DoD Builds New Algorithms for Video Monitoring
The Defense Department is making strides in its effort to harness machine learning and other artificial intelligence technologies to assist with monitoring footage taken from unmanned aerial vehicles.
In spring 2017, the department announced the creation of the algorithmic warfare crossfunctional team, or Project Maven, which is dedicated to utilizing big data analytics, machine learning, artificial intelligence and other advanced computer technologies to gain a competitive advantage over peer adversaries.
The team has recently seen success in developing algorithms that assist in monitoring full-motion video collected by UAVs, said Travis Axtell, informatics technical director and deputy for Project Maven under the office of the under secretary of defense for intelligence.
“The early results definitely show that we can augment [military analysts] and give them some new functionality to things that they are currently doing manually over several hours, and we can now do them in a handful of minutes,” he said at a 2017 webcast conference hosted by Amazon Web Services.
The Project Maven team now wants to apply those algorithms “to a historical context,” he added. Rather than simply monitoring live feeds, new techniques could help intelligence officers extract relevant information from previously recorded footage, he noted.
Advanced cloud computing technologies could assist in “pulling out detections from that footage, storing that and then processing against those detections, so that we can look for more pattern-of-life behavior and understand how an area works and what we need to be looking for when the situation varies in a very dynamic sense,” he added.
Axtell noted that the team uses a process called training quality data, where members review the footage drawn from the UAV and highlight areas and objects of interest.
“Then we put that in front of our analyst community and they are labeling it at a variety of different detail levels … and then putting that into a rudimentary neural network,” he said.
The Project Maven team will then “triage” some of the incoming data to allow the most interesting objects and data to be placed at the front of the review line, he added. “That was one of the ways that we were able to be so successful at getting an algorithm out the door in a matter of months,” he said.
The Defense Department is now looking for new opportunities for technologies such as cloud computing to better organize the mountains of data it gathers from UAVs, in order to more rapidly apply it to the training quality data process, Axtell noted.