A transferable model for pedestrian intention prediction in crosswalks based on mobile phone sensors.

To accommodate the path planning of autonomous vehicle (AV), many researchers utilize on-board sensors to predict pedestrians’ intention, but the on-board sensors cannot detect occluded pedestrians. Roadside sensors can solve this problem, but with a high cost if deployed at every crosswalk. Hence, this paper aims at utilizing mobile phone sensors to predict the pedestrians’ intentions at crosswalks. A transferable prediction model is built in the paper, with a purpose of saving the expense of hardware and data collection. That is, the model trained by the pedestrians’ data at one crosswalk can provide high accuracy at other crosswalks, via extracting the features associating the geometric structure of the crosswalks and the pedestrians’ motion data. Various algorithms were applied to build the prediction model, and results showed that the prediction accuracy exceeds 90% in non-transfer task, 80% in normal transfer tasks, with a distance of 5 meters ahead of crossing.