MALS-Net: A Multi-Head Attention-Based LSTM Sequence-to-Sequence Network for Socio-Temporal Interaction Modelling and...

Area of Science:

• Autonomous Driving
• Deep Learning
• Computer Vision

Background:

• Accurate vehicle trajectory prediction is crucial for autonomous driving safety, especially on highways.
• Existing methods using RNNs, CNNs, and GNNs on the NGSIM dataset face challenges with noise and overfitting.
• Transformers, despite success in NLP, are under-explored for trajectory prediction due to autoregressive decoding errors.

Purpose of the Study:

• To develop a robust trajectory prediction model that overcomes limitations of existing methods.
• To leverage transformer-like attention mechanisms without accumulating errors in time-series forecasting.
• To evaluate the proposed model on a more practical dataset, BLVD.

Main Methods:

• Proposed MALS-Net: a Multi-Head Attention-based LSTM Sequence-to-Sequence model.
• Utilized an attention-based LSTM encoder-decoder architecture to mitigate cumulative errors.
• Evaluated the model on the BLVD dataset, known for its practicality and reduced overfitting issues.

Main Results:

• MALS-Net demonstrated state-of-the-art performance in trajectory prediction.
• Achieved superior results for both short-term and long-term prediction horizons.
• Outperformed existing relevant approaches on the BLVD dataset.

Conclusions:

• MALS-Net offers a significant advancement in autonomous driving trajectory prediction.
• The model effectively combines attention mechanisms with LSTM for accurate and reliable forecasting.
• The findings suggest MALS-Net's potential for enhancing the safety and efficiency of autonomous vehicles.