Design and Implementation of Fall Detection System Based on SoPC

Addressing the challenge of deploying convolutional neural network-based object detection algorithms to edge computing platforms, this paper presents a SoPC embedded system that facilitates the hardware deployment of the YOLOv3-Tiny algorithm for fall detection. In the FPGA hardware domain, the system incorporates a Deep Learning Processing Unit (DPU) designed to expedite the convolutional calculations of neural networks. On the algorithmic front, improvements to the YOLOv3-Tiny neural network yield the Lite-YOLOv3-Tiny model, enhancing detection performance. Experimental results demonstrate a 14.92% increase in mAP value for the improved model compared to the original model on the GPU platform. In the designed SoPC embedded system, the model’s volume post-quantization and compilation is 12.7MB, representing a 74.7% reduction compared to the GPU platform. Achieving a fall detection mAP value of 0.951, with a detection speed of 62.07ms, the system experiences only an 11.67ms increase compared to the GPU platform, making it suitable for providing effective monitoring tools for medical personnel, households, and public space managers.