PROJECT TITLE :

Human Trajectory Forecasting in Crowds A Deep Learning Perspective

ABSTRACT:

Since the last few decades, human trajectory forecasting has been a field of active research due to the numerous real-world applications it has. Some examples of these applications include evacuation situation analysis, deployment of Intelligent Transport systems, and traffic operations. In this line of research, we recast the challenge of human trajectory forecasting as the task of acquiring an accurate representation of human social interactions. In earlier works, this representation was painstakingly crafted by hand using domain knowledge. Nevertheless, the social interactions that take place in crowded environments are not only varied but also frequently subtle. Deep learning techniques have recently outperformed their handcrafted counterparts. This is because deep learning techniques learn about human-human interactions in a manner that is more data-driven and generic. In this work, we present an in-depth analysis of existing deep learning-based methods for modeling social interactions. This work was carried out by the authors of the aforementioned paper. We propose two data-driven methods that are inspired by knowledge specific to the domain in order to effectively capture these social interactions. We develop a large scale interaction-centric benchmark called TrajNet ++ in order to conduct an objective comparison of the performance of these interaction-based forecasting models. This benchmark is a significant yet missing component in the field of human trajectory forecasting. New performance metrics that assess a model's ability to generate socially acceptable trajectories are one of the ideas that we present in this paper. Experiments conducted on TrajNet++ provide evidence that there is a need for the metrics that we have proposed, and our method outperforms competitive baselines on both real-world and synthetic datasets.