=Paper=
{{Paper
|id=None
|storemode=property
|title=MediaEval 2013 Visual Privacy Task: Using Adaptive Edge Detection for Privacy in Surveillance Videos
|pdfUrl=https://ceur-ws.org/Vol-1043/mediaeval2013_submission_76.pdf
|volume=Vol-1043
|dblpUrl=https://dblp.org/rec/conf/mediaeval/EiseleinSKS13
}}
==MediaEval 2013 Visual Privacy Task: Using Adaptive Edge Detection for Privacy in Surveillance Videos==
https://ceur-ws.org/Vol-1043/mediaeval2013_submission_76.pdf
MediaEval 2013 Visual Privacy Task: Using Adaptive Edge
Detection for Privacy in Surveillance Videos
Volker Eiselein Tobias Senst Ivo Keller
Communications Systems Communications Systems Communications Systems
Group Group Group
Technische Universität Berlin Technische Universität Berlin Technische Universität Berlin
eiselein@nue.tu- senst@nue.tu-berlin.de keller@nue.tu-berlin.de
berlin.de
Thomas Sikora
Communications Systems
Group
Technische Universität Berlin
sikora@nue.tu-berlin.de
ABSTRACT For obfuscation of personal data, we propose to show only
In this paper we present a system for preserving the privacy the contour of a person in the video. In case of a low number
of individuals in a video surveillance scenario. While a per- of persons in the scene, their movements and actions could
son’s privacy should not be revealed to a viewer of the video be identified while it is still impossible to identify personal
without special needs, it is still important that the action in details such as the person’s face, color of clothes or skin
a scene as the semantic content of a video remain perceivable color. In this paper, we show a system which uses adaptive
by a human observer. edge detection to determine the contour of a person and
The proposed system uses edge detection and adaptive in the privacy-filtered result video replaces the interior of a
thresholding in order to estimate the persons’ silhouettes in person’s silhouette by background information.
a video scene and thus rendering most of their actions visi-
ble, while hiding sensitive personal information. In order to
obtain a more complete contour around a person, an adap-
tive thresholding scheme using edge histograms is used as
well as background subtraction which limits the edge ex-
traction to foreground masks and thus avoids distraction of
the viewer’s eyes to background structures.
Categories and Subject Descriptors Figure 1: original PEViD frame (left) and privacy-
filtered result (right)
K.4.1 [Computers and Society]: Public Policy Issues
Keywords
2. SYSTEM DESCRIPTION
Privacy preservation, video analysis, obfuscation, adaptive
The proposed system uses a background model in order
edge detection
to exchange the information within a person’s contour by
recent background information (see Fig. 1). The bound-
1. INTRODUCTION ing boxes which are provided as manually-annotated ground
With the increasing usage of Video Surveillance systems truth for the purpose of MediaEval are used in order to iden-
and a continuously growing amount of CCTV data being tify where the privacy filter needs to be applied in the image.
recorded and looked at, the need to apply privacy protection In these regions, edge detection is performed, and the result
techniques in this area rises as well. In order to increase is blended into the background image. As the bounding
acceptance of CCTV-cameras among people being observed, boxes for objects could also be provided automatically (e.g.
it is of special importance to ensure their personal rights be by human detection algorithms or trackers such as [3]), our
not violated, while the surveillance system is still able to system does not necessarily depend on manual annotations.
work and security staff is able to identify critical events in
a video stream. The MediaEval 2013 Visual Privacy Task 2.1 Double Background model
[1] addresses this issue and provides an evaluation based on Our algorithm uses two background models: a) a standard
the PEViD Dataset [5]. Gaussian-Mixture Model (GMM) similar to [4] in order to
obtain accurate foreground masks and b) a very simplified
background model which essentially consists of only a single
Copyright is held by the author/owner(s). frame. The latter is used to maintain a very recent back-
MediaEval 2013 Workshop, October 18-19, 2013, Barcelona, Spain ground information. This allows the viewer (or an analytics
algorithm) to identify e.g. objects left in the scene or graffiti Objective evaluation
sprayed on the walls. As one of the aims of our system is Adaptive edge filter Average (9)
to give the user a clear impression of what is going on in Intelligibility 0.358355 0.502378
the scene, it is crucial to provide very recent and fast back- Privacy 0.693835 0.664903
ground information. We therefore just use a simple on-line Appropriateness 0.368627 0.56048
learning method which adapts the background pixels for the
regions without objects as follows: Subjective evaluation
Adaptive edge filter Average (9)
BGi (x, y) = (1 − α) · BGi−1 (x, y) + α · I(x, y) (1) Intelligibility 0.678333 0.655741
Privacy 0.683750 0.683843
In order to adapt the background model quickly to the Appropriateness 0.532500 0.492130
new frame, we use rather high values for α (e.g. α = 0.35)
and update this model every 5th frame. This also has the Table 1: Evaluation results for the proposed method
advantage that the system can cope quickly with slight cam- compared to 9 other methods in the workshop.
era movements. In some of the PEViD videos, the camera
shakes a bit and the image is shifted for some pixels. While
first the background image in our system gets a bit blurry a person with background information, it is obvious that
in these situations, after a few frames our system maintains a standard person detection and tracking method which is
the current background and is not disturbed by older values usually based on extracting edges by HoG or wavelet filter-
anymore. The GMM uses a much slower learning rate and ing and is trained on normal videos cannot have the same
is only used to mask out background parts within the given performance as in unfiltered images. To apply the filtered
region of interest (ROI) - a task for which the previously videos for automated video analytics the respective prepro-
described background model b) alone would be to simple. cessing of the methods has to be adapted and we thus rec-
ommend a new training step based on feature extraction on
2.2 Adaptive Edge Detection the filtered images. However, based on the results of the
We use Canny edge detection [2] in order to extract the subjective evaluation, the filter is especially appropriate for
silhouettes of the persons in the image. After a noise reduc- display in video surveillance systems that are guided and
tion step, this algorithms computes the gradients in the im- evaluated by human operators.
age and performs thresholding with hysteresis. A common
problem for this algorithm is the choice of the two thresh- 4. CONCLUSIONS
olds T1 , T2 used in the hysteresis step. While T1 sets the In this paper we showed how adaptive edge detection can
minimum edge level accepted for the starting points, T2 de- be used to preserve the privacy of people in CCTV videos.
termines which edge level must be kept during the hysteresis. The subjective test demonstrates that the proposed filter
Usually, it is hard to set general values which give satisfying outperforms most of the other privacy filters proposed in the
results for arbitrary video sequences. MediaEval 2013 challenge. Thus, our method is of special
In our system which has to work under varying conditions interest for semi-automatic video surveillance systems which
(indoor, outdoor, different lighting, changing weather...), we need to hide sensitive personal information while preserving
propose to adapt the threshold according to the gradient the context of actions.
histogram in the ROI. For the given ROI, we compute the
absolute values of the gradient information and set up a 5. ACKNOWLEDGMENTS
histogram over every ROI. Using the assumption, that most
of a human’s silhouette can be recovered by the highest 15 % This work was supported by the European Commission
of the gradient information, we can thus find T1 for the under contracts FP7-261743 VideoSense.
Canny algorithm. In order to close holes in the silhouette,
T2 is set to T2 = 0.9 · T1 . This automatic choice of T1 and T2 6. REFERENCES
adapts well to most scenes in the data set and gives a good [1] A. Badii, M. Einig, and T. Piatrik. Overview of the
contour information of a walking person in most cases. mediaeval 2013 visual privacy task. 2013.
[2] J. Canny. A computational approach to edge detection.
3. EVALUATION RESULTS IEEE Trans. Pattern Anal. Mach. Intell., 8(6):679–698,
1986.
The evaluation of the proposed filter is based on objec-
tive and subjective metrics described in [1]. The results [3] V. Eiselein, T. Senst, I. Keller, and T. Sikora. A
are shown in Table 1. The proposed filter score has been motion-enhanced hybrid probability hypothesis density
compared with the average score of all 9 participants of the filter for real-time multi-human tracking in video
MediaEval Visual Privacy Task challenge. The results are surveillance scenarios. In 15th IEEE International
very promising in terms of the privacy behavior where the Workshop on Performance Evaluation of Tracking and
adaptive edge filter is competitive with other participants. Surveillance (PETS), pages 6–13, 2013.
In regards to the intelligibility and appropriateness the [4] R. Heras Evangelio, T. Senst, and T. Sikora. Detection
results of the subjective and objective evaluations related to of static objects for the task of video surveillance. In
the average of all participants are contrary. While for these IEEE Workshop on Applications of Computer Vision
categories the score of the objective results are less than the (WACV), pages 534–540, 2011.
average, the score of the subjective results is higher than the [5] P. Korshunov and T. Ebrahimi. PEViD: privacy
average score. As the proposed filter extracts and reprojects evaluation video dataset. In Proceedings of SPIE
the edge information into the image and fills the interior of Volume 8856, volume 8856, 2013.