Lina Achaji

Autonomous Vehicles, Deep Learning

Working on the pedestrian action anticipation problem – usage of image-based raw data, pedestrian skeleton sequences, and bounding boxes temporal time-series with Transformer-based networks.

Working on the multi-agent pedestrian trajectory prediction problem – focusing on the interaction modeling, and on the parallel trajectory generation techniques. Proposition of the PRETR model, that achieved SOTA accuracy results and 11x faster trajectory generation.

Working on the Exo-vehicle trajectory prediction problem – focusing on the goal-based multi-modal trajectory prediction, and enhancing the interaction modeling between vehicles for vector-based models.

Supervising intern students working on pedestrian trajectory prediction, and interaction modeling between physical agents.

Research on SOTA models for computer vision and prediction techniques: object detection, segmentation, Transformer-based networks, Inverse-RL, SSL, etc.

Research on Machine learning for graphs: Node embeddings, Graph Neural Networks, Scene, Knowledge graphs, etc.

Development of fault-tolerance techniques by using Extended and Informational Kalman Filters for person indoor localization.

Multi-person tracking using global nearest neighbors methods.