IDM-Follower: A Model-Informed Deep Learning Method for Long-Sequence Car-Following Trajectory Prediction

Yilin Wang, Yiheng FengarXiv 2022 · OpenReview

TL;DR

This paper introduces "IDM-Follower," a novel deep learning framework that combines the strengths of physics-based car-following models (like IDM) with sequence-to-sequence autoencoders. By integrating physical constraints into a learning model, IDM-Follower achieves more robust and accurate long-term trajectory predictions for vehicles, especially in the presence of noisy sensor data.

What Problem Does This Paper Address?

Predicting the future movement of vehicles, particularly their "car-following" behavior (how one vehicle follows another), is crucial for intelligent transportation systems, traffic management, and autonomous driving. Traditional approaches fall into two camps: model-based methods, which use explicit mathematical equations but struggle with individual driving variations and noisy data, and learning-based methods, which can capture complex patterns but often act as "black boxes" and can be sensitive to data quality. Existing models also tend to predict only the next time step, not long sequences. This paper aims to bridge this gap by creating a model that leverages the interpretability and robustness of physics-based models while benefiting from the data-driven flexibility of deep learning for long-sequence prediction.

Methods

The core innovation is the "IDM-Follower" framework, which integrates the Intelligent Driver Model (IDM), a well-established physics-based car-following model, with a sequence-to-sequence recurrent autoencoder.

  • Learning Component: A recurrent autoencoder (specifically, a symmetric encoder-decoder structure) is used to learn the mapping from input sequences (leading vehicle's trajectory) to output sequences (following vehicle's trajectory). To handle long sequences and memory limitations, the architecture incorporates a self-attention mechanism. It features two independent encoders for position and velocity, and a decoder with attention to extract and reconstruct features.
  • Model Component: The IDM model is used to generate a "model-based prediction" of the following vehicle's trajectory. The parameters for the IDM are pre-calibrated.
  • Hybrid Loss Function: A novel loss function is designed to train the autoencoder. It's a weighted sum of two components:
  • Data Loss ($L_{data}$): Measures the discrepancy (using Root Mean Square Error - RMSE) between the neural network's predicted trajectory and the ground truth trajectory.
  • Model Loss ($L_{model}$): Measures the discrepancy (using RMSE) between the IDM's predicted trajectory and the neural network's predicted trajectory. This term acts as a regularizer, guiding the learning process with physical constraints.
  • Optimization: The autoencoder's parameters are updated via backpropagation using the combined loss function.

Datasets

The framework was evaluated on two datasets:

  1. SUMO Simulation Dataset: Generated using the SUMO traffic simulator. This dataset was created to mimic real-world conditions, including the use of IDM as the default car-following model within SUMO. GPS noise was artificially added to simulate localization errors.
  2. NGSIM Dataset: Real-world trajectory data from the Lankershim Boulevard dataset. Similar to the simulation data, GPS noise was added to the ground truth trajectories to create training data.

Key Insights and Ideas

  • Physics-Informed Learning for Robustness: Integrating a physics-based model like IDM into a deep learning framework significantly improves robustness against noisy sensor data (like GPS errors). While pure learning models can become erratic with noise, the IDM-Follower maintains more stable predictions.
  • Synergy of Model and Data: The hybrid loss function is key. It allows the model to learn complex, data-driven patterns while being constrained by the physical plausibility enforced by the IDM. This prevents the learning model from deviating into unrealistic behaviors.
  • Long-Sequence Prediction with Attention: The use of self-attention in the autoencoder architecture is crucial for effectively handling long-term dependencies in vehicle trajectories, a common challenge in sequential prediction tasks.
  • Optimal Weighting is Scenario-Dependent: The effectiveness of the IDM-Follower depends on the weighting ($\mu$) between the data loss and model loss. The optimal weight varies with noise levels and dataset characteristics, suggesting that a careful balance is needed for best performance.

Key Results

  • Improved Prediction Accuracy: The IDM-Follower consistently outperformed both pure learning-based autoencoders and pure model-based IDM predictions on both simulation and real-world (NGSIM) datasets, especially under noisy conditions.
  • Robustness to Noise: On the SUMO dataset, the IDM-Follower showed significant improvement over the learning-based baseline (26.9% vs. 2.1%) when dealing with middle noise levels, demonstrating its robustness.
  • Qualitative Improvements: Visualizations showed that the IDM-Follower produced smoother trajectories with fewer physically implausible fluctuations (e.g., vehicles moving backward) compared to the pure learning-based method. The model-based IDM, while smooth, tended to accumulate errors over time.
  • Optimal $\mu$ Varies: For the SUMO dataset, the best performance was achieved with $\mu=0.7$ (30% IDM) under small noise and $\mu=0.5$ (50% IDM) under middle noise. For the NGSIM dataset, $\mu=0.7$ was consistently optimal across different noise levels.

Limitations and Future Work

  • Optimal Weight Selection: The paper notes that determining the optimal weight ($\mu$) for combining the model and learning components is crucial and scenario-dependent. Further research is needed to develop methods for automatically selecting these optimal weights.
  • Model Parameter Sensitivity: Table 6 suggests that the performance of the IDM-Follower can be sensitive to the accuracy of the pre-calibrated IDM parameters. Using less accurate physical model parameters can degrade the performance of the model-informed framework.
  • Generalizability: While the framework is presented as generic, further testing across a wider range of driving scenarios and environments would be beneficial.
  • Future Work: The authors suggest more comprehensive testing and theoretical analysis on optimal weight selection. They also aim to explore how to choose optimal weights between the model-based and learning-based components under various scenario settings.

Summary updated

BibTeX
@misc{wang2022idm,
  author = {Yilin Wang and Yiheng Feng},
  title = {{IDM-Follower: A Model-Informed Deep Learning Method for Long-Sequence Car-Following Trajectory Prediction}},
  year = {2022},
  url = {https://openreview.net/forum?id=a23h5ut0fI},
}
Figure