Image-based motion analysis

advertisement
SPIE Newsroom
10.1117/2.1200604.0192
Image-based motion analysis
Jack Sanders-Reed
Software has been developed to perform quantitative analysis of moving
targets in sequences of images: abilities include target tracking, multisensor triangulation, and model-based pose estimation.
The quantitative motion analysis of sequences of images poses
a challenge in disciplines ranging from automotive crash-test
analysis to biological motion studies and military test-range instrumentation. The images can be created using high-speed or
standard video cameras in a wide variety of formats. In many
cases, analysis requires not only the image data, but various supporting (ancillary or meta) data that can vary over time: these
include camera location and pointing, camera frame rate, or
radar range data.
Both the processing and analysis requirements can vary
widely, from simple, single-camera 2D (x-y) tracking to
geometric-lens-distortion correction, model-based pose estimation, multi-camera triangulation, target-shape analysis, and
radiometric analysis. Unique to image-based motion analysis is
the need to deal with sequences of images—possibly exploiting
ancillary data—to develop time histories of the desired features
of objects, such as position, shape, and intensity.
In traditional motion analysis, each target in each frame is
manually designated. While this is acceptable for short sequences with only a few targets, it becomes impractical otherwise. To address the needs and challenges of image-based
motion analysis, I developed the Visual Fusion motion analysis
software package (see Figure 1): a software analysis tool suitable
for a wide spectrum of users with expertise ranging from personnel without degrees to Ph.D.-level investigators in a broad
range of application fields. It also required the development and
integration of many kinds of image-analysis technology with
much wider applications: these include multi-target tracking,
triangulation-error analysis, automatic outlining of complexshaped targets, and the correction of complex optical distortions.
Motion analysis can be seen as consisting of the following
steps: image pre-processing; target detection and feature extraction (shape, moments, centroid, leading edge, intensity);
correction for optical distortion; calibration from focal-plane
pixel space to units such as azimuth and elevation or x-y distance units (e.g. meters); calibration from pixel grey scale to
Figure 1. Motion-analysis software user interface showing target
detection and tracking, plotting of position versus time, Fourier transform to determine frequency, and curve fitting to provide more exact
estimation.
Figure 2. Multi-target tracking and surface relief, alternative image
display.
radiometric intensity units; and detect-to-track association to develop time histories or tracks (see Figure 2). This can be followed
by multi-sensor triangulation, model-based pose estimation,
determination of the motion of one object relative to another,
numerical derivatives or curve fitting to estimate velocity and
acceleration, or frequency analysis. Each of these involve a variety of adjustable parameters or even algorithms, necessitating
the development of a user interface that provides simple operaContinued on next page
10.1117/2.1200604.0192 Page 2/2
SPIE Newsroom
tions for many applications, but allows the experienced user to
adjust parameters for more challenging problems.
A basic requirement for most applications is to be able to manipulate and visualize both the image and ancillary data. This
means the software must be able to read a wide range of different image and data formats and allow the user to easily view the
imagery or data. Some of the basic tools available at this level
include the ability to adjust gray scale to enhance features of
interest and perform operations such as edge enhancement or
convolution filtering and electronic zoom. Additionally, it is
sometimes informative to display imagery in a 3D relief format
where bright regions appear as higher elevations.
Basic target detection involves pre-processing, followed by
threshold detection (dark targets on light or light targets on
dark), and then either correlation detection (in which the software finds the various user-selected targets) or detection of
pre-defined targets, such as standard quadrant patterns. Simple radial polynomial optical system correction can be used to
correct for wide angle lens distortion. However, I elected to
implement a more robust bi-linear interpolation on a grid of
points, thus allowing correction of off-axis imaging systems or
external windows, such as curved windscreens. Robust multitarget detect-to-track assignment (basically a modified globalnearest-neighbor algorithm) is used to develop time histories of
targets.1–4
Passive imaging sensors are 2D-angle-only sensors. To generate 3D target tracks, one must either bring in range data from
radar and perform multi-sensor triangulation, or use modelbased pose estimation. While developing the former I was surprised to find that, although triangulation itself is well known,
little work had been published on error analysis to determine
the accuracy of triangulation results. As a result, I developed the
theory for triangulation error analysis 5 and some methodology
for practical application of the theory.6
Though model-based pose estimation is a well known
technique,7 classical approaches suffer from the need for an initial pose guess. This is impractical in many situations and becomes unstable in important scenarios involving approximately
planar objects. Recent work8, 9 has greatly improved performance
in this area.
Image-based motion analysis has recently moved from manual frame-by-frame point and click analysis to much more
automated and sophisticated computer analysis. This brings together a rich set of image-processing and numerical analysis
techniques. While commercial software is available, an understanding of the underlying techniques and principles can be
valuable in accurately obtaining the most information from these
tools.
Author Information
Jack Sanders-Reed
Boeing-SVS
Albuquerque, NM
http://www.swcp.com/∼spsvs
Jack Sanders-Reed is a Technical Fellow with the Boeing Company, specializing in image-based motion analysis, image exploitation, and electro-optic systems. Developer of the Visual
Fusion motion analysis software package, he has more than 20
peer reviewed and conference publications to his name, has published numerous papers in Optical Engineering and the Proceedings of the SPIE, and teaches the image-based motion analysis
course at the SPIE Defense & Security Symposium.
References
1. S. S. Blackman, Multiple-Target Tracking with Radar Applications, Artech House,
1986. ISBN 0-89006-179-3
2. Y. Bar-Shalom and X.-R. Li, Multitarget-Multisensor Tracking: Principals and Techniques, 1995. ISBN 0-9648312-0-1
3. J. N. Sanders-Reed et al., Multi-target tracking in clutter, Proc. SPIE 4724, April
2002.
4. J. N. Sanders-Reed, Multi-target, multi-sensor, closed loop tracking, Proc. SPIE 5430,
April 2004.
5. J. N. Sanders-Reed, Error propagation in two-sensor 3D position estimation, Opt.
Eng. 40 (4), April 2001.
6. J. N. Sanders-Reed, Impact of tracking system knowledge on multi-sensor 3D triangulation, Proc. SPIE 4714, April 2002.
7. E. Trucco and A. Verri, Introductory Techniques for 3D Computer Vision,
pp. 26–32, 279–292, Prentice-Hall, 1998. ISBN 0-13-261108-2
8. D. F. DeMenthon and L. S. Davis, Model-based object pose in 25 lines of
code, International Journal of Computer Vision 15, pp. 123–141, June 1995.
http://www.cfar.umd.edu/ daniel/
9. D. Oberkampf, D. F. DeMenthon, and L. S. Davis, Iterative pose estimation
using coplanar feature points, CVGIP: Image Understanding 63 (3), May 1996.
http://www.cfar.umd.edu/ daniel/
c 2006 SPIE—The International Society for Optical Engineering
Download