This paper describes a physics-based technique for protecting the privacy of people in videos as defined by the MediaEval 2013 Visual Privacy task. We propose a physics-based approach which estimates the full spectrum of the surface spectral reflectance from the video. Whereby the wavelength which corresponds to the global minimum of the spectral curve (an intrinsic feature of the material) at a pixel is calculated and converted to RGB values which are used to filter pixels that belong to a moving object. This effectively implements visual privacy protection by replacing foreground pixel colour by another which is related to intrinsic optical properties of the original pixel. Both objective and subjective evaluations are performed using both video analytics algorithm and user studies in order to evaluate the proposed technique.
In this challenge, we propose a physics-based technique to protect privacy in surveillance videos. Privacy protection approaches may be classified depending on the image representation used as physics-based or non-physics-based. Non-physics based approaches use one of the known colour spaces, with no explicit physics underpinning, as a cue to model the object. The word physics refers to the extraction of intrinsic features about the materials contained in the object based on an understanding of the underlying physics which govern the image formation. This process is achieved by applying physics-based image formation models which attempt to estimate or eliminate the illumination and/or the geometric parameters in order to extract information about the surface reflectance.
Conventional video cameras, analogous to a retina, sense reflected
light so that colour values are the integration of the product of
incident illumination power spectral distribution, object’s surface
Copyright is held by the author/owner(s).
spectral reflectance and camera sensors sensitivities. Humans have
the ability to separate the illumination power spectral distribution
from the surface spectral reflectance when judging object
appearance, such ability is called colour constancy [
Our proposed privacy-protection technique consists of two steps. First, a change detection step segments moving objects (the foreground). Second, a filter changes foreground pixels, so as to achieve visual privacy, by converting the surface spectral reflectance at these pixels into RGB values as perceived by humans. 2.1
Spectral-360 [
In order to build a computational physical model, the illumination
is estimated by segmenting areas in the image which represent
high specularities (highlights); McCamy’s formula [
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
metrics [
This paper argues that image formation models offer interesting new alternative physics-based cues for privacy protection 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 such models, and hence possible unfeasibility of real-time implementation. This challenge was tackled, firstly, by choosing an appropriate reflection model, the dichromatic reflection model. Secondly, by setting a feasible set of assumptions for such model which best match the reduction of model complexity and does not contradict with real-world operational conditions. In this paper we have proposed a physics-based approach which segments the moving objects and then estimates the full spectrum of surface spectral reflectance from the video, where the wavelength that corresponds to the global minimum is calculated and converted to RGB values as perceived by humans. This effectively replaces foreground pixel colours by others, as a simple privacy protection mechanism. Our ongoing work is investigating enhanced pixel replacement schemes, beyond the obfuscation mechanism described here, so as to protect privacy while minimizing the change of the semantic content of the image.
Team: CIISS1
Our Score 0.453317 0.812541