Scholarlink

In this paper, we present an approach and a study to use a model of classification of drone obstacles based on MobileNet architectures to validate the existence of obstacles on aerial image captured by a monocular RGB (Red Green Blue) camera-equipped drone after passing it to simple computer vision algorithms; the models used were MobileNet, V2, V3 Small, and V3 Large. We assemble a dataset of multiple classes of aerial and ground obstacles, specifically drones, airplanes, electric lines/power lines/towers, birds, and buildings. We start training based on the pretrained weights of the ImageNet dataset using Transfer learning techniques to accelerate the training time and accuracy. Finally, we compare the accuracy and performance of all models on the Raspberry Pi 3 model b and the Nvidia Jetson Nano using a unified inference pipeline based on OpenCV DNN tools. MobileNet V3 Small minimalistic version gave the best accuracy/inference time ratio reaching 95.14% (top1) with inference times of 35 ms on the Nano and 56 ms on Pi 3. The training of the original MobileNet and Mobilenet V2 was done using the Darknet framework, while MobileNet V3 Small and Large were trained using TensorFlow due to the absence of pretrained weights on Darknet.

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

Fast Drone Obstacle Detection Approach Based on MobileNet Classification CNN Architectures

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ali El Habchi,
  • Wassim Khiati,
  • Ilham Zerrouk,
  • Mohammed Aissi,
  • Jamal Berrich,
  • Toumi Bouchentouf

摘要

In this paper, we present an approach and a study to use a model of classification of drone obstacles based on MobileNet architectures to validate the existence of obstacles on aerial image captured by a monocular RGB (Red Green Blue) camera-equipped drone after passing it to simple computer vision algorithms; the models used were MobileNet, V2, V3 Small, and V3 Large. We assemble a dataset of multiple classes of aerial and ground obstacles, specifically drones, airplanes, electric lines/power lines/towers, birds, and buildings. We start training based on the pretrained weights of the ImageNet dataset using Transfer learning techniques to accelerate the training time and accuracy. Finally, we compare the accuracy and performance of all models on the Raspberry Pi 3 model b and the Nvidia Jetson Nano using a unified inference pipeline based on OpenCV DNN tools. MobileNet V3 Small minimalistic version gave the best accuracy/inference time ratio reaching 95.14% (top1) with inference times of 35 ms on the Nano and 56 ms on Pi 3. The training of the original MobileNet and Mobilenet V2 was done using the Darknet framework, while MobileNet V3 Small and Large were trained using TensorFlow due to the absence of pretrained weights on Darknet.