Proceedings Article | 5 May 2009
KEYWORDS: Video, Sensors, Cameras, Detection and tracking algorithms, Video processing, Target detection, Photovoltaics, Field programmable gate arrays, Data modeling, Video surveillance
A low cost, lightweight, easily deployable imaging sensor that can dependably discriminate threats from other activities
within its field of view and, only then, alert the distant duty officer by transmitting a visual confirmation of the threat
would provide a valuable asset to modern defense. At present, current solutions suffer from a multitude of deficiencies -
size, cost, power endurance, but most notably, an inability to assess an image and conclude that it contains a threat. The
human attention span cannot maintain critical surveillance over banks of displays constantly conveying such images from
the field.
DigitalTripwire is a small, self-contained, automated human-detection system capable of running for 1-5 days on two AA
batteries. To achieve such long endurance, the DigitalTripwire system utilizes an FPGA designed with sleep
functionality. The system uses robust vision algorithms, such as a partially unsupervised innovative backgroundmodeling
algorithm, which employ several data reduction strategies to operate in real-time, and achieve high detection
rates. When it detects human activity, either mounted or dismounted, it sends an alert including images to notify the
command center.
In this paper, we describe the hardware and software design of the DigitalTripwire system. In addition, we provide
detection and false alarm rates across several challenging data sets demonstrating the performance of the vision
algorithms in autonomously analyzing the video stream and classifying moving objects into four primary categories -
dismounted human, vehicle, non-human, or unknown. Performance results across several challenging data sets are
provided.
