Predicting Time-Series Trajectories of Human Driven Vehicles, using an Unsupervised Attention-Based Recurrent Neural Network

Driving along public roads must always be done safely and efficiently. To avoid potential collisions experienced drivers understand and anticipate how surrounding vehicles behave, allowing them to adjust their vehicles up to a few seconds in advance. This paper presents an attention-based recurrent neural network that is capable of accurately predicting the Cartesian trajectories of multiple human driven vehicles over a 3s prediction-horizon. Using only a 1s input sequence the proposed network can predict trajectories three times this length, with an average root mean square error of just 0.497m and a variance of 0.506m. The network is trained on a dataset that contains various potentially dangerous collision scenarios, where broken down vehicles block lanes on a motorway. Learning how human driven vehicles behave in such scenarios is essential for developing autonomous vehicles. Not just for them to be trusted to on our roads, but to also help reduce the number of road collisions though improved path planning, as well as communication between autonomous vehicles. These two points forming part of the larger "Multi-Car Collision Avoidance" project that this paper contributes towards.