Scholarlink

Improvised explosive devices (IEDs) are used by Armed Terrorist Groups (GATs) to hinder the progress of the military forces. The burial of IEDs causes disturbances to the ground surface perceptible by thermal sensors. However, it happens that GATs dig the ground without burying an object in order to just simulate the presence of an IED. We refer to this as a decoy throughout this work. Environmental factor can also create similar ground disturbances. In this paper, disregarding the detection aspect, we propose an approach based on convolutional neural networks (CNNs) to distinguish a decoy from a buried object. Our approach uses VGG, Inception, ResNet, and ViT architectures. The model selection was based on a comparison of metrics over the training, validation, and testing steps. The best results, with accuracy rates of 88% and 95% during the validation phase, were achieved using respectively ResNet-18 and VGG16 with batch normalization and stride subsampling.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Classification of Buried IEDs and Buried IED Decoys Using Convolutional Neural Networks on Ground Images from Thermal Cameras Embarked on UAVs

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Issaka Daouèga,
  • Sadouanouan Malo,
  • T. Frédéric Ouédraogo

摘要

Improvised explosive devices (IEDs) are used by Armed Terrorist Groups (GATs) to hinder the progress of the military forces. The burial of IEDs causes disturbances to the ground surface perceptible by thermal sensors. However, it happens that GATs dig the ground without burying an object in order to just simulate the presence of an IED. We refer to this as a decoy throughout this work. Environmental factor can also create similar ground disturbances. In this paper, disregarding the detection aspect, we propose an approach based on convolutional neural networks (CNNs) to distinguish a decoy from a buried object. Our approach uses VGG, Inception, ResNet, and ViT architectures. The model selection was based on a comparison of metrics over the training, validation, and testing steps. The best results, with accuracy rates of 88% and 95% during the validation phase, were achieved using respectively ResNet-18 and VGG16 with batch normalization and stride subsampling.