Thesis Topic Details

Topic ID:
Addressing Privacy Concerns in Images Captured by Wearable Cameras
Salil Kanhere
Research Area:
Privacy, Computer Vision, Pervasive computing
Associated Staff
Chun Tung Chou
Topic Details
R & D
Group Suitable:
Knowledge of machine learning, computer vision and algorithms will be useful
Over the last few years, wearable cameras have emerged as a new way to capture and record a wide variety of experiences from a first-person point-of-view (FPPOV) perspective. Due in large part to improvements in camera, battery and storage technologies, wearable cameras can be now packaged in a form factor that allows them to be lightweight, unobtrusive, and easy to mount or carry without restricting the wearer's activity. Recent consumer wearable cameras include the GoPro, Replay and Contour, which are designed specifically for sports activities, and the Looxcie and Memoto, aimed at recording everyday moments for archival and future review. Google's Glass Project has also fuelled the excitement for general purpose consumer wearable cameras.

Motivated by the wearable camera's ability to directly and continuously observe and record real-world settings, researchers have begun to explore the potential of FPPOV images in a number of domains, such as autism support, travel behaviour, and activity recognition.

Accompanying the exciting possibilities of wearable cameras are a host of technical and nontechnical challenges. One of the fundamental issues is how to process the volumes of data (continuous video or sequence of images taken over the course of a day). Analysing these images for activities of interest involves reviewing photos manually, a tedious, time- consuming and error-prone task. Although much progress has been achieved in the area of computer vision, state-of- the-art algorithms do not yet yield the automated performance on real-world images that is required for many practical applications. In light of these limitations, an approach that has been embraced by researchers is to use human computation techniques to analyse FPPOV images, using services like Amazon Mechanical Turk or other commercial solutions.

Although human computation has proved to be a viable image analysis alternative to manual or algorithmic techniques, it introduces some challenges of its own, in particular privacy. Privacy is always front and center when it comes to collecting FPPOV images from wearable cameras. These images might capture sensitive information of the person wearing the camera or reveal the identity of others who are captured in the photos as well. If these images were to re- main in the exclusive possession of the individual wearing the camera, privacy risks would be kept to a minimum. How- ever, if they are uploaded to be inspected by non-trustworthy third-parties, such as Amazon Mechanical Turk workers, additional precautionary steps must be taken to reduce or, ideally, eliminate the possibility of privacy violations.

Although privacy in FPPOV images has been recognised by the community as an area that deserves furthers exploration, additional studies are needed to examine techniques for mitigating the privacy risk that is inherent in this type of image capture. In this project, we will investigate a variety of approaches from computer vision literature in addressing these issues.

An empirical study will be undertaken to collect a large set of FPPOV imagery from volunteers. A number of GoPro cameras and smarphones are available for the experiments.
There is potential to write-up a high-quality conference/journal article based on the results from this project.
Past Student Reports
No Reports Available. Contact the supervisor for more information.

Check out all available reports in the CSE Thesis Report Library.

NOTE: only current CSE students can login to view and select reports to download.