evolve in a Lie group. We present an object tracking algorithm, based on an Unscented Particle Filter, for systems whose state lives in a Lie group and incorporate this motion model. The observations are taken from a single RGB camera and make use of the known 3D model of the object and color characteristics to predict its appearance in the 2D image. We quantitatively evaluate the effectiveness of our proposed methods by means of simulations on synthetic images.