Scholarlink

We present a comparative study of real-time object detection models optimized for low-power devices, specifically the Raspberry Pi 3B and 4B. We evaluated multiple deep learning models including SSD MobileNet V2 [1], EfficientDet-D0, and the YOLO [3] v5/v8/v11 series under both FP32 and INT8 quantized settings. Using a custom dataset with only 977 images across three object classes, we fine-tuned each model and measured inference speed (FPS), latency, and mean average precision (mAP). Our findings show that models such as YOLOv11 and SSD-MobileNet-v2-FPNLite maintain high accuracy even after quantization, making them highly suitable for deployment on low-power edge devices.

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

Efficient Real-Time Object Detection Using Deep Neural Networks: A Comparative Analysis on Low-Power Devices

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Byambabat Batkhuyag,
  • Miho Akiyama,
  • Takuya Saito

摘要

We present a comparative study of real-time object detection models optimized for low-power devices, specifically the Raspberry Pi 3B and 4B. We evaluated multiple deep learning models including SSD MobileNet V2 [1], EfficientDet-D0, and the YOLO [3] v5/v8/v11 series under both FP32 and INT8 quantized settings. Using a custom dataset with only 977 images across three object classes, we fine-tuned each model and measured inference speed (FPS), latency, and mean average precision (mAP). Our findings show that models such as YOLOv11 and SSD-MobileNet-v2-FPNLite maintain high accuracy even after quantization, making them highly suitable for deployment on low-power edge devices.