A Method for Constructing a Person Detection Model Based on Global-Local Alignment and Frequency Domain Adaptive Fusion in Fire Smoke Scenarios

Abstract

This invention proposes a method for constructing a personnel detection model based on global-local alignment and frequency domain adaptive fusion in fire smoke scenarios. First, near-infrared and thermal imaging dual-modal images are acquired, and multi-scale features are extracted using a dual-stream YOLOv11 backbone network. Second, a GLAFC fusion module is designed: first, near-infrared features are corrected using thermal imaging as a spatial reference; then, high-frequency textures from near-infrared imaging and low-frequency thermal radiation from thermal imaging are adaptively enhanced in the frequency domain; next, cross-modal repair and fusion are achieved through confidence-aware cross-attention. Then, the fused features are enhanced by SPPF and C2PSA and input into the neck network, where upsampling, stitching, and a C3k2 module are used to aggregate semantic and localization information to generate a multi-resolution detection map. Finally, a multi-task loss function with structural alignment consistency constraints is used to train the model. Experiments show that this invention can significantly reduce the false negative rate and improve detection accuracy in extremely dense smoke environments.

Images