FirePM-YOLO: Position-Enhanced Mamba for YOLO-Based Fire Rescue Object Detection from UAV Perspectives

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• UAV-based fire rescue object detection faces challenges like small targets, occlusion, smoke, and flame interference.
• Existing models like YOLO prioritize speed over global context, limiting performance in complex fire scenarios.

Purpose of the Study:

• To propose FirePM-YOLO, an optimized object detection architecture for UAV-based fire rescue applications.
• To enhance the detection of small, occluded targets and improve overall scene understanding in challenging fire environments.

Main Methods:

• Introduced a Position-Aware Enhanced Mamba module (PEMamba) with positional encoding, spatial enhancement, and adaptive feature fusion.
• Developed a PEMBottleneck structure to balance local convolutional and global PEMamba features, integrated as PEM-C3K2 modules.
• Utilized a self-built "FireRescue" dataset for training and evaluation.

Main Results:

• FirePM-YOLO demonstrated improved mean average precision (mAP) and recall compared to YOLOv12 and other detectors.
• The model maintained real-time inference capabilities.
• Achieved superior detection performance on challenging samples, including small-scale and partially occluded firefighting vehicles.

Conclusions:

• FirePM-YOLO effectively addresses limitations of mainstream detectors in complex fire rescue scenarios.
• The proposed PEMamba and PEMBottleneck modules enhance perception and contextual understanding for critical emergency response tasks.
• The architecture offers a promising solution for real-time, accurate object detection in UAV-based fire rescue operations.