Big Brother is watching you!
Danish Agency for Science, Technology, and Innovation, CVMT, 2007-2010
The
constantly decreasing price of surveillance cameras and the notion that
more surveillance equals more security have lead to a very large number of
surveillance cameras mounted in both public and private spaces. It is for
example estimated that 200,000 surveillance cameras are in operation in
Automated analysis of such video streams is a hot research topic, but so far without much success for general purposes. One way forward is to concentrate on specific applications and possibly accept constrained scene conditions. Face recognition, or at least generation of frontal facial images of persons from surveillance videos, is one important “specific” application worth pursuing.
Commercial face recognizers are currently in operation around the world. They operate by matching a camera image with known faces in a database. For controlled situations, e.g., for access control, persons face the camera and good quality images can be captured for high performance face recognition. For video recorded by surveillance cameras current
state-of-the-art recognizers fail due to poor quality of the images, i.e. low resolution, motion blur due to head movement, non-frontal face image, strange facial expressions etc.This project will aim at bridging the gab between poor quality surveillance video and technologies processing faces (like a face recognizer), which require good quality images of the face. A successful project will allow for, e.g., automatic and real-time recognition of faces in a standard surveillance camera setup.
The project contains three parts; 1) figure-ground segmentation of moving objects in video, 2) control active ptz cameras to focus on and capture video of faces, and 3) obtaining a good quality face image.
HERMES - Human-Expressive Representations of Motion
and Their Evaluation in Sequences
A 6 partners EU/IST/STREP project (IST-027110), April 2006 - August 2009
HERMES concentrates on how to extract descriptions of
human behaviour from videos in a restricted discourse domain, transform this
into written text, and allow to synthesis a dynamic scene based on a textual
description. Discourse domains are for example pedestrians crossing
inner-city roads and pedestrians approaching or waiting at stops of buses or
trams. These discourse domains allow to explore a coherent evaluation of
human movements and facial expressions across a wide variation of scale.
This general approach lends itself to various cognitive surveillance
scenarios at varying degrees of resolution: From wide-field-of-view
multiple-agent scenes, through to more specific inferences of emotional
state that could be elicited from high resolution imagery of faces. HERMES
aim to consider how cooperating pan-tilt-zoom sensors can enhance the
process of cognition via controlled responses to uncertain or ambiguous
interpretations. The system will be exposed to video recordings from
different parts of Europe in order to prevent over adaptation to local
habits and, in addition, to learn systematically occurring differences
between pedestrian habits in different countries. The system's explanatory
and arguing capabilities are expected to ease an assessment of its strengths
and weaknesses.Within the HERMES project CVMT will primary be working on
extracting body information from sequences, for example arm gestures, head
pose, and body pose. HERMES is a 6 partners EU/IST/STREP project
(IST-027110).
Official HEREMS homepage
CVMT's HEREMS
homepage