MmodalFire: A Multimodal Dataset Comprising Video and Physical Sensing Data for Detecting Indoor Fires

Because no multimodal dataset was previously available for fire detection research, we developed the MmodalFire multimodal fire detection dataset for training and evaluation of indoor fire detection algorithms. This publicly available dataset includes video and physical sensing data for fire detection use. The dataset comprises 65 videos that simultaneously captured six physical sensing data types, including smoke density, temperature, and infrared and ultraviolet radiation at 5 μm, 4.4 μm, and 3.8 μm. All data were acquired using monitoring cameras and fire sensors deployed as part of a fire detection system that was carefully designed to cover all possible variations, including different wind velocities, illumination conditions, common interference types, and occlusions. All videos and corresponding physical sensing data sequences are labeled as either fire or non-fire sequences. Using the MmodalFire dataset, we evaluated four basic baseline fusion models and the proposed dynamic fusion models to provide a reference for multimodal fire detection research under controlled laboratory settings, promoting research on multimodal fire detection algorithms using controlled-setting data.

此前尚无适用于火灾检测研究的多模态数据集，为此我们构建了MmodalFire多模态火灾检测数据集，用于室内火灾检测算法的训练与评估。该公开数据集包含面向火灾检测任务的视频与物理传感数据。数据集包含65段视频，所有视频均与六类物理传感数据同步采集，具体涵盖烟雾浓度、温度，以及5μm、4.4μm、3.8μm波段的红外与紫外辐射。所有数据均通过部署于火灾检测系统中的监控摄像头与火灾传感器采集所得，该系统经精心设计，可覆盖各类潜在变化场景，包括不同风速、光照条件、常见干扰类型与遮挡情况。所有视频与对应的物理传感数据序列均被标注为火灾序列或非火灾序列。依托MmodalFire数据集，我们对四款基础基线融合模型与本文提出的动态融合模型开展了评估，旨在为受控实验室环境下的多模态火灾检测研究提供参考基准，进而推动基于受控场景数据的多模态火灾检测算法研究。