Improved multi-level pedestrian behavior prediction based on matching with classified motion patterns

This paper proposes an improved multi-level pedestrian behavior prediction method based on our previous research work on learning pedestrian motion patterns and predicting pedestrian long-term behaviors as their motion instances are being observed. The improvement mainly focuses on the similarity matching criteria between the trajectory and the clustered MP whose main advantages are that (1) a reasonable similarity range of MP is automatically calculated instead of manually set; (2) the distance feature and the changing angle feature are considered together for similarity matching while only the distance feature is considered before. The improved method has been implemented and a study of how the new prediction method performs in real world scenario is conducted. The results show that it works well in real DCE and the prediction is consistent with the actual behavior.