Dynamic probability integration for electroencephalography-based rapid serial visual presentation performance...

Area of Science:

• Neuroscience
• Computer Vision
• Signal Processing

Background:

• Rapid serial visual presentation (RSVP) is a sensitive EEG-based target detection method.
• Current EEG-RSVP algorithms struggle with attention lapses causing weak or absent event-related potentials (ERPs).
• Fusing human and computer vision offers complementary information for robust object detection, especially during attention lapses.

Purpose of the Study:

• To introduce dynamic probability integration (DPI) for fusing human and computer vision in EEG-RSVP tasks.
• To develop a novel basic probability assignment (BPA) method for weight generation based on classification capabilities.
• To design a spatial-temporal hybrid common spatial pattern-principal component analysis (STHCP) algorithm for decoding EEG signals.

Main Methods:

• Dynamic probability integration (DPI) fuses human and computer vision using a novel BPA method.
• The BPA method considers classification capabilities of heterogeneous information sources.
• A spatial-temporal hybrid common spatial pattern-principal component analysis (STHCP) algorithm decodes EEG signals using spatial-temporal features.

Main Results:

• DPI achieved an average AUC of 0.912 ± 0.041 in nighttime vehicle detection via RSVP, outperforming individual methods and other fusion techniques.
• DPI demonstrated superior balanced accuracy (0.845 ± 0.052), indicating balanced target and non-target detection.
• STHCP achieved the highest AUC (0.818 ± 0.06) compared to baseline methods.

Conclusions:

• The proposed fusion method (DPI) significantly improves detection performance in RSVP tasks compared to individual and existing fusion methods.
• DPI and STHCP show promise for enhancing object detection, particularly under attention-lapsing conditions.
• This approach offers an efficient and generalizable solution for object detection in challenging visual environments.