Ego-Surfing: Person Localization in First-Person Videos Using Ego-Motion Signatures

Area of Science:

• Computer Science
• Human-Computer Interaction
• Computer Vision

Background:

• Wearable cameras offer potential for assistive technologies but raise significant privacy concerns due to passive recording of individuals.
• First-person videos capture everyday life, necessitating methods to manage privacy and enable content retrieval.

Purpose of the Study:

• To develop a self-search technique specifically for first-person videos recorded by wearable cameras.
• To address privacy challenges and unlock applications like content retrieval and privacy filtering.

Main Methods:

• The core technique leverages the observation that egocentric head motion is visible in both the target person's and observer's first-person videos.
• Motion correlation is computed over densely-sampled trajectories to uniquely identify instances of the self.
• The approach compares motion correlation against established face detection and recognition methods.

Main Results:

• The proposed self-search method significantly outperforms existing face detectors and recognizers in identifying individuals in first-person videos.
• Demonstrated the effectiveness of motion correlation for unique self-identification.

Conclusions:

• The developed self-search technique offers a robust solution for managing and utilizing first-person video data.
• This approach facilitates practical applications including privacy filtering, targeted video retrieval, and social group analysis.