Trajectory and behavior forecasting in autonomous driving and robotics

Artificial Intelligence algorithms are increasingly present within our society to improve people's daily lives, especially in the fields of Autonomous Driving, Virtual Reality and Industry 4.0. The forecasting of trajectories and behaviors of pedestrians, 3D poses of robots and head motion of human users is fundamental for reaching the necessary level of safety, avoiding dangerous situations in Autonomous Driving and Robotics and streaming optimization in Virtual Reality. It is not a trivial problem given the high uncertainty of what happens in the future in crowded road scenarios and in Industry 4.0 workplaces. In this thesis, different deep learning neural networks have been developed to predict the correct future states given observed past ones. Memory-augmented neural networks have been developed to generate multiple future trajectories given a single past trajectory and to understand the social interactions between pedestrians in crowded scenarios. A Transformer-based model was developed to understand the most dangerous safety behaviors of people given multi-modal inputs about pedestrians and environments. Generative models were used for the multiple generation of future head motion of users in Virtual Reality. Finally, a new paradigm was introduced that demonstrates the improvement of the current 3D pose of a robot if jointly the future poses are predicted. Experiments were carried out on datasets acquired from real situations to demonstrate the performance and explainability of the developed neural networks.