This paper describes our participation in the Visual Privacy Task of MediaEval 2013, which aims to obscure human occurrence in image sequences. As a result the recorded person should be unrecognisable. We use an approach which pseudo-randomly scrambles pixels within speci ed regions. This technique is reversible and preserves the colour characteristic of each region. So, colour-based approaches will still be able to automatically distinguish between di erently dressed individuals. The evaluations of our results show that the privacy aspect got a very high score in both objective and subjective metrics. Our approach has a lack of intelligibility since it was measured by applying the Histogram of Oriented Gradients which might be fail on scrambled areas since edges are not preserved.
Video surveillance of public spaces is expanding.
Consequently, individuals are increasingly concerned about the
"invasiveness" of such ubiquitous surveillance and fear that
their privacy is at risk. The demands of stakeholders to
prevent criminal activities are often seen to be in con ict with
the privacy requirements of individuals. The main challenge
is to preserve the anonymity of the surveyed individuals and
also to ful l the stakeholders needs. The problem of privacy
protection in video surveillance is concerned in this year's
MediaEval Visual Privacy Task [
Our proposed privacy-protection approach consists of a scrambling module that obfuscates regions of detected persons. A human detector was not built since these regions were provided by the organisers.
These regions are then vertically partitioned into slices having a width of one pixel. Our scrambling module obfuscates these regions by shu ing these slices. So, an obfuscated region di ers from its original version in a changed sequence of these slices.
For this scheme, we choose a column-wise swapping pixels to recognise background structures and action. As seen in Figure 1, the action of stealing can be coarsely recognised by the extended arm visible through the horizontal black line in the scrambled image.
The shu e algorithm is based on a modi ed variant of
the Fisher-Yates method [
So, the permutation of the slices of the detected regions is determined by the order generated by a pseudo-random sequence. The pseudo-random sequence is repeatable due to the characteristics of the pseudo-random number generator. This generator produces a random, but repeatable sequence of integer numbers by specifying a certain, but xed seed. This seed is generated from the hash value of a chosen password. This value is xed for all regions in each frame and video sequence. Since the pseudo-random sequence is repeatable thru the given seed, the permutation of slices is reversible. So, the scrambled image regions can be recovered by knowing the password and the coordinates of each region. Overlapping regions were united before scrambling to ensure its reversibility. We choose for scrambling instead of cryptography to be robust against image compression artefacts and transmission errors. If these errors occur parts of the image will be also distorted, but these errors are not propagated. reversible scrambling SCD SCD
The video sequences of the VPT dataset [
The evaluation of the obscured videos took place using objective and subjective procedures. The objective metrics like face detection, object tracking and person re-identi cation were compared for each sequence pair. Three metrics are used to evaluated the results: Appropriateness, intelligibility, and privacy.
Appropriateness stands for the in uence of the obscuring lter on the human perception of the image distortion. The objective score is measured using SSIM and PSNR, here a score of ca. 0.41 is reached which is below the average. A reason might be that quite large regions are scrambled which a ect the pixel-based calculation of SSIM and PSNR. The subjective score is based on aspects such as pleasantness, distraction, and user acceptance. Here the score of ca. 0.15 is quite low resulting from distraction of the users.
Intelligibility stands for the ability of tracking obscured persons. The objective score is measured by applying the Histogram of Oriented Gradients as a human detector. Since our approach scrambles the slices of detected regions an edge-based human detector is unlikely able to detect person on these scrambled slices. So, a sub-standard value is expected: 0.35. The subjective score is 0.18.
The privacy metric concerns about the identi cation of
individuals through their faces or personal accessories. The
objective score is measured by a re-identi cation score from
a visual model [
Table 1 provides the evaluation scores for these three metrics. 4.
We propose a reversible approach for scrambling regions within images or videos to obscure its content. This approach ensures a maximum of privacy, but fails in the other aspects, like appropriateness and intelligibility. In our mind, coarse action within these scrambled regions are still recognisable as depicted in Figure 1. The clue is that these regions can be de-scrambled, if the coordinates of the regions and the password which generated the seed for the pseudo-random number generation are known.
The research leading to these results has received funding from the European Community's FP7 under grant agreement number FP7-261743 (VideoSense). We would also like to thank Martha Larson for organizing this event.