Our solution to the MediaEval 2014 Visual Privacy Task [4] is a privacy-preserving video lter that is able to maintain a high intelligibility level in surveillance systems while providing a reasonable privacy protection level to monitored people and a pleasant view to observers. This paper describes our context-aware method that is based on cartooning and pixelation e ects. Subjective evaluation results are also presented to demonstrate the performance of our algorithm.
Surveillance cameras play various roles in our everyday
lives and their increasing number attracted attention to
privacy issues. The goal of the MediaEval 2014 Visual
Privacy Task [
Traditional CCTV cameras are continually being replaced
by more modern smart cameras which are usually part of
Visual Sensor Networks (VSNs). Other widespread
videocapable devices such as smart phones, tablets or web-cams
also pose privacy threats due to their frequent use in
public spaces. Processing capabilities of these devices allow the
integration of privacy protection methods directly into the
camera. Our aim is to create such an integrated lter. In
order to simulate the limited computational power of the above
mentioned embedded devices, we ran our privacy-preserving
algorithm on the Jetson TK1 [
Our method is based on a cartooning e ect which is applied both globally and locally. In sensitive regions the lter intensity is adjusted according to the annotation. Faces are further protected with an extra pixelation e ect.
A prototype of our lter is implemented in C++ by using
OpenCV [
First, a medium-intensity cartooning e ect is applied to
the whole video frame. This always ensures a default level of
privacy thereby preparing the lter for real-world use where
privacy loss may occur at sensitive regions due to
inaccurate feature extractors. Additionally, implicit privacy
channels [
The parameters used in the steps above are dependent on desired privacy levels taken from the annotation les. k=17, sp=30, sr=60 are used for high level; k=9, sp=20, sr=40 for medium level; and k=3, sp=10, sr=20 for low level privacy. For global cartooning we used the medium level. 2.2
After global cartooning, protection levels of sensitive
regions are adjusted locally according to Table 1 in [
In the nal step of our processing pipeline an extra pixelation e ect is applied on faces in order to further obscure the identity of people. The region of pixelation is the maximum inscribed ellipse of the face's bounding box and the pixel size is one- fteenth of its larger dimension.
(a) Original. (b) Filtered. (c) Original. (d) Filtered.
Two pairs of video frOavmereasll (original and ltered) in Figure 2 demonstrate the visual e ect of our privacy lter.
In terms of processing speed the Jetson TK1 board is
capable of 5 fps for 320 180, 2 fps for 640 360, 1 fps
for 800 450, 0.8 fps for 1024 576, and 0.2 fps for the
provided full HD resolution videos. It proves that privacy
protection can really be done directly inside the camera.
Despite running the Mean Shift lter on the GPU instead of
the CPU it remains the bottleneck of our algorithm. Thus,
the current version of our prototype cannot lter full HD
videos in real time, although acceptable frame-rates can
be achieved at lower resolutions. More detailed discussion
about achievable frame-rates on embedded devices can be
found in [
Figure 3 shows the subjective evaluation results provided by the Visual Privacy Task organizers. These numbers are calculated from the outcome of a 12-question survey that has been conducted in three di erent groups. The rst group consists of 230 regular people and the questionnaire was lled out in frame of a crowd-sourcing campaign. The second group is constituted of 65 participants from Thales, France. And the third is a focus group with 59 participants from all over the world. Questions of the survey are assigned around the following three criteria: intelligibility, privacy, and pleasantness.
After analysing Figure 3, it is clear that the performance of our method is always better than the median performance among the 8 participants in terms of intelligibility and pleasantness. We also achieved competitive results as for privacy, although we slightly underperform the median. 4.
By introducing a global component to our privacy protection lter we cover implicit privacy channels and ensure a default level of privacy even if inaccurate real-world feature detectors are being used. Our method provides a pleasant view and high intelligibility while reasonably protecting privacy. It works reasonably well on the Jetson TK1 board for lower resolution videos, although further improvements are necessary to reach acceptable frame-rates for full HD videos. 5.
This work was partly funded by the European Regional Development Fund and the Carinthian Economic Promotion Fund (KWF) under grant KWF-3520/23312/35521.