An enhanced motion planning approach by integrating driving heterogeneity and long-term trajectory prediction for automated driving systems

Ni Dong, Shuming Chen, Yina Wu, Yiheng Feng, Xiaobo LiuarXiv 2023 · OpenReview

TL;DR

This paper proposes an enhanced motion planning system for autonomous vehicles (AVs) that accounts for the diverse driving styles of human-driven vehicles (HDVs) and predicts their long-term trajectories. By integrating these predictions, the AV can make safer and more proactive decisions, particularly in complex scenarios like highway merging.

What Problem Does This Paper Address?

Navigating complex driving environments is a major hurdle for autonomous driving systems (ADSs). A key challenge is the uncertainty introduced by surrounding human-driven vehicles (HDVs). ADSs need to accurately predict the future behavior of these HDVs to plan safe and efficient paths. Existing approaches often struggle with the heterogeneity of human driving styles (e.g., aggressive vs. normal drivers) and have limitations in predicting trajectories over longer horizons, which is crucial for proactive decision-making. This paper aims to bridge this gap by developing a system that explicitly models driving heterogeneity and provides accurate long-term trajectory predictions for improved ADS motion planning.

Methods

The proposed approach is hierarchical, with two main components:

  • Driving Style Prediction:
    • An unsupervised clustering algorithm (K-means) is used to group HDVs into "aggressive" and "normal" driving styles based on characteristics like merging duration and distance/speed differences with surrounding vehicles.
  • A logistic regression model then predicts the driving style of an HDV using features derived from the clustering, such as the rate of change in lateral position deviation.
  • Long-Term Trajectory Prediction:
    • A transformer-based model is employed for predicting the future trajectories of HDVs. Transformers are chosen for their ability to capture complex sequential dependencies using self-attention mechanisms, making them suitable for long-sequence predictions.
  • A sliding window approach is used to feed historical trajectory data into the transformer model for continuous prediction.
  • Enhanced Motion Planning:
    • The predicted driving styles and trajectories of surrounding HDVs are integrated into an intelligent decision-making strategy for the ADS.
  • This strategy utilizes four driving states: cruising, following, real-time avoidance (RTA), and avoidance in advance (AIA). The ADS dynamically adjusts its state based on the predicted HDV behavior and its own position relative to the HDV.
  • The time-to-collision (TTC) metric is used to evaluate the safety performance of different prediction horizons and to determine the optimal prediction horizon for planning.

Datasets

  • Real-world traffic data: Collected from drones on a highway ramp in Xi'an, China. This dataset was used for training and evaluating the proposed models. The data includes vehicle position (latitude, longitude), speed, acceleration, and lateral position deviation.

Key Insights and Ideas

  • Driving Style Matters: Explicitly modeling and predicting the driving style (aggressive vs. normal) of HDVs significantly improves trajectory prediction accuracy. Aggressive drivers exhibit more erratic behavior, which needs to be accounted for.
  • Long-Term Prediction is Crucial for Proactivity: Predicting trajectories up to 30 seconds in advance, as achieved by the transformer model, allows the ADS to move beyond reactive maneuvers and adopt proactive strategies like "avoidance in advance."
  • Transformers Excel at Long Sequences: The transformer architecture, with its self-attention mechanism, proves effective for capturing long-range dependencies in vehicle trajectories, outperforming traditional RNN/LSTM models in prediction horizon.
  • Hierarchical Approach Synergizes: Combining driving style prediction with trajectory prediction in a hierarchical manner leads to a more robust and safer motion planning system.

Key Results

  • Trajectory Prediction Accuracy: The proposed approach achieved a mean squared error (MSE) smaller than 0.05 at thirty seconds ahead of the merging point, outperforming existing methods.
  • Optimal Prediction Windows: For normal driving, an optimal time window of 6.67 seconds yielded an MSE of 0.0314. For aggressive driving, an optimal window of 11.67 seconds resulted in an MSE of 0.028.
  • Prediction Horizon: The transformer model achieved a prediction horizon 3.5 times longer than LSTM models while maintaining an MSE below 0.05.
  • Safety Performance: The proposed approach significantly improved safety, as indicated by higher TTC values. For example, TTC increased from 90.49 to 300.4 for aggressive driving and from 101.69 to 383.88 for normal driving when using HDV prediction compared to planning without it.
  • Proactive Avoidance: The "avoidance in advance" (AIA) strategy, enabled by accurate long-term predictions, demonstrated better gap management and smoother velocity profiles compared to real-time avoidance (RTA).

Limitations and Future Work

  • Limitations:
    • The current study focuses on a highway merging scenario with a limited number of surrounding vehicles.
  • The comfort of ADS motion planning is not explicitly investigated.
  • Future Work:
    • Extend the proposed approach to scenarios with three or four lanes and more surrounding vehicles.
  • Test the transferability of the method to more complex urban environments.
  • Investigate and optimize the comfort aspect of ADS motion planning.

Summary updated

BibTeX
@misc{dong2023enhanced,
  author = {Ni Dong and Shuming Chen and Yina Wu and Yiheng Feng and Xiaobo Liu},
  title = {{An enhanced motion planning approach by integrating driving heterogeneity and long-term trajectory prediction for automated driving systems}},
  year = {2023},
  url = {https://openreview.net/forum?id=miy9doizsL},
}
Figure