Adaptive Indoor People-Counting System Based on Edge AI Computing

Mao Hsu Yen, Bor Shyh Lin, Yu Lun Kuo, I. Jung Lee, Bor Shing Lin

研究成果: Article同行評審

摘要

Since the outbreak of COVID-19 in 2019, people counting in confined spaces has become essential for controlling the flow of people and reducing viral spread. Many people-counting systems use a fisheye lens to achieve a wide viewing range. However, the images captured with a fisheye lens can have different sizes and deformations due to different camera heights, which consequently affects the accuracy of people counting. In this study, a Raspberry Pi 4 development board attached to a fisheye lens was used to construct an edge computing platform for people counting by using artificial intelligence. A modified You Only Look Once (YOLO) model, which has three output layers and a spatial pyramid pooling module, is proposed to recognize people in images captured at various heights. For efficiently running this deep learning model on the proposed edge computing platform, a TensorFlow Lite framework was used to reduce the model size. The final experimental results indicated that the mean average error and mean squared error of the modified tiny version of the YOLO v4 model were 0.249 and 0.292, respectively, and the detection speed of the platform was 2 frames per second. The results indicate that the proposed edge computing platform can efficiently and accurately count the number of people. Because the proposed system uses edge computing, it can effectively solve personal privacy and transmission bandwidth problems.

原文English
頁(從 - 到)1-9
頁數9
期刊IEEE Transactions on Emerging Topics in Computational Intelligence
DOIs
出版狀態Accepted/In press - 2023

指紋

深入研究「Adaptive Indoor People-Counting System Based on Edge AI Computing」主題。共同形成了獨特的指紋。

引用此