Scholarlink

In this paper, we develop a drone-based solution for detecting productivity characteristics of tomato crops inside agricultural greenhouses using the YOLO8 computer vision model; a mobile phone is used to deploy the trained model. The implementation leverages the Apple Neural Engine (NE), a hardware accelerator module embedded in recent Apple mobile phones, to enable fast and efficient inference. Our video acquisition component also employs a DJI remote controller that streams live video from the drone to the mobile app for processing. The main objective is to perform rapid and precise detection of tomatoes within greenhouses, where drones can improve efficiency and coverage. We describe the model architecture and various optimization techniques suitable for embedded-platform deployment. The experimental study demonstrates the system’s effectiveness in detection accuracy and inference time when utilizing NE compared to CPU-based inference. We also compare accuracy, model size, and inference speed across variants of the YOLO algorithm.

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

Drone-Based Tomato Fruit Detection Through Hardware-Accelerated YOLO Deployment

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Abdellah Islam Kafi,
  • Antonio P. Sanfilippo,
  • Raka Jovanovic,
  • Sa’d Abdel-Halim Shannak

摘要

In this paper, we develop a drone-based solution for detecting productivity characteristics of tomato crops inside agricultural greenhouses using the YOLO8 computer vision model; a mobile phone is used to deploy the trained model. The implementation leverages the Apple Neural Engine (NE), a hardware accelerator module embedded in recent Apple mobile phones, to enable fast and efficient inference. Our video acquisition component also employs a DJI remote controller that streams live video from the drone to the mobile app for processing. The main objective is to perform rapid and precise detection of tomatoes within greenhouses, where drones can improve efficiency and coverage. We describe the model architecture and various optimization techniques suitable for embedded-platform deployment. The experimental study demonstrates the system’s effectiveness in detection accuracy and inference time when utilizing NE compared to CPU-based inference. We also compare accuracy, model size, and inference speed across variants of the YOLO algorithm.