BLSTM based night-time wildfire detection from video

Area of Science:

• Computer Vision
• Artificial Intelligence
• Wildfire Detection

Background:

• Distinguishing fire from non-fire objects in low-light night videos is challenging due to spatial feature disruption.
• Limited dynamic range of cameras significantly impacts visual analysis in dark environments.
• Temporal behavior analysis is essential for accurate night-time fire classification.

Purpose of the Study:

• To propose and evaluate a Bidirectional Long Short-Term Memory (BLSTM) based algorithm for night-time wildfire event detection from video.
• To address the limitations of spatial feature analysis in low-light conditions for fire detection.
• To provide insights into the unique characteristics of night-time wildfire videos and sources of detection errors.

Main Methods:

• Development of a novel algorithm utilizing Bidirectional Long Short-Term Memory (BLSTM) neural networks.
• Focus on analyzing temporal dynamics within video frames for fire identification.
• Testing the algorithm against diverse real-world recordings of night-time wildfire incidents.

Main Results:

• The proposed BLSTM algorithm achieved an accuracy of 95.15% in detecting night-time wildfire events.
• The algorithm demonstrated an efficient detection time of 23.7 milliseconds per frame.
• Experimental validation confirmed the effectiveness of temporal analysis for this task.

Conclusions:

• The BLSTM-based approach offers a robust solution for night-time wildfire detection in videos.
• Temporal feature analysis is a critical component for overcoming low-light challenges in fire detection.
• Further research can build upon these findings for more targeted and improved wildfire monitoring systems.