This paper describes Imcube's participation in the DroneProtect Task of MediaEval 2015, which aims to obscure privacy-concerned image regions in videos sequences captured with drones. As a result persons and vehicles should be unrecognisable, but the semantic meaning of the scene should remain understandable to viewer. We use an approach which replaces the privacy-concerned region with an automatically computed composite of inpainted background and foreground contour. Before obfuscation, the image region to be hidden is extracted and steganographically embedded into the processed frame leading to a reversible solution. The evaluation shows that the developed solution achieves good privacy protection while preserving the intelligibility and aesthetic pleasantness of the original video.
Since drones become a ordable, these devices are increasingly used for security applications. Due to their exibility videos captures by a drone may contain highly sensitive personal data. 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 DroneProtect Task of MediaEval 2015 deals with
the problem of privacy protection in dynamic surveillance
videos [
A common way to protect privacy in images and videos is
to apply techniques such as blurring or masking, as shown
in [
Our approach combines both masking and steganography to obtain a visually appealing obfuscated video and to have the possibility to recover the original frame. An example is shown in Figure 1. Inpainting, edge detection
DCT-based steganography 2.1
Since the videos captured by the mini-drones [
In order to remove the original foreground object, the
region of interest needs to be lled with reasonable
background information. We rely on the inpainting algorithm by
Telea [
The obfuscated region of interest is obtained by blending the extracted foreground contour with the reconstructed background texture. Depending on the application the contour may be emphasized with di erent colours. 2.2
Since the masking algorithm itself is irreversible, the
original image data contained within the region of interested
must be embedded into the obfuscated image. Therefore, we
make use of a steganography library, which is based on the
F5-algorithm [
The video sequences of the DroneProtect dataset [
The evaluation of the obscured videos took place using subjective procedures. Two groups of subjects with di erent experience in surveillance applications were asked to survey the videos and respond to questions concerning the content. Based on the answers to these questions three di erent metrics (privacy, pleasantness, intelligibility) and the deviation between were computed. The average scores of the di erent subject groups (experts, novices, overall) for all 38 videos are summarized in Table 1. In the following we will analyse the results for the di erent metrics and di erent subject groups.
The privacy metric measures how well the identity of the persons and vehicles was protected through the obfuscation or in other words how di cult the obfuscation made the identi cation of a person or vehicle by hiding relevant visual information. The proposed method achieves a medium overall score (0.50) which suggests that even though only the contour of the object of interest is preserved in some cases it can still be identi ed. A deeper analysis of these cases is needed to identify potential improvements.
Intelligibility stands for the ability of classifying objects and actions within a video sequence and evaluates how well the activities within a scene are preserved even if the object of interest is obfuscated to prevent its identi cation. The proposed method achieves a good overall score (0.61) which shows that contour information alone is enough to understand most of the semantics of a scene from a surveillance perspective. A deeper analysis of the individual videos is needed to understand what additional information is needed to improve the intelligibility further.
Pleasantness evaluates the in uence of the obfuscation method on the visual quality of the video or by how much the quality of the video is degraded by distortions and artefacts within the region of interest. The subjective score is based on the level of user acceptance. Here, the proposed methods achieves a good overall score (0.67), since the black foreground contours composed over an inpainted background blend well with the original content outside the region of interest. This score may be further improved by using a more sophisticated inpainting algorithm which reconstructs a more plausible background texture.
Since the three metrics mentioned above evaluate quite contrary requirements, the deviation evaluates di erence between these metrics by computing the standard deviation. As it can be expected from the similar scores for the di erent metrics, the proposed approach has a very good deviation score (0.09). This shows that it strikes a good balance between the di erent criteria (privacy, intelligibility and pleasantness).
Comparing the results between the di erent subject groups shows that the scores of the novices are more equal across the di erent metrics, while the experts evaluate the intelligibility and pleasantness higher and the privacy lower. This follows the intuition that experts will be able to recognize actions and identities better than novices. The higher pleasantness score suggests that experts value the content of a video higher than its quality. 4.
A reversible approach for protecting the privacy based on masking and steganographic embedding of the region of interest has been proposed and evaluation on the DroneProtect dataset. The results shows that the approach strikes a good balance between privacy, intelligibility and pleasantness. For developing potential improvements a more detailed analysis of the scores for the individual videos is needed.