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

A multi-institution research team has published a novel framework for industrial fire safety, implementing a smart camera system specifically designed for the real-time segmentation and characterization of dangerous jet flames. The core innovation is a hardware-software co-design that leverages a System-on-Chip Field-Programmable Gate Array (SoC FPGA)—the Ultra96 platform—to run an optimized AI model directly on the edge. This eliminates the need to send video feeds to a cloud server, drastically reducing processing latency and enabling instant hazard detection.

The team's key technical achievement was using the Xilinx Vitis AI development platform to massively optimize a standard UNet segmentation model for efficient deployment on the FPGA's reconfigurable logic. They compressed the model from 7.5 million parameters down to just 59,095—a 125x reduction in size. Through further optimizations like multi-threading and batch normalization, they achieved a 7.5x improvement in processing latency. The final system performs inference at 30 frames per second while maintaining accuracy, as measured by the Dice Score metric, providing a reliable, real-time monitoring solution.

This work demonstrates a complete, replicable pipeline for deploying vision AI in demanding industrial environments where low latency is critical. By moving from a generic computer vision setup to a dedicated, optimized hardware implementation, the researchers have created a prototype that addresses a significant gap in real-time industrial fire safety management. The framework is also designed to be adaptable, suggesting it could be extended to other safety and monitoring applications beyond jet flame analysis.