Multi-Model and Learning-Based Framework for Real-Time Trajectory Prediction

Accurate and real-time trajectory prediction of traffic participants is important in autonomous driving systems, especially for decision making and risk assessment. Existing models such as physics-based and maneuver-based models are mainly used for short-term prediction. Deep-learning-based methods have been applied as novel alternatives for trajectory prediction. This problem can be viewed as a sequence generation task, where the future trajectory of vehicles is predicted based on their past positions. Following the recent success of Recurrent Neural Network (RNN) models for sequence prediction tasks, especially Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), in this paper an approach that combines LSTM for driving sequences classification and GRU for trajectory prediction is proposed. The obtained experimental results show the effectiveness of the proposed approach.