Drone camera uses AI to detect plastic landmines without internet

Standard metal detectors struggle to locate modern antipersonnel landmines due to their small size and plastic casings. Techniques like ground-penetrating radar are also less effective against these non-metallic threats, particularly scatterable mines like the Soviet-era PFM-1, known as the "butterfly mine." These mines are designed to evade detection and inflict injury rather than immediate death.

A team from Binghamton University has developed a novel approach utilizing machine-learning algorithms to detect these plastic mines over broad areas. The system employs a drone-mounted camera to capture low-resolution images, which are then processed by the You Only Look Once (YOLO) object-detection algorithm. This AI model was trained using inert PFM-1 mines and 3D-printed replicas in various environmental conditions to build a comprehensive dataset.

The research involved training two YOLO models: one exclusively on PFM-1 mines and another on PFM-1 mines alongside other random objects. The latter model, while exhibiting lower performance values which may better reflect real-world scenarios where cameras capture extraneous elements like leaves, aims to identify specific targets within a cluttered environment.

Crucially, the system is designed for field readiness, with much of the processing occurring during the algorithm training phase. In deployment, only a lightweight laptop, drone, and camera are required. This allows for real-time or near real-time data analysis, eliminating the need to transport data for processing. This capability is vital for active and post-conflict zones where internet connectivity is often unreliable due to damaged infrastructure or signal jamming.