Surround Vehicle Motion Prediction Using LSTM-RNN for Motion Planning of Autonomous Vehicles at Multi-Lane Turn Intersections

Abstract

This paper presents a surround vehicle motion prediction algorithm for multi-lane turn intersections using a Long Short-Term Memory (LSTM)-based Recurrent Neural Network (RNN). The motion predictor is trained using the states of subject and surrounding vehicles, which are collected by sensors mounted on an autonomous vehicle. Data on 484 vehicle trajectories were collected from real traffic situations at multi-lane turn intersections. 11,662 and 4,998 samples acquired from the vehicle trajectories were used to train and evaluate the networks, respectively. A motion planner based on Model Predictive Control (MPC) is designed to determine the longitudinal acceleration command based on the predicted states of surrounding vehicles. The future states of the subject vehicle derived by MPC is used as an input feature to reflect the interaction of subject and target vehicles in LSTM-RNN based motion predictor. The proposed algorithm was evaluated in terms of its accuracy and its effects on the motion planning algorithm based on the driving data sets. The improved prediction accuracy substantially increased safety by bounding the prediction error within the safety margin. The application results of the proposed predictor demonstrate the improved recognition timing of the preceding vehicle and the similarity of longitudinal acceleration with drivers. INDEX TERMS Autonomous vehicle, intersection driving data, motion prediction, machine learning, recurrent neural network, long short-term memory, model predictive control.

Virtual Target-Based Overtaking Decision, Motion Planning, and Control of Autonomous Vehicles

“…There are probabilistic prediction methods that could perform better than EKF-based prediction [23]- [25], [45]. However, in this paper, the EKF-based probabilistic prediction model is adopted in consideration of vehicle implementation.…”

“…The calculation time of the EKF-based model is compared with other data-driven prediction models. These predictors were learned from the data of the perception module used in this paper [25], [45]. One is the RNN-based method, which predicts behavior through the accumulated trajectory of the target vehicle [25].…”

“…These predictors were learned from the data of the perception module used in this paper [25], [45]. One is the RNN-based method, which predicts behavior through the accumulated trajectory of the target vehicle [25]. The other is a predictor, which combines RNN and EKF methods to reduce the computational load [45].…”

“…Because deterministic prediction methods have limits, probabilistic prediction methods have been developed to augment robustness [21], [22]. Also, advanced probabilistic prediction algorithms have been devised utilizing data-driven approaches [23]- [25]. Inverse reinforcement learning has been employed to predict interactive motions considering discrete and continuous driving decisions [23].…”

“…The multi-modal probabilistic model could consider behavior intention based on deep neural networks [24]. Long short-term memory-based recurrent neural networks have been utilized for interactive prediction in multi-lane turn intersections [25]. Data-driven probabilistic prediction models considering interactive behavior are powerful in complex situations, such as ramp-merging, roundabout, and multi-lane turn intersections [23]- [25].…”