For robots to execute their navigation tasks both fast and safely in the presence of humans, it is necessary to make predictions about the route those humans intend to follow. Within this work, a model-based method is proposed that relates human motion behavior perceived from RGBD input to the constraints imposed by the environment by considering typical human routing alternatives. Multiple hypotheses about routing options of a human towards local semantic goal locations are created and validated, including explicit collision avoidance routes. It is demonstrated, with real-time, real-life experiments, that a coarse discretization based on the semantics of the environment suffices to make a proper distinction between a person going, for example, to the left or the right on an intersection. As such, a scalable and explainable solution is presented, which is suitable for incorporation within navigation algorithms.
Improving Word Recognition using Multiple Hypotheses and Deep Embeddings
