Deep neural networks for Markovian interactive scene prediction in highway scenarios

In this paper, we compare different deep neural network approaches for motion prediction within a highway entrance scenario. The focus of our work lies on models that operate on limited history of data in order to fulfill the Markov property 1 and be usable within an integrated prediction and motion planning framework for automated vehicles. We examine different model structures and feature combinations in order to find a model with a good tradeoff between accuracy and computational performance. We evaluate all models with standard metrics like the negative log-likelihood (NLL) and evaluate the performance of each model within a closed-loop simulation. We find a neural network only operating on spatial features of the current state to have the best closed-loop prediction performance, despite the NLL suggesting otherwise.