Active Vision in Driving: Joint Modeling of Scanpaths and Risk Perception.

Area of Science:

• Cognitive Science
• Human-Computer Interaction
• Robotics

Background:

• Eye movements are actively guided by task demands, not passive responses.
• Driver scanpaths may reflect information sampling for risk assessment.
• Existing models often separate scanpath prediction from risk assessment.

Purpose of the Study:

• To investigate joint modeling of driver scanpaths and traffic risk perception.
• To propose a unified computational framework for gaze behavior and risk assessment.
• To explore the cognitive coupling between visual attention and safety perception.

Main Methods:

• Developed an Adversarial Inverse Reinforcement Learning (AIRL) framework.
• Used a generator for simulating fixation/saccade sequences and a discriminator for reward approximation.
• Constructed the BDDA dataset with over 13,000 spatio-temporal gaze points and risk annotations.

Main Results:

• Simultaneously modeled scanpath dynamics and risk perception.
• Achieved superior performance compared to baseline methods on the BDDA dataset.
• Demonstrated that generated scanpaths synergistically inform risk perception.

Conclusions:

• Provided computational evidence for a functional coupling between visual attention and risk perception in driving.
• Supported the Active Vision hypothesis in safety-critical environments.
• Highlighted the role of eye movements in acquiring task-relevant information for risk assessment.