[A Gaussian mixture-hidden Markov model of human visual behavior].

Area of Science:

• Cognitive Science
• Computer Science
• Neuroscience

Background:

• Human vision is crucial for information intake and interaction with the environment.
• Understanding visual behavior requires effective modeling of eye movement patterns (scanpaths).
• Existing models may not fully capture temporal dynamics in scanpath data.

Purpose of the Study:

• To develop and evaluate an optimized Gaussian mixture-hidden Markov model (GMM-HMM) for scanpath analysis.
• To introduce and assess the efficacy of a novel time-shifting segmentation (TSS) method for scanpath modeling.
• To compare the performance of the proposed model against baseline methods and analyze task-specific visual behavior.

Main Methods:

• Utilized Gaussian mixture-hidden Markov models (GMM-HMM) to model human eye movement scanpaths.
• Proposed and implemented a time-shifting segmentation (TSS) technique to enhance temporal feature extraction.
• Employed linear discriminant analysis (LDA) for multi-dimensional feature pattern recognition and model validation.
• Conducted comparative trials using GMM-HMM alone, TSS alone, and the combined GMM-HMM with TSS approach.

Main Results:

• The baseline GMM-HMM achieved 0.507 classification accuracy, surpassing chance (0.333).
• The TSS method alone improved accuracy to 0.610.
• The combined GMM-HMM and TSS model reached 0.602 accuracy with enhanced stability.
• The proposed modeling approach significantly outperformed basic saccade amplitude analysis.

Conclusions:

• The GMM-HMM model demonstrates strong performance in scanpath pattern recognition.
• The integration of TSS method effectively enhances scanpath characteristic differences and model stability.
• The optimized model shows particular advantages for recognizing scanpaths in search-type tasks.
• This research offers a novel solution for analyzing single-state eye movement sequences.