PRIVACY PROTECTION FOR VIDEO
CONFERENCE
M.Vijay Venkatesh, Jian Zhao and Sen-ching S. Cheung
Objective
The
aim of this technology is to provide privacy for selected individuals in a
video-conferencing environment by obfuscating their presence.
Problem setting and approach
The figure below shows a typical video conferencing scenario.
Classes of participants:
Active Participant (AP) :
People participating interactively
Passive Participant
(PP) : People who passively observe and
whose privacy needs to be protected
Assumptions:
1. We restrict the movement of the PPs
so that they never occlude APs. This
process alleviates the need to perform the time consuming process of dynamic
foreground inpainting. We only need to locate the position of the PP’s at all
time instances and replace them with the background.
2. We assume that there is no occlusion between a moving PP and a moving AP.
This is a reasonable assumption for video-conference because the participants
are usually disjoint in the video and the occlusion occur only when some
participants enter or leave the conference as shown by the red and green arrow
paths in Figure 1
Approach:
We develop a fast object
segmentation scheme by combining a stationary background model to detect all
the participants and an adaptive background model to identify moving
participants. Based on our second assumption, each of the foreground blobs
detected by the adaptive background model corresponds to a single moving
participant. All the blobs are then tracked from frame to frame. Using our
first assumption, the genesis of the object track can tell us if it corresponds
to an active or passive participant. Objects classified as passive participants
are replaced by the background frame obtained from the adaptive background
model. The block diagram shown below describes the segmentation using
background subtraction and tracking for object identification.
Algorithm flow of the system
We present here some images
and video of our experiments below.
Images:
Fig 3: Privacy protection under 3 different scenarios. The top row
shows the result of the stationary PP. The middle row shows the result of
moving AP while the bottom row shows the result of moving PP.