Vehicle trajectory prediction and generation using LSTM models and GANs.

Area of Science:

• Computer Science
• Artificial Intelligence
• Transportation Engineering

Background:

• Vehicle trajectory prediction is crucial for autonomous driving, traffic management, and urban planning.
• Existing methods struggle with challenges like multimodality and generalizability in predicting vehicle movements from Floating Car Data (FCD).

Purpose of the Study:

• To address the limitations in current vehicle trajectory prediction models, particularly concerning multimodality and generalizability.
• To propose novel datasets, evaluation metrics, and deep learning models for improved trajectory prediction.

Main Methods:

• Developed and compared Long Short-Term Memory (LSTM) and Generative Adversarial Network (GAN) based deep learning models.
• Introduced new datasets and evaluation metrics (N-ADE, N-FDE) to better assess prediction accuracy.
• Conducted experiments in four Italian cities using real-world FCD.

Main Results:

• GAN-based models, specifically GAN-3, demonstrated superior performance in multimodal traffic scenarios.
• LSTM models proved effective in unimodal traffic situations.
• New metrics (N-ADE, N-FDE) provided a normalized evaluation, reducing biases in standard metrics.

Conclusions:

• The proposed generative models offer a significant advancement for multimodal vehicle trajectory prediction.
• The study highlights the importance of specialized models for different traffic complexities.
• The methodology is validated for real-world applications in traffic management and urban planning.