Predicting driving speed using neural networks

Predicting the speed of a vehicle for a future point on the road ahead is an important subtask of many advanced safety systems. We propose a two-stage neural net approach: first, (a small number of) characteristics of the overall speed distribution at a given location are estimated from road features alone. Second, for the case of a particular trip the speed at the current location, together with the speed characteristics output by the first stage for both the current and a future location, is used to predict the speed at the latter. Our approach parallels the previous empirical constant-percentile approach. It achieves nearly the same predictive accuracy, while at the same time reduces the data requirement to a feasible amount and additionally is able to generalize to extreme speeds not previously seen in the training set.