Environment-aware Pedestrian Trajectory Prediction for Autonomous Driving

People’s safety is a primary concern in autonomous driving. There exist efficient methods for identifying static obstacles. However, the prediction of future trajectories of moving elements, such as pedestrians crossing a street, is a much more challenging problem. A promising direction of research is the use of machine learning algorithms with location bias maps. Our goal was to further explore this idea by training an interchangeable location bias map, a location-specific feature that is added into the middle of a convolutional neural network. For different locations, we used different location bias maps to allow the network to learn from different setting contexts without overfitting to a specific setting. Using pre-annotated video footage of pedestrians moving around in crowded areas, we implemented a pedestrian behavior encoding scheme to generate input and output volumes for the neural network. Using this encoding scheme, we trained our neural network and interchangeable location bias map. Our research demonstrates that the network with an interchangeable location bias map can predict realistic pedestrian trajectories even when trained simultaneously in multiple settings.