Privacy-Preserving Visual Localization With Event Cameras.

Area of Science:

• Computer Vision
• Robotics
• Edge Computing

Background:

• Client-server localization is vital for AR/VR and mobile apps.
• Resource-limited edge devices face challenges with large 3D maps and efficient localization.
• Existing systems struggle with computational efficiency, robustness, and privacy.

Purpose of the Study:

• To develop a client-server localization pipeline using event cameras.
• To address computational efficiency, robustness, and privacy concerns in edge device localization.
• To enable practical location-based services on resource-constrained devices.

Main Methods:

• Utilized event cameras for low energy consumption and small memory bandwidth.
• Implemented event-to-image conversion for leveraging established image-based localization techniques.
• Introduced dual-level privacy protection: network-level split inference and sensor-level filtering.

Main Results:

• Achieved robust localization in challenging conditions like low-light and fast motion.
• Demonstrated significant mitigation of privacy concerns with minimal performance loss.
• User study confirmed reduced feelings of insecurity.

Conclusions:

• The proposed event camera-based system offers an efficient and private solution for client-server localization.
• This pipeline serves as a foundational component for future location-based services on edge devices.
• The system enhances user experience by addressing privacy and security in visual data transmission.