|
Video Tracking
Inexpensive, highly capable video cameras and networks are rapidly being deployed throughout the world for a
wide variety of law enforcement, security, and military applications. Many of these applications require systems
that automatically track objects of interest over extended time periods in rural and urban terrain. SET’s team
of computer vision and image processing experts have developed award winning solutions for automatically tracking
vehicles and humans with networks of visible and infrared video sensors. Our tracking solutions employ SET’s
software camera stabilization technology, and are designed for robust, real-time performance in challenging
ground and airborne environments. 
Video Stabilization and Mosaicking
Unmanned Air Vehicles capabilities are being extended through the proliferation of micro-air vehicles (MAVs),
which may be tasked by individual warfighters in the field to provide immediate situation awareness. These small,
portable platforms present new challenges for video stabilization algorithms. Traditional methods for recovering
ego-motion parameters do not perform well for MAVs due to jitter caused by poorly constrained platform motion and
poor video quality. SET has designed and demonstrated a powerful video stabilization and mosaicking algorithm that
provides robust MAV video stabilization. In contrast to competing solutions implemented on custom hardware, our
solution runs real time on a general purpose PC processor. Our video stabilization software module The module
includes multiple selectable user displays for optimal task based utilization, and incorporation of IMU/GPS data
when available. Written as a software “.dll,” the module can be readily incorporated by ground-station vendors
into their operator interface systems.
Intelligent Video Monitoring
Conventional video security systems generate large numbers of costly false alarms. They also lack the ability to
intelligently filter relevant information from large quantities of extraneous video data, creating cognitive
overload conditions in even the most proficient human operators. SET is addressing these problems by developing
intelligent video monitoring technologies that automatically monitor specific behaviors in predefined regions of
interest. Our technology allows an operator to quickly identify desired behaviors and interest regions, such as
the red passenger deplaning corridor in the associated picture. Once defined, our system automatically monitors
the video stream to detect deviations from allowable behavior patterns in and around these regions.
3D Modeling
Automatic acquisition of 3–D models of humans, vehicles, and structures is in high demand for gaming, navigation,
and computer vision systems. However, manual construction of such models takes far too long to be of practical
use, and specialized sensors such as Interferometric Synthetic Aperture Radar (IFSAR) and Laser Induced Differential
Absorption Radar (LIDAR) are not feasible for many applications. SET researchers are developing shape-from-motion
techniques for automatically constructing detailed 3D models from image sequences captured by a single, consumer
grade, handheld video camera equipped with a commercial inertial measurement unit. These techniques will allow
for the rapid and inexpensive acquisition of 3D models in unconstrained indoor and outdoor environments.
Hyperspectral Target Detection
Broadband sensors are often insufficient to enable accurate automatic characterization of scene content. Hyperspectral
imaging sensors provide high dimensional spectral responses at each spatial location that yield a rich feature set to
enable accurate scene characterization. The voluminous data typically requires large computational complexity for scene
analysis. SET, in conjunction with the Johns Hopkins University Applied Physics Laboratory (APL), has developed and
demonstrated technologies that have significantly advanced computation and operational performance for HSI detection.
A new non-parametric approach has been developed to process this data in real time in a fashion that provides detection
accuracy surpassing existing technologies. This represents two orders of magnitude reduction in computational complexity
relative to current approaches. The system facilitates real-time operation (5-10 Hz for 256x256x100 hypercubes) on a COTS
PC platform. Subpixel target detection has also been demonstrated in a fashion that enables recognition of targets from
space-based sensors with greater accuracy than existing methodologies.
|
