Mitigating Bystander Privacy Concerns in Egocentric Activity Recognition with Deep Learning and Intentional Image Degradation

Abstract

Recent advances in wearable camera technology and computer vision algorithms have greatly enhanced the automatic capture and recognition of human activities in real-world settings. While the appeal and utility of wearable camera devices for human-behavior understanding is indisputable, privacy concerns have limited the broader adoption of this method. To mitigate this problem, we propose a deep learning-based approach that recognizes everyday activities in egocentric photos that have been intentionally degraded in quality to preserve the privacy of bystanders. An evaluation on 2 annotated datasets collected in the field with a combined total of 84,078 egocentric photos showed activity recognition performance with accuracy between 79% and 88% across 17 and 21 activity classes when the images were subjected to blurring (mean filter k=20). To confirm that image degradation does indeed raise the perception of bystander privacy, we conducted a crowd sourced validation study with 640 participants; it showed a statistically significant positive relationship between the amount of image degradation and participants' willingness to be captured by wearable cameras. This work contributes to the field of privacy-sensitive activity recognition with egocentric photos by highlighting the trade-off between perceived bystander privacy protection and activity recognition performance.