Detection of Buried IEDs Using Convolutional Neural Networks on Ground Images from Thermal Cameras Embarked on UAVs

Buried improvised explosive devices (IEDs) are commonly used by terrorist in Burkina Faso. Burial requires the soil to be disturbed, which can be detected effectively using thermal cameras. In this paper, we propose an IED detection approach that consists of discriminating between ordinary soil and deliberately disturbed soil. The approach uses convolutional neural networks and images acquired by thermal cameras on board unmanned aerial vehicles (UAV). We employed VGG, Inception, and ResNet architectures for soil texture classification. The VGG architecture with a shallow model yields better results. However, images captured with a large field of view (FOV) limit the detection capacity due to the multitude of noisy objects in this environment. We then use transfer learning to improve the efficiency of the models. In our paper, we also make the link between the depth of CNN architectures and their ability to discriminate thermal images of disturbed and ordinary soils.