In this paper, a pixel-based method for personal anonymisation in visual surveillance applications, is presented. The proposed method tries to tackle the problem of balancing the intelligibility with privacy, by including some form of contextual information in the anonymisation process. Objective and subjective evaluations show promising results in intelligibility and appropriateness, but they also show that privacy could be further improved.
Proactive e orts to ensure citizens' security lead to
widespread adoption of invasive video surveillance systems. The
ever-increasing amount of recorded information poses a
direct threat to citizens' privacy and their right to preserve
their personal information. Thus, a general social concern
has emerged for the loss of privacy, demanding new
approaches to preserve and protect it, ensuring their anonymity
and freedom of action whilst maintaining the surveillance
performance. We propose a new approach to preserve
persons' identity in visual surveillance information for the
MediaEval Privacy Task [
When designing the anonymisation method, the main goal is to maximise the privacy of the person while maintaining a very good intelligibility. With this in mind, we listed the possible types of ROIs and the information they carried to identify a person. After this, we classi ed the possible types of Regions of Interest (ROIs) into three categories, each category containing more information than the previous.
The core of the method applies a pixelisation lter to all ROIs, but additional steps are performed to each \level", becoming more and more speci c in the type of lter applied. The description of the di erent levels and the lters applied are described in the following subsections, followed by a brief discussion of the method. 2.1
This level carries the less information about the person being anonymised, so the more general lter is applied in this step, which is the pixelisation lter. In our tests, a pixel size of 24x24 yielded the best results for the presented scenarios, but di erent conditions might need a di erent pixel size, and the proposed method is exible enough to allow a di erent pixel size via a parameter, so it could be changed easily to adapt to a di erent scenario. 2.2
In this step, a skin colour detector in the hue, saturation, and value (HSV) colour space is used to detect and change the subject's skin. We used a xed range in the colour space in which all pixels within this range are considered skin. Since we are only applying this to skin ROIs, the risk of false positives is low, while the amount of true positives is maximised. In our experiments, this range was H 2 [0 ; 28:23 ]; S 2 [0:04; 1]; V 2 [0; 1]:
The colour of each detected skin pixel is then changed to a single colour. This ensures everyone to have the same skin colour, e ectively concealing their true race. In this particular scenario, we choose the colour (14.11 ; 0:31; 0:39):
This step also introduces other parameters to further adapt the method to di erent scenarios, which are the skin detection range and the skin colour changing. 2.3
The face is considered apart from the skin to improve not
the privacy but the intelligibility. In this step, after the
skin colour change and the pixelisation, edges of the face,
detected by the Canny Edge Detector[
When developing the anonymisation lter, several things
were considered. Pixelisation, as the main lter, was
considered because it has shown to have a very good balance of
privacy and intelligibility [
The lter was implemented in C++ using the OpenCV library for image processing and Xerces-C++ to load the ROIs from an XML le. In particular, the implementation uses the parallel programming paradigm to perform the different stages of the algorithm concurrently to achieve a near real-time performance.
Results of the objective and subjective evaluations are shown in Figures 2 and 3 respectively. The gure shows the score of the proposed method paired against the average score of the other methods presented in the challenge. The use of pixelisation combined with the face's edges combined paid of in the high score seen in intelligibility and appropriateness, while the change of the colour skin did not a ect this measures. The evaluation results also show that there is still room for improvement in the privacy preserving aspect of the lter.
A pixel-based anonymisation method of visual surveillance information has been presented for the MediaEval 2013 privacy task. The proposed method applies di erent lters depending of the level of privacy information carried in a ROI. Objective and subjective evaluations show that the lter performs very well in intelligibility and appropriateness but there is still opportunities to improve the privacy preserving aspect of the lter. Because the lter is developed in a modular way, di erent parts of the lter can be improved separately. The inclusion of parameters also allows to an improvement in results without modifying the method itself. For example, the pixelisation lter could produce smaller or bigger pixels. The range of colour in the colour detection, and the target colour in the skin colour change, could be all changed to include more (or less) tones of skins or to produce a darker or lighter skin tone. Finally, the Canny Edge Detection algorithm used in the face introduces it's own set of parameters to detect the edges. More advanced adjustments to the lter include the colour change part of the method. For example an advance colour transfer technique could be used to produced more natural results. The edge detection could also be improved by softening the detected edges to produce a more natural result, or using a di erent method to detect the edges altogether.