SDN-YOLO: Research on the Fall Detection Algorithm for Privacy Scenarios Based on Optimized YOLO11

With the increasing degree of population aging, the rapid rise in the elderly population will lead to a serious issue of fall detection for elderly individuals living alone. Fall detection using images from conventional cameras exposes the privacy of elderly individuals and suffers from performance degradation due to insufficient lighting. Therefore, it was decided to use depth images for fall detection. The current methods for fall detection still face issues such as inconvenience, low accuracy, and high missed detection rates. This paper presents SDN-YOLO, a fall detection algorithm designed for privacy-preserving scenarios, which is based on an enhanced version of YOLO11, aimed at improving the detection rate. First, the C3k2 module is redesigned in the backbone network. By integrating the SConv branch for edge information extraction and the convolution branch for spatial information, this module enables the model to learn more comprehensive image feature representations. Secondly, to address the potential loss of detailed information due to the pooling operation in SPPF, convolution with different dilation rates is used to capture information in the image. Additionally, parameter sharing in the convolution helps reduce redundancy, thereby improving model efficiency. Finally, the BiFPN concept is introduced in the neck part, enabling the model to effectively fuse feature information from different layers, thereby improving the model's performance. Experimental results demonstrate that SDN-YOLO enhances the [email protected] by 8.5%, Recall by 5.9%, and Precision by 2.7% over the baseline YOLO11 model on the dataset.