Neural Network Aided Potential Field Approach For Pedestrian Prediction

Autonomous driving is one of the key challenges in recent time. As pedestrians are the most vulnerable traffic participants, collisions with pedestrians have to be avoided under all circumstances. Hence, prediction of pedestrian trajectories is of high interest for automated vehicles. For this purpose, a plethora of algorithms has been proposed to model the pedestrian in the last decades, reaching from simple kinematic models to advanced microscopic models. In addition, the machine learning community started to learn the behavior of pedestrians and showed major improvements in complex scenarios or unexpected situations. However, as most of the machine learning algorithms are treated as black boxes, the safeguarding of the software is one key challenge which has to be solved. This contribution proposes to combine classic modeling of pedestrians with machine learning algorithms by learning the model errors between a simple physical model and real data. In particular, it is proposed to combine a physical model based on potential fields with a neural network to predict the future behavior of pedestrians. It is shown that the combined approach outperforms the physical model in learnable areas, whereas the physical model without the neural network is more robust in areas where almost no training data is available. In addition, different structures of neural networks are analyzed.