Pedestrian Trajectory Prediction with Deep Learning Transformers and Kalman Filters

In recent years, transform neural networks have shown great promise in many machine learning applications. This project explores the application of these networks on a challenging problem: pedestrian trajectory prediction, which is prevalent in many applications including the autonomous navigation of vehicles. A new prediction algorithm is developed, using pedestrian pose information to generate two predictions, one via a Kalman filter, to benefit from its computationally efficient model-based estimation and the other with transformers, to leverage their advanced pattern recognition. By fusing the predictions from both methods, the algorithm generates the future trajectory of pedestrians. The proposed algorithm demonstrates its effectiveness through a Python-implemented ensemble model that predicts pedestrian paths with reduced data reliance and computational demand. The results of this integration are promising, exhibiting competitive accuracy compared to cutting-edge deep learning models, despite being trained on a considerably smaller dataset. This research lays the foundation for more efficient data-driven systems in scenarios where real-time processing and adaptability are crucial.