HFD-YOLO: Improved YOLO Network Using Efficient Attention Modules for Real-Time One-Stage Human Fall Detection
Adri Priadana, Duy-Linh Nguyen, Xuan-Thuy Vo, Jehwan Choi, Russo Ashraf, Kang-Hyun Jo
Abstract
Detecting human falls has become a crucial technology for intelligent surveillance systems, particularly in smart factories. A one-stage fall detection model based on the YOLO network provides a solution for resource-limited environments, offering real-time detection at high speeds. This work proposes HFD-YOLO, an improved version of YOLOv8n, for one-stage human fall detection. It introduces a Sequential Efficient Attention Module (SEAM) and Residual Bottleneck Self-Attention (RBSA) for feature enhancement, leading to improved detection precision. The SEAM includes efficient channel and spatial attention and applies a large separable kernel to consider a wider receptive field in the spatial attention module. The RBSA uses a bottleneck mechanism to mitigate issues related to excessive parameters and computations. The proposed HFD-YOLO outperforms other methods on two benchmark datasets (Fall Detection and Human Fall Detection) based on mean average precision (mAP). With a low computation load and relatively few parameters, HFD-YOLO achieves an average of 23.54, 18.71, and 29.18 frames per second on an Intel Core i7-9750H CPU, Intel Core i5-6600 CPU, and NVIDIA Jetson Orin Nano GPU devices, respectively, making it suitable for real-time human fall detection to support intelligent surveillance systems in resource-constrained environments. The demo video is presented at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://bit.ly/HFD-YOLO-AP</uri>.