Predicting Occupancy Distributions of Walking Humans With Convolutional Neural Networks

As robots are increasingly entering human environments, many subtleties of socially compliant navigation are still unsolved. To behave in a socially compliant way, robots need to have an understanding of the natural motion paths of humans in the shared environment. Humans intuitively follow social norms, which allows them to navigate smoothly even in crowded environments. For example, when humans enter a previously unseen building, they are still able to infer from their surroundings where humans would typically walk and use this information to obviate interference. In this letter, we propose an approach to learn such a predictive method. A robot could use this information to find nondisturbing waiting positions, avoid crowded areas, or clean heavily frequented areas more often. We propose the use of convolutional neural networks to predict average occupancy maps of walking humans even in environments where no human trajectory data are available. In experiments, we show that our method transfers from simulation to real-world data and performs better than several baseline methods. We demonstrate the applicability on a real robot to find good waiting positions near narrow passages as well as a planner, which avoids areas where human interference is likely.