=Paper=
{{Paper
|id=None
|storemode=property
|title=MediaEval 2013 Visual Privacy Task: Physics-Based Technique for Protecting Privacy in Surveillance Videos
|pdfUrl=https://ceur-ws.org/Vol-1043/mediaeval2013_submission_84.pdf
|volume=Vol-1043
|dblpUrl=https://dblp.org/rec/conf/mediaeval/SedkyCM13
}}
==MediaEval 2013 Visual Privacy Task: Physics-Based Technique for Protecting Privacy in Surveillance Videos==
https://ceur-ws.org/Vol-1043/mediaeval2013_submission_84.pdf
MediaEval 2013 Visual Privacy Task: Physics-Based
Technique for Protecting Privacy in Surveillance Videos
Mohamed Sedky Claude C. Chibelushi Mansour Moniri
Staffordshire University Staffordshire University Staffordshire University
Beaconside, Stafford, ST18 0AD Beaconside, Stafford, ST18 0AD Beaconside, Stafford, ST18 0AD
United Kingdom United Kingdom United Kingdom
+44(0)1785353260, +44(0)1785353802, +44(0)1785353634,
m.h.sedky@staffs.ac.uk c.c. chibelushi@staffs.ac.uk m.moniri@staffs.ac.uk
ABSTRACT spectral reflectance and camera sensors sensitivities. Humans have
This paper describes a physics-based technique for protecting the the ability to separate the illumination power spectral distribution
privacy of people in videos as defined by the MediaEval 2013 from the surface spectral reflectance when judging object
Visual Privacy task. We propose a physics-based approach which appearance, such ability is called colour constancy [5].
estimates the full spectrum of the surface spectral reflectance from
the video. Whereby the wavelength which corresponds to the 2. SYSTEM DESCRIPTION
global minimum of the spectral curve (an intrinsic feature of the Our proposed privacy-protection technique consists of two steps.
material) at a pixel is calculated and converted to RGB values First, a change detection step segments moving objects (the
which are used to filter pixels that belong to a moving object. This foreground). Second, a filter changes foreground pixels, so as to
effectively implements visual privacy protection by replacing achieve visual privacy, by converting the surface spectral
foreground pixel colour by another which is related to intrinsic reflectance at these pixels into RGB values as perceived by
optical properties of the original pixel. Both objective and humans.
subjective evaluations are performed using both video analytics
algorithm and user studies in order to evaluate the proposed 2.1 Change Detection
technique. Spectral-360 [6], a novel change detection algorithm has been
used. The algorithm uses colour constancy techniques to
1. INTRODUCTION computationally estimate a consistent physics-based colour
While advances in surveillance technologies are generally descriptor model of surface spectral reflectance from the video
welcomed by employers, there is a growing sense of unease and then to correlate the full-spectrum reflectance of the
amongst academics, employees and interest groups as to the background and foreground pixels to segment the foreground
ethical boundaries that such technologies may cross [1]. There is a from a static background. The rationale behind this approach is
concern that new technologies will leave people under that the segmentation between foreground and background objects
surveillance open to abuse and discrimination. In addition to this, can be done through the matching between the surface spectral
video surveillance represents a threat to individuals’ privacy and reflectance over the visible wavelengths of a reference
dignity [2]. Social scientists started more than two decades ago, a background frame and each new frame. The challenge of this
discussion about the implications of video surveillance and the approach arises from the new idea of processing the full-spectrum
privacy of people under surveillance [3]. of the surface spectral reflectance instead of the three samples
used by other colour spaces. The spectral representation uses a
The Visual Privacy task focuses on the problem of privacy
linear model, which consists of a number of basis functions (pre-
protection in video surveillance, aiming to find new technologies
trained from a set of materials) and weights (calculated for the
to ensure an appropriate level of privacy protection [4]. object under investigation). A numerical estimation of the
In this challenge, we propose a physics-based technique to protect physics-based model for image formation and the real-time
privacy in surveillance videos. Privacy protection approaches may transformation from the video to the physical parameters is carried
be classified depending on the image representation used as out.
physics-based or non-physics-based. Non-physics based
approaches use one of the known colour spaces, with no explicit
2.2 Privacy Filter
In order to build a computational physical model, the illumination
physics underpinning, as a cue to model the object. The word
is estimated by segmenting areas in the image which represent
physics refers to the extraction of intrinsic features about the
materials contained in the object based on an understanding of the high specularities (highlights); McCamy’s formula [7] is then
underlying physics which govern the image formation. This applied and the correlated colour temperature is calculated. The
process is achieved by applying physics-based image formation illumination spectral power distribution is then calculated using
models which attempt to estimate or eliminate the illumination Plank’s formula [5]. Using the dichromatic model, and by
and/or the geometric parameters in order to extract information assuming diffuse-only reflection and the existence of a dominant
about the surface reflectance. illuminant, the surface spectral reflectance is then recovered. The
wavelength that corresponds to the global minimum of the surface
Conventional video cameras, analogous to a retina, sense reflected spectral reflectance is then calculated and converted to RGB
light so that colour values are the integration of the product of values as perceived by humans.
incident illumination power spectral distribution, object’s surface
Copyright is held by the author/owner(s).
MediaEval 2013 Workshop, October 18-19, 2013, Barcelona, Spain
Figure 1, shows the recovered surface spectral reflectance for one
pixel for the foreground and the background as well as the
corresponding RGB value of the foreground surface spectral
reflectance global minimum. The rationale behind this approach is
that the global minimum of the surface spectral reflectance
represents an intrinsic feature of the material. This would allow
objects such as the face to be easily segmented and filtered.
3. RESULTS
The resulting filtered videos have been evaluated using both
objective and subjective procedures. The objective metrics
compare each pair of original and filtered video in terms of: face
detection accuracy, human body tracking accuracy, image quality
FG
metrics [4].
Table 1 provides the average objective evaluation values for the
videos analyzed. The subjective evaluation, shown in Table 2,
provides a measure for the protection of privacy of individuals,
intelligibility of activities recorded and the visual appropriateness.
As a whole, our technique shows a high degree of privacy, and an
average level of intelligibility, but low level of visual BG
appropriateness. The visual appropriateness of our proposed filter
can be adjusted for future improvement.
4. CONCLUSIONS
This paper argues that image formation models offer interesting
Figure 1. Reconstructed Surface Spectral Reflectance for one
new alternative physics-based cues for privacy protection
pixel, Foreground (FG) and Background (BG).
compared with other representations. Features such as surface
spectral reflectance have not been applied yet in the field of
privacy protection. The reason is the computational complexity of 5. REFERENCES
such models, and hence possible unfeasibility of real-time [1] Hoghai Liu, Shengyong Chen and N. Kubota. Intelligent
implementation. This challenge was tackled, firstly, by choosing Video Systems and Analytics: A Survey. In IEEE
an appropriate reflection model, the dichromatic reflection model. Transactions on Industrial Informatics, volume 9, issue 3, pp.
Secondly, by setting a feasible set of assumptions for such model 1222 – 1233, 2013.
which best match the reduction of model complexity and does not [2] M. Sedky, C. C. Chibelushi and M. Moniri. Smart Video
contradict with real-world operational conditions. In this paper we Surveillance for Workplace Applications: Implications,
have proposed a physics-based approach which segments the Technologies and Requirements. In Proceedings of The 5th
moving objects and then estimates the full spectrum of surface IASTED International Conference on Visualization, and
spectral reflectance from the video, where the wavelength that Image Processing, Spain, 2005.
corresponds to the global minimum is calculated and converted to
[3] A. Cripps. Workplace Surveillance. Intern at NSW Council
RGB values as perceived by humans. This effectively replaces
for Civil Liberties 2004.
foreground pixel colours by others, as a simple privacy protection
mechanism. Our ongoing work is investigating enhanced pixel [4] A. Badii, M. Einig, T. Piatrik. Overview of the MediaEval
replacement schemes, beyond the obfuscation mechanism 2013 Visual Privacy Task. MediaEval 2013 Workshop,
described here, so as to protect privacy while minimizing the Barcelona, Spain, October 18-19, 2013.
change of the semantic content of the image. [5] J. Ho, B. V. Funt, and M. S. Drew. Separating a colour
Table 1. Objective evaluation signal into illumination and surface reflectance components:
Theory and applications. In: IEEE Transaction on Pattern
Team: CIISS1 Our Score Average (9) Score
Analysis and Machine Intelligence, vol. 12, pp. 966-977,
Intelligibility 0.453317 0.502378 1990.
Privacy 0.812541 0.664903 [6] M. Sedky, M. Moniri and C. C. Chibelushi. Object
Appropriateness 0.284307 0.560480 Segmentation Using Full-Spectrum Matching of Albedo
Derived from Colour Images. PCT patent application
Table 2. Subjective evaluation international application no. PCT/GB2009/002829,
Team: CIISS Our Score Average (9) Score EP2374109, 2009. US patent no. 2374109 12.10.2011 US,
2011.
Intelligibility 0.561667 0.655741
[7] McCamy and S. Calvin. Correlated Colour Temperature as
Privacy 0.790000 0.683843
an Explicit Function of Chromaticity Coordinates. In:
Appropriateness 0.304167 0.492130 Journal of Colour Research & Application, vol. 17, no. 2, pp.
142-144, 1992.
1
Centre for Information, Intelligence and Security Systems (CIISS)