Vehicle Speed Prediction with RNN and Attention Model under Multiple Scenarios

Abstract

Motion prediction is an essential feature for autonomous vehicle to understand the intention of nearby vehicles so as to plan its route. Several works have been proposed to predict the motion of nearby vehicles using Kalman filter, RNN, and other machine learning methods. However, many of them rely on either the communication among vehicles or the global information of all the vehicles on the road and have limited applicability. This paper presents the design and implementation of a new machine learning model to predicate the motion of nearby vehicles by observing their motions in last few seconds. The proposed network consists of encoder- decoder, LSTM, and attention model to tackle the challenges so as to predicted the vehicle speed using limited observations. The network was trained based on data set collected on public roads. Comparing with Kalman filter, the developed method reduces the prediction error up to 50% and the prediction error is up to 6.5KPH(1.8 meter per second) under all evaluated scenarios, which is less than the tolerance of speedometers on vehicles.