Cyclist Motion State Forecasting - Going beyond Detection

Maarten Bieshaar, Stefan Zernetsch, Katharina Riepe, Konrad Doll, Bernhard Sick

Abstract: In this article, we present two novel methods to forecast the motion states of cyclists. The states we aim to anticipate are waiting, starting, moving, and stopping. This information can be utilized to increase road safety when used in an automated vehicle. We classify the cyclist motion state for every step in a discrete-time horizon using a single neural network in the first method. In our second approach, we consider a two-stage model, i.e., a neural network predicts the current and the next motion state, and a second quantile regression neural network (Q RNN) forecasts the time to transition between these two motion states. Our results show that both methods have advantages and disadvantages. The first method can forecast multiple changes in motion state while the second is restricted to a single transition. However, the first method is limited to a fixed forecast horizon. The two-stage approach, which forecasts motion and time separately, is more flexible regarding the forecast horizon, i.e., it can forecast very long as well as short time spans. Regarding the transition detection performance, both methods perform equally well. Our experiments show that the time to transition to the next motion state can be forecasted accurately, especially for short-time horizons.