Kernelized convolutional transformer network based driver behavior estimation for conflict resolution at unsignalized...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Robotics

Background:

• Driver behavior modeling is vital for Advanced Driver Assistance Systems (ADAS).
• Accurate behavior estimation of surrounding vehicles is critical for autonomous vehicle navigation, especially in unsignalized intersections.
• Unsignalized three-way roundabouts present complex driving scenarios requiring sophisticated behavior prediction.

Purpose of the Study:

• To propose a novel Kernelized Convolutional Transformer Network (KCTN) for driver behavior estimation.
• To enhance the model's capacity to capture higher-order feature interactions in non-linear spaces.
• To improve the accuracy and lead time of behavior prediction in challenging driving environments.

Main Methods:

• Developed a Kernelized Convolutional Transformer Network (KCTN) incorporating a multi-head attention (MHA) mechanism.
• Introduced a kervolution operation, generalizing convolution into non-linear space using a Gaussian kernel function.
• Validated the proposed model using the real-world ACFR dataset.

Main Results:

• The KCTN model demonstrated superior performance compared to current state-of-the-art methods.
• Achieved higher accuracy in driver behavior prediction.
• Provided a significant lead time for identifying potential conflict situations.

Conclusions:

• The proposed KCTN with kervolution and MHA is effective for driver behavior estimation at unsignalized roundabouts.
• The model enhances the safety and reliability of autonomous navigation systems.
• This approach offers a significant advancement in predicting complex driver behaviors for ADAS and autonomous vehicles.