Finding a Needle in a Digital Haystack

Surveillance systems need to be able to detect distant objects for a variety of important reasons. First, distant object detection improves the system's overall coverage, allowing it to monitor larger areas without the need for an extensive network of cameras. This broader scope is essential in vast landscapes, such as border regions, industrial complexes, or sprawling urban centers, where threats can emerge from afar. Second, the ability to identify distant objects improves the system's effectiveness in early threat detection and prevention. Whether it be potential intruders, suspicious activities, or hazardous situations, spotting distant objects quickly allows for timely response and intervention. Furthermore, in various scenarios, such as traffic management or environmental monitoring, being able to detect distant objects becomes essential for ensuring public safety and averting potential disasters.

But recognizing objects from great distances has proven to be very challenging. Distant objects appear very small to a camera. They may only be a few pixels in size. This has confounded traditional machine learning object detection algorithms that excel at recognizing larger objects. It has also proven to be too difficult a problem for traditional real-time moving object detection techniques involving background subtraction because the signal-to-noise ratio is simply too low. Other methods have emerged in recent years, but each is fraught with problems, whether in accuracy, processing speed, or otherwise.

Innovation in this area is sorely needed to support a number of critical applications. Fortunately, a key advancement recently made by researchers at Sandia National Laboratories may be exactly what the field needs. They have created a software system called Multi-frame Moving Object Detection System (MMODS) that can analyze video from satellites, drones and long-range security cameras to locate and track moving objects. And those objects can be as small as a single pixel.

Whereas most present systems rely on the information present in a single frame to detect objects, MMODS leverages the wealth of information available by analyzing multiple frames at a time. The system detects regions of movement, then matches it up with other video frames to see if that movement can be correlated across frames over time. This process improves signal-to-noise ratio over time, gradually becoming more and more certain of its findings. MMODS also is useful in ignoring irrelevant background noise that is introduced by factors like the wind. Since these forces move randomly over time, they will not be flagged as objects of interest.

In a simulated environment, the team set up some scenarios to test their techniques. The tests even included single-pixel objects with a signal-to-noise ratio of one, which means that they are undetectable to both sensors and the human eye. Traditional moving object detectors were found to detect such objects on average 30% of the time. MMODS, on the other hand, was observed to be capable of recognizing these nearly invisible objects 90% of the time. Moreover, this high degree of accuracy did not come with any increase in false positive detections.

A real-world test was also conducted that involved the use of a camera installed on the peak of a mountain. This camera watched distant roads to assess how well it could determine patterns of automobile traffic. MMODS detection sensitivity was found to be improved by 200% to 500% when compared with existing options. It also showed itself to be a versatile system as it was able to detect objects that moved rapidly or slowly, and even under conditions of poor visibility.

There are a few limitations of the present system. First, an upper bound needs to be set for the velocity and acceleration of target objects to prevent wasting CPU cycles and slowing the algorithm down. In practice, this is generally a reasonable requirement, but does leave open the possibility that unexpected activity could be missed. Second, the MMODS user needs to configure how many frames are considered together in locating objects. More frames make the system more accurate, but also slow down processing, which can hinder the presentation of real-time results. The user must make a trade-off, which can lead to suboptimal performance of MMODS.

In any case, this new technology fills a major gap in current remote sensing surveillance systems. And the team is still hard at work to make it even better.