Research on high-rise building fire detection algorithm based on short-wave infrared camera

Short-wave infrared (SWIR) imaging has emerged as an ideal data source for fire detection due to its unique smoke penetration capability and high sensitivity to flame radiation. However, existing algorithms exhibit limitations in complex scenarios, such as small target detection and motion blur. This study proposes an enhanced YOLOv5 algorithm optimized for SWIR image characteristics, providing a foundational solution for future UAV-based fire monitoring systems. First, the original backbone network is replaced with the lightweight MobileNetv3, leveraging depthwise separable convolutions to improve computational efficiency. Second, a channel and spatial attention mechanism (CBAM) is integrated into the neck network to amplify flame feature responses, effectively mitigating missed detections under smoke occlusion. Furthermore, the Focal-EIoU loss function is introduced to replace the original CIoU, addressing sample imbalance by dynamically adjusting gradient weights for high-quality anchor boxes. A real-time image stabilization module is designed, combining adaptive ORB feature detection and bidirectional optical flow tracking to compensate for motion blur caused by UAV vibrations. Experimental results demonstrate that the improved algorithm achieves a 9.5% increase in mean average precision ([email protected]) compared to the original YOLOv5, while maintaining high real-time performance (33 FPS) and reducing model parameters by 65%. This work provides a lightweight, high-precision, and dynamically robust solution for UAV-based fire detection in high-rise buildings.