Abstract:Traffic state perception is key to controlling autonomous vehicles, and predicting the future trajectories of agents in the context is critical in perception of traffic states. However, agents are not separated but interacting with other agents, as well as the surrounding physical scene. Currently, the best models are based on graph neural network, but prior works normally include only position and timestep without other available and valuable information. Our model incorporates more information, conforming to evolving V2I (Vehicle-to-Infrastructure) technology. We have proposed a novel relative-position-wise attention mechanism and constructed the graph not only based on distance but on yaw angle and velocity. In the main dataset used in the trajectory predictions, our model has achieved state-of-the-art performance. We have contributed a simulated dataset named Shougang Dataset upon the map of Beijing Shougang Industrial Park. We have applied our model to Shougang Dataset and obtained an excellent performance in trajectory prediction.

Visual-Angle Attention Predictor: A Multi-Agent Trajectory Predictor Based on Variational Auto-Encoder