=Paper=
{{Paper
|id=None
|storemode=property
|title=MediaEval 2013 Visual Privacy Task: Pixel Based Anonymisation Technique
|pdfUrl=https://ceur-ws.org/Vol-1043/mediaeval2013_submission_78.pdf
|volume=Vol-1043
|dblpUrl=https://dblp.org/rec/conf/mediaeval/PantojaAI13
}}
==MediaEval 2013 Visual Privacy Task: Pixel Based Anonymisation Technique==
https://ceur-ws.org/Vol-1043/mediaeval2013_submission_78.pdf
MediaEval 2013 Visual Privacy Task: Pixel Based
Anonymisation Technique
Cesar Pantoja Virginia Fernandez Ebroul Izquierdo
Queen Mary University of Queen Mary University of Queen Mary University of
London London London
Mile End Road Mile End Road Mile End Road
E1 4NS, London, UK E1 4NS, London, UK E1 4NS, London, UK
cesar.pantoja@eecs virginia.fernandez@eecs ebroul.izquierdo@eecs
.qmul.ac.uk .qmul.ac.uk .qmul.ac.uk
ABSTRACT egory containing more information than the previous.
In this paper, a pixel-based method for personal anonymi- The core of the method applies a pixelisation filter to all
sation in visual surveillance applications, is presented. The ROIs, but additional steps are performed to each “level”,
proposed method tries to tackle the problem of balancing becoming more and more specific in the type of filter applied.
the intelligibility with privacy, by including some form of The description of the different levels and the filters applied
contextual information in the anonymisation process. Ob- are described in the following subsections, followed by a brief
jective and subjective evaluations show promising results in discussion of the method.
intelligibility and appropriateness, but they also show that 2.1 Accessories and Hair
privacy could be further improved.
This level carries the less information about the person
being anonymised, so the more general filter is applied in this
1. INTRODUCTION step, which is the pixelisation filter. In our tests, a pixel size
Proactive efforts to ensure citizens’ security lead to wide- of 24x24 yielded the best results for the presented scenarios,
spread adoption of invasive video surveillance systems. The but different conditions might need a different pixel size, and
ever-increasing amount of recorded information poses a di- the proposed method is flexible enough to allow a different
rect threat to citizens’ privacy and their right to preserve pixel size via a parameter, so it could be changed easily to
their personal information. Thus, a general social concern adapt to a different scenario.
has emerged for the loss of privacy, demanding new ap-
proaches to preserve and protect it, ensuring their anonymity 2.2 Skin Regions
and freedom of action whilst maintaining the surveillance In this step, a skin colour detector in the hue, satura-
performance. We propose a new approach to preserve per- tion, and value (HSV) colour space is used to detect and
sons’ identity in visual surveillance information for the Me- change the subject’s skin. We used a fixed range in the
diaEval Privacy Task [5]. The proposed approach tries to colour space in which all pixels within this range are con-
balance the privacy and the intelligibility of the scene by sidered skin. Since we are only applying this to skin ROIs,
combining different filters for different parts of the scene. the risk of false positives is low, while the amount of true
Both objective and subjective evaluations show promising positives is maximised. In our experiments, this range was
results in intelligibility and appropriateness, but a low score H ∈ [0◦ , 28.23◦ ], S ∈ [0.04, 1], V ∈ [0, 1].
in privacy shows there is still room for improvement of the The colour of each detected skin pixel is then changed
filter. The rest of the paper is organised as follows: Section to a single colour. This ensures everyone to have the same
2 presents the proposed method, the objectives and design skin colour, effectively concealing their true race. In this
choices behind it’s development. Section 3 presents the eval- particular scenario, we choose the colour (14.11◦ , 0.31, 0.39).
uation of the method using both objective and subjective This step also introduces other parameters to further adapt
metrics. Finally, section 4 draws some closing remarks and the method to different scenarios, which are the skin detec-
states future research opportunities. tion range and the skin colour changing.
2.3 Face
2. PROPOSED METHOD DESCRIPTION The face is considered apart from the skin to improve not
When designing the anonymisation method, the main goal the privacy but the intelligibility. In this step, after the
is to maximise the privacy of the person while maintaining skin colour change and the pixelisation, edges of the face,
a very good intelligibility. With this in mind, we listed the detected by the Canny Edge Detector[1], are overlaid, which
possible types of ROIs and the information they carried to allows to keep some information of the subject and the class
identify a person. After this, we classified the possible types of the ROI. A final set of parameters are introduced in this
of Regions of Interest (ROIs) into three categories, each cat- step with those belonging to the Canny Edge Detector.
2.4 Method Discussion
Copyright is held by the author/owner(s). When developing the anonymisation filter, several things
MediaEval 2013 Workshop, October 18-19, 2013, Barcelona, Spain were considered. Pixelisation, as the main filter, was con-
Figure 1: Output produced by the filter.
Figure 3: Subjective Evaluation
vacy task. The proposed method applies different filters
depending of the level of privacy information carried in a
ROI. Objective and subjective evaluations show that the fil-
ter performs very well in intelligibility and appropriateness
but there is still opportunities to improve the privacy pre-
serving aspect of the filter. Because the filter is developed in
a modular way, different parts of the filter can be improved
separately. The inclusion of parameters also allows to an
improvement in results without modifying the method itself.
For example, the pixelisation filter could produce smaller or
bigger pixels. The range of colour in the colour detection,
Figure 2: Objective Evaluation
and the target colour in the skin colour change, could be all
changed to include more (or less) tones of skins or to produce
sidered because it has shown to have a very good balance of a darker or lighter skin tone. Finally, the Canny Edge De-
privacy and intelligibility [3][4]. Additionally, Skin colour is tection algorithm used in the face introduces it’s own set of
regarded as one of the most important features when iden- parameters to detect the edges. More advanced adjustments
tifying humans[2], which is why an additional step was con- to the filter include the colour change part of the method.
sidered to conceal the person’s real skin colour before the For example an advance colour transfer technique could be
pixelisation filter. Finally, the face’s edges are imposed over used to produced more natural results. The edge detection
the colour change and pixelisation of the face ROI as a mea- could also be improved by softening the detected edges to
sure of intelligibility. This allows to keep some information produce a more natural result, or using a different method
on the subject’s face while keeping his true identity con- to detect the edges altogether.
cealed. Figure 1 shows an example output produced by the
filter, next to the original, unfiltered ROI. 5. REFERENCES
The filter was implemented in C++ using the OpenCV [1] J. Canny. A computational approach to edge detection.
library for image processing and Xerces-C++ to load the Pattern Analysis and Machine Intelligence, IEEE
ROIs from an XML file. In particular, the implementation Transactions on, PAMI-8(6):679–698, 1986.
uses the parallel programming paradigm to perform the dif- [2] M. Demirkus, K. Garg, and S. Guler. Automated
ferent stages of the algorithm concurrently to achieve a near person categorization for video surveillance using soft
real-time performance. biometrics. Proc. SPIE, 7667:76670P–76670P–12, 2010.
[3] P. Korshunov, C. Araimo, F. De Simone, C. Velardo,
3. EVALUATION RESULTS J. Dugelay, and T. Ebrahimi. Subjective study of
Results of the objective and subjective evaluations are privacy filters in video surveillance. In Multimedia
shown in Figures 2 and 3 respectively. The figure shows Signal Processing (MMSP), 2012 IEEE 14th
the score of the proposed method paired against the average International Workshop on, pages 378–382, 2012.
score of the other methods presented in the challenge. The [4] P. Korshunov, S. Cai, and T. Ebrahimi. Crowdsourcing
use of pixelisation combined with the face’s edges combined approach for evaluation of privacy filters in video
paid of in the high score seen in intelligibility and appropri- surveillance. In Proceedings of the ACM multimedia
ateness, while the change of the colour skin did not affect 2012 workshop on Crowdsourcing for multimedia,
this measures. The evaluation results also show that there is CrowdMM ’12, pages 35–40, New York, NY, USA,
still room for improvement in the privacy preserving aspect 2012. ACM.
of the filter. [5] T. Reuter, S. Papadopoulos, V. Mezaris, P. Cimiano,
C. de Vries, and S. Geva. Social Event Detection at
4. CONCLUSIONS AND FUTURE WORK MediaEval 2013: Challenges, datasets, and evaluation.
A pixel-based anonymisation method of visual surveillance In MediaEval 2013 Workshop, Barcelona, Spain,
information has been presented for the MediaEval 2013 pri- October 18-19 2013.