A reconfigurable smart camera implementation for jet flames characterization based on an optimized segmentation model

This research presents an innovative system for enhancing fire safety management in industrial environments through the deployment of a smart camera platform tailored for jet flame characterization. The primary objective is to address the current limitations in real-time solutions for early fire segmentation and detailed analysis. The proposed platform achieves this by optimizing a UNet segmentation model, making it suitable for implementation on a System-on-Chip Field-Programmable Gate Array (SoC FPGA). This optimization facilitates highly parallel execution and on-device image processing, thereby significantly reducing latency and video processing overheads. The framework utilizes optimized AI models on the SoC FPGA to establish a comprehensive edge processing pipeline for analyzing jet flames. The methodology involves training a high-precision model using an infrared dataset of jet flames, subsequently optimizing this model using Vitis, and then porting it to an FPGA-based SoC architecture for thorough evaluation. This strategy not only aims for improved accuracy but also ensures robust on-device inference, mitigating latency concerns in critical fire safety applications.