Robust object tracking with depth data
J.A. Van Egmond · Research Repository (Delft University of Technology) · 2014
Object tracking plays a vital role in the future of Robotics as it is used for motion planning and decision making. In recent years, research in object tracking has focussed on designing tracking methods that do not rely on a pre-learned model of the object and are therefor able to track any kind of object. As a result, online learning of an object model is used in virtually all new methods. However, these tracking methods often fail when presented with articulated objects or when an object rotates enough that parts become visible that haven’t been seen before. Furthermore, these methods rely purely on texture and have no notion of object shape. Additionaly, almost no existing tracking method processes images fast enough to be practically applied in Robotics applications. Affordable depth sensors have started to become popular. These sensors work as normal cameras, but with an additional depth channel. This depth information contains information about the shape of objects, which can be used to recognize the difference between object and background, or to learn a more accurate model of the object of interest. In this work, novel methods are introduced to improve tracker accuracy. Acknowledging that introducing novel methods, tracking speeds are likely to decrease, two additional methods of increasing the tracker speed are proposed. K-means clustering and local searching are applied to increase tracker speed, while depth information is used in two ways to improve accuracy. Gaussian Mixture Modelling of the depth histogram is applied to segment object from background and shape features based on local curvature are used in a Random Forests classification method to learn an object model that includes shape information.