In this paper, we describe EPFL privacy protection tool submitted to the MediaEval 2013 Visual Privacy task. The goal of the task is to obscure faces and other personal items of people in the provided surveillance clips to preserve their personal privacy. In the privacy protection tool, we used a combination of reversible privacy protection filter based on geometric warping transformation, randomized saturation filter, masking with partial opacity, and pixelization. The aim of the implementation was to achieve an acceptable balance between privacy and intelligibility, as well as, privacy and appropriateness. The results of both objective and subjective evaluations provided by the organizers of the task demonstrated that our privacy protection tool leads to high appropriateness and intelligibility (the surveillance task can be performed with high accuracy) while keeping strong privacy protection.
The problem of privacy protection in video surveillance systems
gaining more and more attention from research and industry. Many
privacy protection tools were proposed to mitigate privacy
intrusiveness of modern surveillance systems. These protection
techniques vary from such simple approaches like blurring,
pixelization, or masking to more advanced methods satisfying the
following desirable practical properties: reversibility, robustness, and
security. The advanced methods can be divided into several
categories: encryption-based [
Despite wide availability of visual privacy protection tools, with
an exception of a few works [
In our submission to MediaEval 2013 Visual Privacy task, we
aimed to address both tradeoffs. We have built a privacy protection
tool based on reversible and secure warping filter [
Organizers of the task provided video dataset [
The proposed privacy protection tool adopted a two-stage
approach (see Figure 1 for an illustration): (i) warping [
Warping filter makes the details of the visible object unrecognizable (i.e., privacy is increased), but, by controlling its strength, we can keep its overall general shape preserved, so we can still understand what is going on in the surveillance scene (i.e., intelligibility is not decreased). Higher intelligibility is also insured by not distorting the whole body regions keeping intact the less privacy sensitive visual information. Randomized saturation and opacity masking allow us to decrease the color level of skin, hair, or accessories, so they are not recognizable (i.e., privacy is increased), yet the original colors can be recovered if needed. 2.1
The best privacy preserving filter would be a blacked out camera with no video feed, but, in such case, there would be no surveillance possible and intelligibility would be zero. Therefore, a usable privacy protection filter should have a balance between privacy and intelligibility. Similarly, an encryption or scrambling based privacy filters could lead to high privacy but can be annoying or even scary, resulting in very low appropriateness.
Aiming to balance between these tradeoffs, we made the following implementation decisions:
Warping is applied to faces and accessories. The aim is to distort facial features and details of accessories, preserving, in the same time, a general appearance of people to keep understanding of the scene and actions clear.
Strong pixelization is applied to hair and skin regions. Since these regions are not as important for surveillance purposes as, for example, faces, but still carry information about gender and race, higher degree of protection is required.
Randomized saturation is applied to faces, skin, and hair and opacity masking is applied to face (opacity value 0:3) and accessories (opacity value 0.8) to hide color and detailed information about these regions.
No filter is applied to body regions to keep visible as much intelligibility information as possible.
The evaluation results provided by the organizers of the task are summarized in Table 1, where our objective and subjective evaluation results are compared with the average result of the total 9 submissions. The results are favorable and demonstrate that our privacy protection tool achieves a reasonable balance between privacy, intelligibility, and appropriateness, with all scores except for privacy being well above the average. 4.
The proposed privacy protection tool combined several privacy protection filters achieving a balance between privacy, intelligibility, and appropriateness. For the future work, the strength of adopted filters in the tool should be adjusted, so the privacy increased while intelligibility stays at the same level.
This work was conducted in the framework of the EC funded Network of Excellence VideoSense. Special thanks to Andrea Melle (Eurecom) for providing parts of the code for our implementation. 5.