HL-YOLOv8s: A High-Performance and Lightweight Network for Fall Detection

AYOLOv8s-based fall detection method, termed HL-YOLOv8s, is presented in this paper. It achieves high-performance and lightweight prediction on edge devices via two steps: architectural design and model pruning. In the first step, three key components - the spatial collaborative perception fusion module (SCPFM), spatial channel dual attention network (SCDAN), and the dynamic decoupled head (DDH) - are designed to ensure considerable performance. SCPFM enhances global information perception via expanding receptive fields, while SCDAN aims to learn the relationship between different positions and channels with an attention mechanism. DDH, motivated by the recently popular Dynamic Head method, alleviates the negative effect caused by multiple down samplings of features. In the second step, random channel pruning is adopted to reduce the dependency on computing resources. Experimental results on the Fall Detection Image Dataset (FDID) indicate that HL-YOLOv8s obtains a mean average precision (mAP@0.5:0.95) improvement of 3.0% compared to YOLOv8-s, while the number of parameters, floating point operations per second (FLOPS), and model size are reduced by up to 45.10%, 52.82%, and 43.93%, respectively. Additionally, the final model deployed on the Jetson Xavier NX satisfies the requirement of real-time detection.