Research on fall detection algorithm for older adults in complex lighting environments based on improved YOLOv8n.

Background: With the continuous intensification of global aging, the issue of older adult(s) falls has become a significant public health challenge. Real-time and accurate fall detection technology is of great importance for ensuring the safety of the older adult(s). To address the problems of difficult discrimination of fall features caused by human movement in complex backgrounds and low detection accuracy, this study proposes a fall detection method for the older adult(s) based on the improved YOLOv8n, aiming to effectively identify fall behaviors of the older adult(s) in complex lighting and shadow environments. Methods: Firstly, a multi-pose human fall database containing different light intensities and diverse real-life scenarios was constructed. Secondly, based on the YOLOv8n network architecture, seven advanced lightweight and attention mechanism modules, namely the Reparameterization Visual Geometry Group (RepVGG) module, the Spatial Pyramid Pooling with Fast Large Separable Kernel Attention (SPPF_LSKA) module, the Shuffle Attention (SA) module, the C2f module with Spatial and Channel Reconstruction Convolution (C2f_ScConv), the Simplified Spatial Pyramid Pooling-Fast (SimSPPF) module, the Squeeze-and-Excitation Attention (SEAttention) module, and The C2f module with Partial Kernel Interaction (C2f_PKI), were introduced, and 20 groups of improvement experiments covering single-module, dual-module, and triple-module fusions were conducted. The impact of different improvement strategies on model performance was systematically analyzed. Finally, the optimal model was selected for field verification. Results: Experimental results demonstrate that different improvement strategies exert varying effects on the performance of the YOLOv8n model, among which the dual-module enhancement combining C2f_PKI and SimSPPF has proven to be the most effective. Under the condition of maintaining the same parameter count, this improved model achieved an mAP@0.5 of 91.8% on a self-constructed dataset, representing a 2.1% increase over the baseline YOLOv8n model, with an inference speed reaching 41.6 frames per second. The detection accuracy and efficiency of this model surpass those of mainstream models such as YOLOv5s and Faster-RCNN. Furthermore, its detection performance was validated on the publicly available UR Fall dataset, where it achieved an accuracy improvement of up to 10% in real-world scenarios. Conclusion: The study demonstrates that the improved model enables efficient and reliable fall detection, which has significant practical implications for ensuring the safety of older adults and promoting healthy aging.