Exploring Edge Vision-Language Model for Privacy-Aware Activity Recognition
Student Contest:
No
Affiliation Type:
Academia
Keywords:
Vision-Language Model, Edge computing, Human Activity Recognition
Abstract:
Activity recognition with wearable cameras has progressed thanks to large egocentric datasets, but capturing faces and other identifying details raises privacy issues. Existing methods like blurring or masking often reduce reusability. To address this, we propose a privacy-preserving approach using Edge VLM (a lightweight Vision-Language Model) to generate descriptive captions in place of raw images, then combine them with sensor data for recognition. Tested on six desk activities, our method showed higher user acceptability (p < 0.05) than blurring or cartoonization, and achieved 77.1% accuracy with SVM—comparable to using raw images—demonstrating the promise of Edge VLM-based captioning for privacy-aware activity recognition.
Track ID:
5.8
Track Name:
Japan-Korea Joint Special Session on Intelligent Systems & Applications for Future Networked Society
