Will the Pedestrian Cross? Probabilistic Path Prediction Based on Learned Motion Features

Future vehicle systems for active pedestrian safety will not only require a high recognition performance, but also an accurate analysis of the developing traffic situation. In this paper, we present a system for pedestrian action classification (walking vs. stopping) and path prediction at short, sub-second time intervals. Apart from the use of positional cues, obtained by a pedestrian detector, we extract motion features from dense optical flow. These augmented features are used in a probabilistic trajectory matching and filtering framework. The vehicle-based system was tested in various traffic scenes. We compare its performance to that of a state-of-the-art IMM Kalman filter (IMM-KF), and for the action classification task, to that of human observers, as well. Results show that human performance is best, followed by that of the proposed system, which outperforms the IMM-KF and the simpler system variants.