Design and Implementation of Real-Time Fire Segmentation Algorithm on PYNQ Z2 FPGA

Field Programmable Gate Arrays (FPGA) are being employed for various hardware designs required for aerospace, autonomous vehicles, industries, defence, biomedical applications and many more. Fire detection has always been of utmost significance and early detection prevents loss of life and damage to property. Similarly, Machine Learning (ML) is being employed almost everywhere. In this paper, an attempt is made to design a real-time fire segmentation system for fire detection using an FPGA and hardware Machine Learning accelerator from Tensil. The proposed design is implemented and evaluated on a Xilinx Zynq 7000 FPGA based PYNQ Z2 evaluation board. Details about the design flow of proposed work is reported in this paper, followed by experimental results. Several models were designed, implemented on hardware and their performances are compared and reported in the paper. It is observed that the best response time of the proposed model is 0.63 seconds with 2.199W power consumption. The proposed work can be employed for various other applications and generalizes the procedure for deploying an ML model at the edge on an FPGA.