Hyper real-time flame detection: Dynamic insights from event cameras and FlaDE dataset

Abstract

Bio-inspired sensors known as event cameras offer significant advantages over traditional frame-based RGB cameras, particularly in overcoming challenges like static backgrounds, overexposure, and data redundancy. In this paper, we explore the potential of event cameras in flame detection. Firstly, we establish an open-access Flame Detection dataset based on Event Cameras (FlaDE). To mitigate noise in extreme conditions with event cameras, we then propose a denoising preprocessing module termed Recursive Event Denoiser (RED). By leveraging distinctive probability distributions between signals and noise, RED achieves 0.974 (MESR) on the E-MLB benchmark, outperforming than other statistical denoising methods. Furthermore, we delve into the physical meaning behind the event rates, enabling statistical extraction of flame amidst static background and other dynamic sources. Based on this insight, we develop the hardware-efficient BEC-SVM flame detection algorithm. Benchmarked against other prominent detection modules using the FlaDE dataset, our approach highlights the feasibility of leveraging event data for robust flame detection, achieving a detection accuracy of 96.6% (AP.50) with a processing speed of 505.7 FPS on CPU. This research contributes valuable insights for future advancements in flame detection methodologies. The implementation of the code is available at https://github.com/KugaMaxx/cocoa-flade.