SMI-LSTM: A vehicle trajectory prediction network based on spatial motion interaction modeling

Vehicle trajectory prediction is a critical task in intelligent vehicle behavior decision-making and motion planning. It involves estimating the future position change of a vehicle using its historical trajectory and motion states. However, accurately predicting trajectories is still a challenging task due to the diversity of traffic scenes and the limitations of autonomous vehicle perception. Most existing models use a occupancy grid-based method to represent the surrounding environment and select surrounding vehicles, then relying solely on the absolute motion state in static space to represent the interaction relationship between vehicles. This method fails to consider vehicles as nodes in the graph structure to capture hidden interactions between vehicles, and ignores the relative motion characteristics of vehicles in dynamic space. Moreover, the influence of vehicles beyond visual range on object vehicle movement has not been fully considered since data acquisition mainly relies on on-board equipment. This paper proposes a LSTM network model based on spatial motion interaction (SMI-LSTM) that leverages the relative motion characteristics and the graph neural network to address these issues. Our model creates a global dynamic spatial representation based on data sampled by roadside equipment from an overhead perspective, allowing the relative motion characteristics of important vehicles to be input into the network. We also use the graph attention mechanism to capture hidden interaction relationship of vehicle interactions.