Long-Term Prediction Of Vehicle Trajectory Using Recurrent Neural Networks

The expectations regarding autonomous vehicles are very high to transform the future mobility and ensure more road safety. Autonomous driving system should be able in the short term to detect dangerous situations and respond appropriately and thus increase driving safety. Understanding the intentions of drivers has recently received growing interest. A long-term prediction method based on gated unit-recurrent neural network model is proposed for the problem of trajectory prediction of surrounding vehicles. A deep neural network with Long-short term memory (LSTM) and Gated Recurrent Units (GRU) structure is used to analyze the spatial-temporal features of the past trajectory. Through sequences learning, the system generates the future trajectory of other traffic participants for different horizons of prediction. We evaluate all models with standard metric (Root mean square error RMSE), loss function convergence and processing time. After comparing the different models, our experiments revealed that the proposed GRU based models is indeed better than LSTM based models in term of accuracy and processing speed.