Self-Growing Spatial Graph Network for Context-Aware Pedestrian Trajectory Prediction

This video was presented at the ICIP'21 conference at the poster session "Machine learning for image and video analysis, synthesis, and retrieval". The video walk-through the proposed approach STR-GGRNN.

The problem of predicting pedestrians' future walking trajectories has received great attention in the research community. This task makes up an essential component for the smart social robots and smart vehicles path planning system.

Across the literature, one can realize the pervasive application of Spatio-temporal graphs as a modeling structure.

Due to their flexibility, existing works have shown a great success. Besides that, there was a bit of engineering effort on how to perceive and model pedestrians' trajectories and their interactions.

The concern here is how to automate the modeling process with less tailoring. In a more specific manner, we want to find the best possible edge set that reflects the ground-truth relationships between pedestrians and minimizes the prediction errors as much as possible.