Transformer-Based Trajectory Prediction Using LiDAR Data for Situational Awareness in Complex Urban EnvironmentsPublished

With the rise of intelligent systems in urban transportation, the ability to predict agent behavior in real time has gained increasing research attention. Accurate trajectory prediction plays an important role in improving safety and decision-making in self-driving vehicles and smart city infrastructure. This study focuses on LiDAR-sensor-based trajectory prediction of agents at a hyperlocal level using a Transformer architecture. A large-scale dataset was collected using an Ouster OS1 LiDAR sensor at a busy urban intersection in West Palm Beach, Florida. This experiment captured more than 12,390 real-world trajectories which include vehicles, pedestrians, and bicycles.

After obtaining experimental results from the sensor, the proposed framework first performs object detection to extract agent trajectories from LiDAR point-cloud data. Afterwards, data curation was performed to filter out the reflections of pedestrians and vehicles on the glass storefronts, or they were almost stationary. In the next stage, a Transformer model is developed to learn and predict spatial-temporal patterns of agent trajectories.

By performing a hyperparameter tuning, the Transformer model was able to achieve a 15.24% improvement in the average displacement error in comparison with the traditional LSTM method. Results are visualized to display predicted and ground-truth paths on a geo-referenced map. With a higher convergence rate compared to the LSTM approach, the proposed results showed the effectiveness of attention-based models in complex multi-agent urban environments.