Probabilistic Lane Change Prediction using Gaussian Process Neural Networks

Finding patterns in the motion of surrounding vehicles is one essential challenge to master automated driving. It is necessary to estimate their future behavior to plan a safe and comfortable trajectory for the ego vehicle. Several approaches exist for this prediction task, where a large part of those methods use machine learning techniques. Due to the continuous motion of vehicles especially during lane changes, any kind of discrete prediction introduces the danger of a significant Bayes error, especially in the transition area from lane keeping to lane change. Therefore, this paper proposes to combine neural networks with Gaussian processes. While the neural network part works as an encoder to extract information from the input data and encodes it in a lower-dimensional space, the final Gaussian process part takes this information and treats it in a probabilistic fashion. The main advantage is that in this way not only a mean value is predicted that represents the lane change probability but a distribution of the prediction consisting of the mean and the standard deviation. So the future development of the traffic situation can be described in a more complex way, considering uncertainties in the input data which lead to an improved prediction performance. Therefore, in situations with an uncertain prediction a special maneuver could be performed by the ego vehicle to reduce the chance of an incident.