DPU Accelerated Thermal Facial Detection for Improved Accuracy and Performance Efficiency

This paper presents a method for thermal facial detection utilizing the YOLO v8 model running on the ONNX (open neural network exchange) framework model, which is deployed to the Xilinx KR260 developer board. The system uses the Deep Processing Unit (DPU) on the Xilinx KR260 Field Programmable Gate Array (FPGA) to accelerate detection tasks, enhancing performance and power efficiency. The FLIR K53 thermal camera captures thermal data for detection. The model is trained on the YOLO-v8-Nano model, which achieves 99% precision, 99% mAP50 (mean Average Precision at 50% IoU (Intersection over Union)), and 86% mAP50-95 accuracy scores. Results show that the KR260 DPU performs approximately 17 times faster than the KR260 ARM CPU. The DPU consumes less than 5 watts, offering 30% higher performance per watt (PPW) than the Nvidia RTX 30 Gen graphics card and around 8 times higher PPW than the Intel i7 CPU. The DPU can process detection at approximately 17 frames per second (FPS).