Scholarlink

This research presents an intelligent system for student counting using computer vision and convolutional neural networks (CNNs). An intelligent system that can automatically count the number of students was developed by applying digital image processing techniques and convolution neural networks to classroom camera images, even in occlusion. The intelligent system was implemented in Python using YOLOv5, Keras, TensorFlow, and Tkinter. The innovative nature of this project is the automatic verification of the number of students in educational spaces considering occlusion. A student database obtained from Roboflow was used to test the intelligent system. YOLOv5 was fine-tuned to handle 640 x 640 resolution images of students, with a confidence level (conf >= 0.35) and a window size of 320. The system was tested using the YOLOv5n and YOLOv5x6 models, working on CPUs and GPUs to identify and count students. The first model achieved a recall of 0.88, and the second demonstrated a higher recall of 0.99, indicating its superior performance. YOLOv5x6 working on GPUs reduced the inference time by 14.04X the CPU version.

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

Intelligent Student Counting System Tolerant to Occlusions Based on Convolutional Neural Networks

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • José Alberto Hernández Aguilar,
  • Yasmín Hernández Pérez,
  • Javier Ortiz-Hernandez,
  • René García Reyes,
  • Felipe Bonilla Sánchez,
  • Jesús del Carmen Peralta-Abarca

摘要

This research presents an intelligent system for student counting using computer vision and convolutional neural networks (CNNs). An intelligent system that can automatically count the number of students was developed by applying digital image processing techniques and convolution neural networks to classroom camera images, even in occlusion. The intelligent system was implemented in Python using YOLOv5, Keras, TensorFlow, and Tkinter. The innovative nature of this project is the automatic verification of the number of students in educational spaces considering occlusion. A student database obtained from Roboflow was used to test the intelligent system. YOLOv5 was fine-tuned to handle 640 x 640 resolution images of students, with a confidence level (conf >= 0.35) and a window size of 320. The system was tested using the YOLOv5n and YOLOv5x6 models, working on CPUs and GPUs to identify and count students. The first model achieved a recall of 0.88, and the second demonstrated a higher recall of 0.99, indicating its superior performance. YOLOv5x6 working on GPUs reduced the inference time by 14.04X the CPU version.