Segmentation-Free, Area-Based Articulated Object Tracking

We propose a novel, model-based approach for articulated object detection and pose estimation that does not need any low-level feature extraction or foreground segmentation and thus eliminates this error-prone step. Our approach works directly on the input color image and is based on a new kind of divergence of the color distribution between an object hypothesis and its background. Consequently, we get a color distribution of the target object for free. We further propose a coarse-to-fine and hierarchical algorithm for fast object localization and pose estimation. Our approach works significantly better than segmentation-based approaches in cases where the segmentation is noisy or fails, e.g. scenes with skin-colored backgrounds or bad illumination that distorts the skin color. We also present results by applying our novel approach to markerless