Intention-Aware Trajectory Prediction with Class-Balanced Learning for Highway Scenarios

In highway autonomous driving, data-driven trajectory prediction models suffer from Long-tailed Distributions, where straight-driving samples (>90%) dominate the expected gradient (termed Gradient Dominance), suppressing the learning of rare but critical intentions like lane changes. This leads to Intention Collapse, where models default to conservative straight trajectories. We propose an Intention-Aware Class-Balanced Framework to resolve this. Our approach introduces an Intention-Guided Distribution Rebalancing strategy using inverse-frequency weighting to break the gradient dominance, and an Intention-Conditioned Recurrent Decoder that maps discrete intentions to a continuous latent space for controllable generation. Experiments on the HighD dataset show our method reduces the Average Displacement Error (ADE) in safety-critical lane-changing scenarios by 21% (1.15m → 0.91m) compared to the Standard Encoder-Decoder, demonstrating superior robustness in tail events, and validating the efficacy of class-balancing in regression tasks.