Trajectory prediction method based on a graph model for autonomous driving

Abstract

A method for predicting the trajectory of an unmanned vehicle using graphs and a network of long short-term memory (LSTM) is developed. The learning model of LSTM network uses an encoder and decoder structure. Based on the graph and the attention mechanism, the encoder encodes information about the received trajectory to form a feature vector, which is puted to the decoder to predict future trajectories. To cope with multimodality in predicting vehicle maneuvers, module of convolutional network (CNN) is used. These two networks: LSTM and CNN are integrated for multimodal trajectory prediction. Keywords unmanned vehicle, trajectory prediction, maneuver, graph, LSTM, CNN, attention mechanism