Predicting Pedestrian Intentions: Synthetic Data and Efficient Models

Navigating streets alongside unpredictable pedestrians poses a challenge for autonomous vehicles. Accurately predicting their crossing intentions is crucial for safe and smooth operation.

At the 2023 IEEE Intelligent Transportation Systems Conference (ITSC), a paper describing PedSynth, a large-scale synthetic dataset specifically designed for pedestrian intention prediction, was presented. It also described the PedGNN, a highly efficient deep-learning model utilizing a GNN-GRU architecture. This model analyzes sequences of pedestrian skeletons extracted from video data to accurately anticipate their crossing intentions.

I participated in the creation of the CARLA-based data generation framework and co-mentored Naveed during his work on this publication.