Vehicle Trajectory Prediction on NGSIM Using Kalman Filter and LSTM

Where will the car next to us be in five seconds? This project predicts the path of a vehicle from its last 3 seconds on real freeway trajectories (NGSIM US-101, video-recorded at 10 Hz) and compares classical physics predictors - constant velocity and Kalman filters - with an LSTM encoder-decoder and a social-pooling LSTM that also sees the six neighbouring vehicles. Training and test vehicles are kept apart, lane-keeping and lane-changing cars are scored separately, and the models are also tried on another freeway (I-80) and on simulated SUMO traffic. On unseen US-101 vehicles the 5-s position error (RMSE) is 7.47 m for constant velocity, 6.55 m for the LSTM and 4.62 m for the social LSTM; the specification's constant-acceleration Kalman filter is worse than plain constant velocity (11.34 m), and on the other roads the plain LSTM loses to constant velocity. A web application shows every prediction next to the truth and runs a time-gap safety check. A teaching prototype, not a driver-assistance function.

Category: Automobile Engineering

Automobile Engineering Project AU08 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.