Early Prediction of a Pedestrian's Trajectory at Intersections

This paper focuses on the early prediction of a pedestrian's short time trajectory in the course of gait initiation at a crosswalk. We present a comprehensive study on trajectories of adults measured at a public urban intersection using 3D triangulation of stationary video-based marker- and head-detection data. Based on the results of this study we propose two models, a piecewise linear model and a sigmoid model, for predicting the trajectory starting at heel-off. At heel-off, the mean prediction error of the absolute walking distance in real intersection scenarios increases from 12 em to 26 em for prediction times of 600 ms to 2.4 s, respectively, for the piecewise linear model. The results provide a basis for vehicle based collision risk estimation resulting in a possible warning of the driver, autonomous emergency braking or evasive steering.