Road User Position Prediction in Urban Environments via Locally Weighted Learning
This paper focuses on the problem of predicting the future position of a target road user given its current state, consisting of position and velocity. A weighted average approach is adopted, where the weights are determined from data containing the state trajectories of previously observed road users. In particular, a similarity function is introduced to extract from data those previously observed road users' states that are most similar to the target's one. This formulation results in an easily interpretable model with few parameters to calibrate. The performance of this weighted average model(WAM) is evaluated on the same real-world data as state-of-the-art methods, showing promising results. WAM outperforms the baseline constant velocity model at longer prediction horizons, making WAM suitable for motion planning applications. WAM and a baseline neural network model performs comparably. Still, WAM has only three parameters which are easily interpretable, while the complex neural network model has thousands of parameters which are difficult to analyze.
READ FULL TEXT