A deep learning and edge computing integrated approach for fall behavior detection in buildings

Abstract

This study proposes a model integrating YOLOv11 and STGCN for accurate and real-time detection of fall behaviors within buildings. A specialized video dataset comprising fall behaviors performed by six volunteers was developed and used to validate the model’s effectiveness in cloud computing and edge computing environments. The results obtained in the cloud computing environment were characterized by ample computational resources and the absence of real-time constraints. The model achieved precision, recall, and F1-score for fall behaviors exceeding 0.98. The model was integrated into edge computing devices in an actual test environment to directly process real-time video stream data. A missed detection rate of 18 % was observed on the Jetson ORIN NX 16GB device, while the Jetson AGX Orin 64GB recorded a lower missed detection rate of 15 %. Similarly, a false alarm rate of 16 % was observed on the Jetson ORIN NX 16GB device and 12 % on the Jetson AGX Orin 64GB device. These performance differences between the high-performance cloud computing cluster and edge computing devices, as well as among different edge computing devices, may be attributed to variations in computational resources, data quality, and device parameters. The results demonstrate the potential of the proposed model for real-time fall detection in resource-constrained environments.