UC San Diego

As a part of developing autonomous vehicles and better Advanced driver assistance systems(ADAS), it is important to consider how the spatio-temporal activities of other agents in the environment like pedestrians, vehicles, etc. which are competing for space on roads might impact the motion planning performance of the vehicle . A system which can predict future obstacle trajectories as well as warn the driver or the autonomous vehicle about an impending collision will lead to safer roads and save lives.

Previous vehicle trajectory prediction approaches use motion models which have assump- tions like constant velocity or constant acceleration which doesn’t generalize well. Our approach is completely data driven and gives promising results for predicting trajectory of the obstacle up to 2 seconds in the future using a deep recurrent neural network.

Taking inspiration from the recent success of sequence-to-sequence models in language translation we apply sequence-to-sequence recurrent neural networks to the new problem of trajectory prediction. The proposed scheme feeds the sequence of obstacles’ past trajectory data obtained from sensors like LIDAR and GPS to the LSTM and predicts the position of the obstacle at future time steps. We use the KITTI dataset which provides us with annotated trajectory data for learning and evaluation.