Neural Network-based Prediction of Pedestrian Crossing Behavior at Unsignalized Crosswalks

This paper proposes a neural network prediction model related to pedestrian crossing decisions, which is aimed to be employed within autonomous vehicle safe speed control strategies. The model is of stochastic nature in order to capture the inherent uncertainty and variability typically present in real pedestrian behavior. Instead of predicting the whole pedestrian crossing trajectory, only ego-vehicle control-related quantities are targeted for prediction, which include the pedestrian entry and exit time to/from the crossing area. To this end, two independent feedforward neural network submodels are designed for prediction of conditional probability distributions of the entry and exit time variables in dependence of different vehicle- and pedestrian-related inputs. Finally, the proposed model is parameterized and verified for a single-vehicle/single-pedestrian interaction case, based on the data generated from numerous simulations of an available game theory-based pedestrian model, where the vehicle is driven in an open-loop manner.