Real-time human detection and behavior recognition using low-cost hardware

Cameras are becoming more pervasive and ubiquitous. The daily activities of individuals are being captured by millions of cameras in public spaces, while individuals are obtaining massive amounts of egocentric videos by employing wearable cameras intended for life-logging. However, recording devices are inexpensive, highly computational, and inconvenient for privacy. We used a low-resolution infrared sensor to detect human activity, including sitting, standing, and lying down, and to locate humans. We acquired the data from a low-cost infrared device and preprocessed them to train the YOLO-v5 network. We developed and tested an infrared technology-based system consisting of $32 \times 24$ thermal input. Our proposed model is trained on 3,864 low-resolution images and made publicly available. The trained YOLO-v5 achieved 96.34% mean Average Precision (mAP) using our designed lightweight and low-cost activity recognition device. We proposed Artificial Intelligence of Things (A-IoT) system can be used either as a stand-alone data collection such as an IoT device or as a data processing and analysis sub-center. Our system consists of a low-power edge computing device and a cost-effective low-resolution infrared module. Our proposed dataset is now available at https://github.com/InfoLab-SKKU/Thermal-Human-Detection