Abstract
In this paper, we propose a fingerprint analysis
algorithm based on using product manifolds to create
robust signatures for individual targets in motion imagery.
The purpose of target fingerprinting is to re-identify a
target after it disappears and then reappears due to
occlusions or out of camera view and to track targets
persistently under camera handoff situations. The
proposed method is statistics-based and has the benefit of
being compact and invariant to viewpoint, rotation, and
scaling. Moreover, it is a general framework and does not
assume a particular type of objects to be identified. For
improved robustness, we also propose a method to detect
outliers of a statistical manifold formed from the training
data of individual targets. Our experiments show that the
proposed framework is more accurate in target re-
identification than single-instance signatures and patch-
based methods.
1. Introduction
Large amounts of video data have become more
available to defense and security analysts as unmanned
aerial vehicles (UAVs) with sensor payloads are used in
many ISR (intelligence, surveillance, and reconnaissance)
operations and large numbers of networked cameras are
used for urban surveillance. Analysts are looking for
significant events/targets that may be of importance for
their surveillance missions. However, most of this video
footage will be eventless and therefore of no interest to the
analysts responsible for checking it. If a computer system
could scan the videos for potential events/targets of
interest, it would greatly lessen the analysts’ workload,
allowing them to focus only on the events of possible
importance.
Recent video surveillance paradigms for multi-object
detection and tracking [15, 23] in video sequences are
developed as an integrated framework to not only spot
unknown targets and track them, but also handle target
reacquisition and target handoff to other cameras [4, 13,
19, 24]. A common process for these paradigms can be
divided into four steps: detection of objects of interest,
extraction of features or signatures, dimensionality
reduction of the feature space, and target template or
signature matching. A variety of different tracking or
fingerprinting techniques have been proposed to address
one or more of these key capabilities.
The tracking problem can be very challenging under
unconstrained environment [7], because the target may
change its appearance quickly, due to lighting conditions
and frame rate, there might be other objects with similar
appearance, and the background could be cluttered. In
addition, occlusions and other nuisance factors of targets
can further add the complexity to the problem [1]. Similar
to single-object tracking, many existing multiple-object
tracking methods are feature-based [15, 23] or image-
patch based [17]. The latter applies sparse representation
to predefined templates and random projection to speed up
the solutions. Using motion and spectral clustering of
trajectories instead of feature extraction is an alternative
approach for tracking [10].
To handle target exiting and reappearing in videos or
camera handoff situations, there are several methods
working towards a unified framework of detection,
tracking and identifying targets. For example, [4]
demonstrates tunnel surveillance applications and shows
the ability to follow vehicles through tunnels with non-
overlapping cameras. Other methods that handle multi-
camera networks are [13] for person re-identification, [19]
for vehicles, and [24] for wide-area surveillance settings.
Compact fingerprints are extracted from targets in these
methods, such as Haar-features [4], covariance [24],
implicit shape models, and SIFT features [13]. To handle
significant scale variations and partial occlusions, [16]
detects and tracks an object from a training video
sequence by matching salient components of local
descriptors and salient locations on a new image.
In this paper, we present a video exploitation
framework to spot unknown targets, track them, segment
the target blobs to obtain useful features, and then create
signature fingerprints for the targets so that they can be
reacquired. We focus on developing a manifold-based
framework of robust and compact signature fingerprinting
for the target re-identification task. Fingerprinting is the
acquisition of a distinct description of an object of interest
Manifold-based Fingerprinting for Target Identification
Kang-Yu Ni, Terrell N. Mundhenk, Kyungnam Kim, Yuri Owechko
HRL Laboratories
3011 Malibu Canyon Road, Malibu, CA 90265
{kni,tnmundhenk,kkim,yowechko}@hrl.com
978-1-4673-1612-5/12/$31.00 ©2012 IEEE 1