Tracking humans interacting with the environment using efficient hierarchical sampling and layered observation models

We present a markerless tracking system for unconstrained human motions which are typical for everyday manipulation tasks. Our system is capable of tracking a high-dimensional human model (51 DOF) without constricting the type of motion and the need for training sequences. The system reliably tracks humans that frequently interact with the environment, that manipulate objects, and that can be partially occluded by the environment. We describe and discuss two key components that substantially contribute to the accuracy and reliability of the system. First, a sophisticated hierarchical sampling strategy for recursive Bayesian estimation that combines partitioning with annealing strategies to enable efficient search in the presence of many local maxima. Second, a simple yet effective appearance model that allows for the combination of shape and appearance masks to implicitly deal with two cases of environmental occlusions by (1) subtracting dynamic non-human objects from the region of interest and (2) modeling objects (e.g. tables) that both occlude and can be occluded by human subjects. The appearance model is based on bit representations that makes our algorithm well suited for implementation on highly parallel hardware such as commodity GPUs. Extensive evaluations on the HumanEva2 benchmarks show the potential of our method when compared to state-of-the-art Bayesian techniques. Besides the HumanEva2 benchmarks, we present results on more challenging sequences, including table setting tasks in a kitchen environment and persons getting into and out of a car mock-up.