Vehicle Interaction Behavior Prediction with Self-Attention.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Predicting vehicle interaction behavior on structured roads is challenging due to high behavioral uncertainty.
• Accurate prediction is crucial for effective ego-vehicle control and collision avoidance.

Purpose of the Study:

• To propose a novel interaction behavior prediction model, termed VC-Attention, to enhance prediction performance.
• To improve the safety and efficiency of autonomous driving systems through better behavioral forecasting.

Main Methods:

• A five-vehicle based cluster structure was designed to extract interactive features like Deceleration Rate to Avoid a Crash (DRAC) and lane gap.
• The sliding window algorithm was employed to extract vehicle cluster (VC) behavior information.
• Two layers of self-attention encoder with six heads captured temporal characteristics of interactive features, followed by a fully connected layer and SoftMax module for prediction.

Main Results:

• The VC-Attention model achieved accuracy, precision, recall, and F1 scores exceeding 92% on the Next Generation Simulation (NGSIM) dataset.
• The model demonstrated a time to event of 2.9 seconds.
• High performance was maintained even in class-imbalanced prediction scenarios and across various driving conditions.

Conclusions:

• The proposed VC-Attention model significantly improves vehicle interaction behavior prediction accuracy.
• This method offers a robust solution for real-time prediction in complex driving environments.
• The model's adaptability to diverse scenarios makes it suitable for advanced driver-assistance systems and autonomous vehicles.