/VTP

Vehicle Trajectory Prediction with Deep Learning Models

Primary LanguagePython

VTP: Deep Learning for Vehicle Trajectory Prediction

Attention-LSTM: An LSTM model in a Mixed Traffic Flow Environment (Connected and Autonomous Vehicle, Human-driven Vehicle)

  • Attention-LSTM assumes a Connected and Autonomous Vehicle (CAV) utilizes all the CAVs within its communication range (the front and back 200 meters in the left, middle and right lanes) to predict future trajectories of a target Human-driven Vehicle (HDV).

  • Data: The training/validation/test datasets extracted from Next Generation Simulation (NGSIM) Vehicle Trajectories can be downloaded here.

Performance of Attention-LSTM under different CAV Market Penetration Ratios

STA-LSTM: An LSTM model with spatial-temporal attention mechanisms

  • STA-LSTM utilizes T-step historical trajectories of all vehicles within a 3×13 grid centered around the target vehicle to predict its future trajectories. The architecture of STA-LSTM is shown below:

STA-LSTM achieves comparable prediction performance against other state-of-the-art models

Models RMSE per prediction time step
1st 2nd 3rd 4th 5th
physics-based model 0.1776 0.3852 0.6033 0.8377 1.0888
naive LSTM 0.1012 0.2093 0.3384 0.4830 0.6406
SA-LSTM 0.1026 0.2031 0.3157 0.4367 0.5643
CS-LSTM [1] 0.1029 0.2023 0.3146 0.4364 0.5674
STA-LSTM 0.0995 0.2002 0.3130 0.4348 0.5615

[1] Nachiket Deo and Mohan M. Trivedi,"Convolutional Social Pooling for Vehicle Trajectory Prediction." CVPRW, 2018

Average temporal-level attention weights of the past six time steps

Spatial-level attention weight analysis

  • Distributions of spatial-level attention weights by target vehicle class (excluding weights in the target vehicle's cell):

  • Averaged spatial-level attention weights by the number of neighboring vehicles

  • Maximum spatial-level attention weight frequency by target vehicle location

  • Maximum spatial-level attention weights regrading the lane-changing behaviors of the target vehicle 2858. The target vehicle 2858 executed two lane-changing maneuvers around the 996th and 1220th time steps, respectively. The vehicle relocated the maximum attention from the current lane to (Right, 1) and (Right, 2), as it was preparing to change to the right lane.

Citation

You are more than welcome to cite our papers:

@article{lin2021long,
  title={Long Short-Term Memory-Based Human-Driven Vehicle Longitudinal Trajectory Prediction in a Connected and Autonomous Vehicle Environment},
  author={Lin, Lei and Gong, Siyuan and Peeta, Srinivas and Wu, Xia},
  journal={Transportation Research Record},
  pages={0361198121993471},
  year={2021},
  publisher={SAGE Publications Sage CA: Los Angeles, CA}
}

@article{Lin2020Attention,
  title={Vehicle Trajectory Prediction Using LSTMs with Spatial-Temporal Attention Mechanisms},
  author={Lin, Lei and Li, Weizi and Bi, Huikun and Qin, Lingqiao},
  journal={IEEE Intelligent Transportation Systems Magazine (ITSM)}
}