Home Americas US Army-developed drone uses machine learning to sniff out explosive hazards

US Army-developed drone uses machine learning to sniff out explosive hazards

Counter-IED drone project
Army researchers develop and transition technologies that provide Soldiers with explosive hazard indicators from safe standoff using unaltered sensor imagery and UAS platform thermal imagery. Photo: US Army

US Army researchers have successfully employed machine-learning technology to enable a drone-based multi-sensor system to perform standoff detection of explosive hazards.

The research program, which was carried out by the US Army Combat Capabilities Development Command, now known as DEVCOM, Army Research Laboratory, was funded by the Defense Threat Reduction Agency, via the Blood Hound Gang Program, which focuses on a system-of-systems approach to standoff explosive hazard detection.

Route reconnaissance in support of convoy operations remains a critical function to keep soldiers safe from roadside explosive hazards, such as improvised explosive devices, unexploded ordnance and landmines. These hazards continue to threaten operations abroad and continually prove to be an evolving and problematic adversarial tactic.

Kelly Sherbondy, program manager at the lab, said the reality is that there is no single sensor solution that will rule them all in regards to different emplacement scenarios and also emphasized the significance of continued collaborative research efforts.

In Phase I of the program, researchers took 15-months to evaluate mostly high-technology readiness level standoff detection technologies against a variety of explosive hazard emplacements. In addition, a lower-TRL standoff detection sensor, which was focused on the detection of explosive hazard triggering devices, was developed and assessed. The Phase I assessment included probability of detection, false alarm rate and other figures of merit which will ultimately lead to a down-selection of sensors based on best performance for Phase II of the program.

A fixed-wing aircraft platform equipped with detection algorithms was executed with various sensor permutations and dual-polarization synthetic aperture radars to enable standoff detection of explosive hazards. Photo: US Army

The sensors evaluated during Phase I included an airborne synthetic aperture radar, ground vehicular and small unmanned aerial vehicle LIDAR, high-definition electro-optical cameras, long-wave infrared cameras and a non-linear junction detection radar.

Researchers carried out a field test in real-world representative terrain over a 7-kilometer test track and included a total of 625 emplacements including a variety of explosive hazards, simulated clutter and calibration targets. They collected data before and after emplacement to simulate a real-world change between sensor passes.

Terabytes of data was collected across the sensor sets which was needed to adequately train artificial intelligence/machine learning algorithms. The algorithms subsequently performed autonomous automatic target detection for each sensor. The sensor data is pixel-aligned via geo-referencing and the AI/ML techniques can be applied to some or all of the combined sensor data for a specific area. The detection algorithms provide confidence levels for each suspected target, which is displayed to a user as an augmented reality overlay. The detection algorithms were executed with various sensor permutations so that performance results could be aggregated and determine the best course of action moving forward into Phase II.

“The accomplishments of these efforts are significant to ensuring the safety of the warfighter in the current operation environment,” said Lt. Col. Mike Fuller, US Air Force Explosive Ordnance Disposal and DTRA program manager.

The Army says future research will enable real-time automatic target detection displayed with an augmented reality engine. The three year effort will culminate with demonstrations at multiple testing facilities to show the technology’s robustness over varying terrain.

“We have side-by-side comparisons of multiple modalities against a wide variety of realistic, relevant target threats, plus an evaluation of the fusion of those sensors’ output to determine the most effective way to maximize probability of detection and minimize false alarms,” Fuller said. “We hope that Army and the Joint community will both benefit from the data gathered and lessons learned by all involved.”