Human Motion trajectory for robot navigation

Our lives are becoming increasingly influenced by robots. They are no longer limited to working in factories and increasingly appear in shared spaces with humans, to deliver goods and parcels, ferry medications, or give company to elderly people. Therefore, they need to perceive, analyze, and predict the behavior of surrounding people and take collision-free and socially-acceptable actions.

 In this thesis, we address the problem of (short-term) human trajectory prediction, to enable mobile robots, such as Pepper, to navigate crowded environments.

We propose a novel socially-aware approach for prediction of multiple pedestrians. Our model is designed and trained based on Generative Adversarial Networks, which learn the multi-modal distribution of plausible predictions for each pedestrian.

Additionally, we use a modified version of this model to perform data-driven crowd simulation. Predicting the location of occluded pedestrians is another problem discussed in this dissertation. Also, we carried out a study on common human trajectory datasets. A list of quantitative metrics is suggested to assess prediction complexity in those datasets.