In this paper, we describe a privacy filter proposed for VPT 2014 in an attempt to provide a context-aware solution. The proposed solution comprises three different techniques applied to the face, skin, and body regions separately. The proposed combination of filtering techniques aimed to produce an adaptive solution and provide an example of a context-aware-like privacy filtering capability. The results demonstrated the effectiveness of the proposed techniques in maintaining a high level of Intelligibility and retaining the appeal of the video i.e. Pleasantness. However the still identifiable gender and race of certain individuals contributed to the perception of lower levels of privacy of the person in the case of some of the video frames.
The Visual Privacy Task at MediaEval 2014 [
We proposed a privacy filter which primarily aims to achieve a balance in the well-addressed Privacy-Intelligibility trade-off. In addition, the Pleasantness and the appropriate filter application criteria are also considered alongside the real-time applicability. Accordingly our three different filtering techniques for face, skin, and body regions of the subject featured in the video were applied as follows: Face filter (H): The face being a highly identity revealing region has been ranked high (H) for privacy protection compared to other body parts. Accordingly, we sought to ensure the anonymity of the visible face. First an empirical threshold was set to examine the minimum size of the face in which it could be identifiable (i.e. 50 pixels in our case, rightmost snapshot in Figure 1). Below the said threshold, a simple median blur filter proved sufficient to protect the person’s identity. Once the face size exceeded the threshold, a key point detector was applied on the face region followed by adaptive colour quantisation and circle texturing to produce the effect as shown in Figure 1.
Skin filter (M): The exposed skin regions could provide sufficient information to enable the detection of the ethnicity of the subject. Skin provides a focus of attention. Exposed skin regions, e.g. hands, are also important in activity recognition and detection of weapons. Therefore, morphologic changes or non-homogenous colour changes are not suitable in this case. To manipulate the skin region in a unified fashion, we reduced the colour saturation and luminance values in the R, G, and B colour channels separately within the oriented boxes enclosing the skin regions. The filtered skin regions were still recognisable. There are only two identifiable skin colours: light-like and dark-like. The skin texture which is responsible for the skin attractiveness was also eliminated.
Person filter (L): The last stage of privacy protection is applied to the bounding box enclosing the subject region. This region is ranked low in the provided annotation. An edge-based analysis is implemented on the foreground region(s) within the subject bounding box. The procedure begins with morphological operations to enhance the foreground mask to reduce the background noise and minimise the holes in the foreground region. Canny edge detector is subsequently applied and further refined to eventually draw the subject’s contour. The final effect produced as depicted in Figure 1 is the result of a distance transformation which calculates the distance of each pixel of the resulting binary contour map with the closest zero pixel in the image. OpenCV implementation based on [4] was used which calculate the Euclidean distance to the nearest zero pixel consisting of basic shifts: horizontal, vertical, diagonal, or knight’s move. A mask of size (5 X 5) was used for the best results.
A subjective evaluation consist of three (3) streams were
conducted and the performance of the proposed privacy protection
solution is examined in terms of the defined criteria namely,
Privacy, Intelligibility, and Pleasantness as described in [
Stream_1 Stream_2 Stream_3
Intelligibility 75.10% 84.07% 73.27%
Privacy 42.80% 35.55% 39.01%
Pleasantness The proposed solution scored above the average for Intelligibility criterion and well above the region of 70% in the three streams which ensures that the processed video will still serve the main purpose of CCTV security objective.
However, the Privacy scores were slightly below the median which is in general below the value of 50% in this competition. One possible explanation of the overall low score of Privacy is the fact that this criterion has been measured based on the ability to identify the gender and the ethnicity of the person which could be hard to conceal without significantly manipulating the person
In this paper we have proposed a video privacy filter using a combination of filtering techniques to simulate a context-aware solution. The filter aims to achieve the highest privacy with minimum content distortion and viewer distraction. The obtained scores were comparable to the average values of the scores for all the privacy filtering solutions as proposed for the Visual Privacy Task 2014. One possible future work would be the addition of person re-identification test to be included in evaluating the Privacy criterion as an important aspect.
We would like to thank Lucas Teixeira and Kevin Lelu for their valuable inputs. This work was supported by the European Commission under contracts FP7-261743 VideoSense project.