Reducing false-positives in multi-sensor dataset of landmines via sensor fusion regularization

Abstract

In a post-war scenario, humanitarian demining is an important and dangerous task that consists in basically 4 phases: Scanning the terrain, detecting potential landmines, distinguishing what is a real landmine from other objects and removing the landmine. Scanning the terrain is usually done in small areas by using the human arm or a robotic arm before touching the object; detecting potential landmines is usually achieved by thresholding the sensor signal, removing the landmines are often achieved by blowing in place method. The biggest challenge is to distinguish the landmines from other objects, this step is normally achieved by touching the object with a pressure probe, but this approach is dangerous and time consuming, specially because the number of false-positives in such a detection is often very high, and one false-negative would be enough to cause a death or a bad injury. Thus, this paper focus on optimizing machine learning techniques to reduce the false positives of such detected objects.