Trajectory-Guided Driving Behavior Prediction for Autonomous Driving

Driving behavior prediction (DBP) plays an important role in autonomous driving systems. Accurately anticipating driving behaviors of ego-vehicles, such as lane changes or left/right turns, several seconds before their actual occurrences enables AI pilots to plan safer trajectories or alert drivers to potential dangers. However, existing DBP benchmarks, e.g., the Brain4Cars, are restricted by the limited amount of data samples and the narrow range of categories for driving behaviors of ego-vehicles. In this study, we extract over 5000 clips from untrimmed multi-modal data sequences of the HRI Driving Dataset (HDD) to create a new large-scale DBP benchmark, named HDD4DBP. Furthermore, a Trajectory-Guided Driving Behavior Prediction (TGDBP) approach is proposed within a two-branch network architecture. One branch encodes visual information of driving scenarios using a vision transformer (ViT) backbone, while the other predicts the future trajectory of the ego-vehicle based on historical trajectory and control signals. Finally, the predicted trajectories and visual information are fused to predict the driving behaviors. In experiments, the proposed TGDBP achieves a significant performance improvement over previous baselines on the HDD4DBP. Moreover, fine-tuning the pre-trained visual encoder of the TGDBP model on the Brain4Cars dataset results in state-of-the-art (SOTA) performance, further showcasing the advantages of the HDD4DBP dataset.