Latest News

Defense Artificial Intelligence: Change Detection Beyond Human Proficiency

Defense technology is ever evolving in order to find solutions to minimize or eliminate varying threats. Some of the greatest challenges arise in the effort to replace humans with artificial intelligence systems. Taking people out of the loop can reduce the risk to human life while providing the opportunity to improve upon the human thought process. Frequently, the goal is real-time decisions that can be made with reliability and accuracy, any time of the day or night. A significant aspect of this technology is known as change detection.

Change detection is a collection of methods in signal or image processing which detect an event within a stochastic process during which the statistical properties of the process undergo a change. The change in question could be suspicious human behavior in video surveillance, an imperfection found in an image of a textile cloth, or the directional change of a vehicle being tracked with a sensor. Some of these changes are easy for a human to detect, but much more difficult for a computer or sensor that digitizes its data. The goal of change detection can be to recognize changes in a data set as a person might, or even detect changes that are not observable to the human eye.

Autonomous Vehicle Mapping

Autonomous vehicles have been a research interest for defense, commercial, and industrial applications for many years. One of frequent missions of such vehicles is to map its environment, consistently surveying a region for potential changes. The ability of a robot to detect changes in its environment is of special interest to Naver Labs, an ambient intelligence technology company developing tools for autonomous driving, robotics, and artificial intelligence. Intent on developing indoor maps using an autonomous robot, they have teamed up with Here Technologies to employ their Scalable & Semantic Indoor Mapping (SSIM) technology [1].

SSIM allows the system to automatically detect changes in indoor environments and to update Points of Interest (POI) accordingly. The data collected is evaluated using AI technology which updates the service in real time. This work could extend to mapping of various other types of environments via autonomous vehicles for reconnaissance missions.

Aiding Adaptable Systems

More often than not, change detection occurs as a part of online processing for a causal system. In other words, the algorithm applied must detect a change as it reads in data. The computer only has access to past and present data while making calculations, not the whole picture. This naturally adds to the challenge of the problem, especially when it comes to applications like vehicle tracking where the system needs to adjust in real time.

Change detection is often considered an extension of adaptive filtering for non-stationary signals. This is certainly true for vehicle tracking, where one vehicle is following another. In this case, the change detected in the path of the target vehicle motivates the system to adapt to those changes. If you want your system to be able to follow the vehicle that it is tracking, then the state of a vehicle and the probability of detected change must be considered simultaneously. The probability that a change in the system has occurred before any new observations are taken can be thought of as the vehicle maneuverability [2].

Bayesian techniques are aptly applied to change detection because they capture the intuitiveness of human reasoning by inferring conclusions based on known information. A vehicle tracking system can use Bayesian changepoint detection to estimate vehicle states with reasonable accuracy, as proposed by researchers at the Naval Air Warfare Center, Weapons Division [2]. This method was demonstrated to provide robust estimations of vehicle position using both simulated and real maneuvering vehicle data.

Rendering Human Saliency

Another goal for a defense AI system might be to observe whether human behavior might be unusual or suspicious. A clever idea for detecting irregularities in images and video was developed by researchers at the Weizmann Institute of Science. In an effort to capture spatial and behavioral saliency, they considered the problem of describing new observed image or video in terms of spatiotemporal patches which are obtained from a database [3].

The database is used to establish combinations of patches which could be the most probable. The database presented is composed of three pictures: a man sitting in a chair, a man sitting on the floor with one hand raised, and a man sitting on the floor with the other hand raised. New images are compared to the database, and if they cannot be formed using the database, they are classified as irregular. So, a man with both arms raised is considered valid as it can be inferred as a combination of patches from the database, whereas a man sitting to the floor holding his legs to his chest is flagged as unusual because the patches from the database cannot sufficiently describe it.


Figure: Demonstration of detection of irregular image configurations. [3]
This method is called “inference by composition,” and it relies on graph-based Bayesian inference [3]. In some cases, it is possible to apply inference by composition onto an image for which there is no database. This exciting result was demonstrated for images of textiles and food as a quality control application, where repeating patterns in the image were used to locate defects in the items depicted [4].

Change detection is one of many challenging problems in the realm of signal processing and AI, but it holds enormous potential for defense applications. Techniques for change detection as well as emulating saliency have the capacity to not only replace a human’s role in observing abnormalities, but also improve upon it, particularly when combined with advanced sensor technology. As automated defense technologies continue to develop, there is no doubt that change detection will be a critical part of their success.

To read more about Rock West’s applications of signal processing to defense, see: https://www.rockwestsolutions.com/industries/defense/

References

  1. J. Kim, “https://www.naverlabs.com/en/storyDetail/54,” NAVER LABS, 2 February 2018. [Online].
  2. M. R. Kirchner, K. Ryan and N. Wright, “Maneuvering vehicle tracking with Bayesian changepoint detection,” in IEEE Aerospace Conference, Big Sky, MT, USA, 2017.
  3. O. Boiman and M. Irani, “Detecting Irregularities in Images and in Video,” International Journal of Computer Vision , vol. 74, no. 1, pp. 17-31, 2007.
  4. “http://www.wisdom.weizmann.ac.il/~vision/Irregularities.html,” [Online].