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

Area of Science:

• Computer Science
• Engineering
• Safety Science

Background:

• Multimodal datasets are crucial for advancing fire detection technology.
• Existing datasets lack the comprehensive, synchronized data needed for robust algorithm development.

Purpose of the Study:

• To introduce the MmodalFire dataset, a novel multimodal resource for indoor fire detection research.
• To provide a standardized benchmark for training and evaluating fire detection algorithms.

Main Methods:

• Collected 65 synchronized videos and six types of physical sensor data (smoke density, temperature, IR/UV radiation).
• Ensured data diversity by varying wind velocity, illumination, interference, and occlusion.
• Labeled all data sequences as either fire or non-fire.

Main Results:

• The MmodalFire dataset enables comprehensive evaluation of multimodal fire detection.
• Baseline fusion models and novel dynamic fusion models were tested on the dataset.
• Established a performance baseline for multimodal fire detection in controlled environments.

Conclusions:

• The MmodalFire dataset addresses the need for multimodal data in fire detection research.
• Facilitates the development and validation of advanced fire detection algorithms.
• Promotes further research in multimodal sensor fusion for enhanced fire safety.