An Extension of GHMMs for Environments With Occlusions and Automatic Goal Discovery for Person Trajectory Prediction

Abstract—Robots navigating in a social way should use some knowledge about common motion patterns of people in the environment. Moreover, it is known that people move intending to reach certain points of interest, and machine learning techniques have been widely used for acquiring this knowledge by observation. Learning algorithms such as Grow-ing Hidden Markov Models (GHMMs) usually assume that points of interest are located at the end of human trajectories, but complete trajectories cannot always be observed by a mobile robot due to occlusions and people going out of sensor range. This paper extends GHMMs to deal with partial observed trajectories where people’s goals are not known a priori. A novel technique based on hypothesis testing is also used to discover the points of interest (goals) in the environment. The approach is validated by predicting people’s motion in three different datasets. I