Road User Position Prediction in Urban Environments via Locally Weighted Learning

Prediction of the future position of a target road user given its current position, velocity and type is formulated as a weighted average. Weights are determined from data of previously observed road users, specifically from those that are most similar to the target. This formulation results in an interpretable model with few parameters. The model is validated on a dataset of vehicles, bicycles, and pedestrians in real-world traffic situations. The model outperforms the baseline constant velocity model, wheras a baseline neural network model shows comparable performance, but this model lacks the same level of interpretability. A comparison is made with state-of-the-arts, where these show superior performance on a sparse dataset, for which it is expected that the weighted average model works less well. With some further refinements a weighted average formulation could yield a reliable and interpretable model, in constrast to one which is difficult to analyze and has a huge number of uninterpretable parameters.