Seite - 98 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

Inspired by [5] we describe a fast and accurate 3D object tracking algorithm for rigid objects. The proposed approach is model-based, uses only depth data and achieves very good accuracy utilizing a framework that combines object localization and object tracking. 2. State of the Art With the introduction of RGB-D sensors like Kinect, various approaches for object tracking in 3D were proposed, which ranged from tracking humans [10], hand tracking [8], and tracking rigid and non-rigid objects. For a better comparison of our approach with the state of the art, the scope of this section is limited to approaches that focus on frame-to-frame tracking of rigid objects using RGB-D data. The proposed approaches can be broadly classified into two categories: a) approaches that are based on 3D models and b) approaches that do not assume pre-defined object models. For example, an approach that does not rely on prior knowledge of the target object representation is described in [14]. The approach uses adaptive Gaussian Mixture Models (GMM) to represent multiple objects that move independently. The object model is updated incrementally at each time instant with the help of the feedback results from the robust tracking process. To correct falsely detected objects in presence of occlusions and various types of interactions among multiple objects, an approach that exploits component-level spatiotemporal association is proposed in [10]. However, the approaches of “individuation-by-feature” [14] and “individuation-by-location” [10] require high computation time to learn each object model at every time instant and would exponentially increase with the number of objects and their spatial relations. Moreover, in an industrial environment which involves human actions, situations keep changing every time instant. To achieve robustness and computational efficiency in such scenarios, applying one individuation method is not sufficient. To alleviate this problem, an approach that determines individuation strategy (by location and/or by feature) depending on the object situation is proposed in [15]. The main assumption of the approach is that falsely segmented objects can be detected and rectified using both location and position information. It also assumes that objects do not change substantially in terms of shape or position from one frame to the other. A probabilistic framework for simultaneous tracking and reconstruction of rigid 3D objects using RGB-D sensor is proposed by [7], where the probabilistic method is used to statistically determine occlusions. Intensity images are used to model appearance of an object while modeling occlusions. With the availability of reliable, fast and simple object reconstruction solutions like ReconstructMe1, 3D object models can be obtained in real-time. A popular approach for model- based object tracking is based on the particle filters [10][18]. For example, the authors in [4] propose a 3D model-based visual tracking approach using edge and keypoint features in a particle filtering framework. This approach does not assume the initial pose of the object. It uses given 2D- 3D keypoint correspondences to calculate a set of possible pose hypothesis of the object. Once the intial pose is estimated, edge points are used to track movement of the object from frame-to-frame. This approach is extended by [5] where an RGB-D object tracking method using a particle filter on GPU is proposed. Another popular method for 3D object tracking is the Iterative Closest Point (ICP) approach which has many variants [16]. The algorithm uses a set of initial parameters and refines them iteratively to reach a set of optimal parameters by minimizing the object function. This approach has problems in dealing with occlusions and object clutter, which result in a local-minimum. To overcome this problem, a model-based learning approach is proposed in [18]. This approach learns the relation between the parameters that induce object’ motion and the change they induce on the 3D point cloud using random forests. In order to track the object in motion, the change in the 3D depth data 1 ReconstructMe http://reconstructme.net 98