A Privacy-Preserving Wearable Camera Setup for Dietary Event Spotting in Free-Living

摘要

We designed a wearable head-mounted egocentric camera setup for dietary data collection in free-living. We addressed the problem of privacy-sensitive image content by fixing a camera on a cap’s visor pointing downwards. Salient content was maintained while drastically constraining unwanted privacy-infringing content. The privacy preservation capability of our setup was compared with literature using a modified privacy-saliency matrix. Furthermore, we implemented a dietary event spotting algorithm to reduce the amount of workload for human operator while performing analysis on a large volume of data. Transfer learning on a deep neural network was employed to perform dietary object detection and, subsequently, dietary event spotting. Average recall performance over 90% suggested the feasibility of the method.