Automated Learning of Pedestrian Walking Speed Profiles for Improved Movement Prediction

Abstract

Every year, about 310,500 pedestrians still lose their lives in traffic accidents worldwide. Cooperative pedestrian collision avoidance represents a promising approach to reduce those accident numbers. This approach assumes that pedestrians are equipped with mobile devices to obtain and exchange their current movement information with nearby vehicles and use those to predict and prevent possible collisions. However, the ability to predict collisions between a pedestrian and a vehicle also depends on the assumptions about the pedestrian’s future behavior. One important aspect of those assumptions is a pedestrian’s individual walking pattern, like his common or maximum speed. Thus, learning and applying individual walking speed profiles of pedestrians to improve movement prediction may increase the accuracy of a collision detection algorithm and could, in turn, reduce the probability of missing or erroneously triggering an alarm. In this publication, we propose an approach to learn individual walking speed profiles of a pedestrian based on smartphone Global Navigation Satellite System (GNSS) data and evaluate the ability to predict collisions based on those profiles. Therefore, we first conducted experiments to estimate the error of walking speed obtained from smartphone GNSS. Second, using our Pedestrian Monitor application, we recorded real-world walking speed information from nine participants. Based on these data, we show that individually learned walking speed profiles are able to increase the accuracy of predicting an impending collision.