Efficient Trajectory Prediction for Automated Driving with a flexible Deep Learning Architecture

Highly automated vehicles are expected to be widespread in the near future. However, as their availability increases, technical challenges persist, particularly in urban environments where predicting the behavior of surrounding traffic participants is critical for safe motion planning. This paper presents a novel motion forecasting architecture that generates marginal trajectory predictions for multiple traffic agents over a six-second horizon. The architecture utilizes different encoders and decoders that employ concepts from Wayformer, Scene Transformer, and Forecast MAE. The framework generates six potential trajectories per agent with marginal probabilities. The scene representations are created from agent tracks, lane features from HD maps, and agent interactions. Trained on the largescale Argoverse 2 Motion Forecasting dataset, the system reduces the input sample rate to enhance efficiency with minimal loss of accuracy compared to similar approaches. The system is evaluated with the Argoverse benchmarks, demonstrating its capability for fast training and inference while producing highly accurate trajectories. The low inference time facilitates the potential for real-world applications.