Lightweight network for small target fall detection based on feature fusion and dynamic convolution

The accurate and prompt detection of falls in the elderly holds significant importance in building a fall detection system based on artificial intelligence. However, the current research has many limitations, including poor performance in low-light conditions, missed detection for small targets, excessive parameters, and slow detection speed. This paper combines feature fusion, dynamic convolution, and the SCYLLA-IoU (SIoU) loss function to overcome these challenges. First, FasterNet is employed to ensure a balance between lightweight and accuracy. Second, the bi-directional cascaded feature pyramid network is introduced, incorporating a module to enhance feature representation and improving the perception capability for targets in dark images. Furthermore, dynamic convolution is implemented based on attention mechanisms to enhance the perception and localization accuracy for small object detection tasks. Finally, the SIOU loss function is introduced to expedite convergence speed and improve target localization accuracy. Experimental results demonstrate that the improved model outperforms the original YOLOv5s model, achieving a 6.6% increase in precision and a 15.3% enhancement in detection speed, while reducing parameter count by 24%. It exhibits superior performance compared to other networks, including Faster-R-CNN, SSD, YOLOXs, and YOLOv7.