Privacy-preserving crowd counting with thermal imaging alone

Crowd counting in public spaces raises privacy concerns from continuous RGB capture. This work proposes the first thermal-only framework that removes RGB dependency at inference time while maintaining competitive accuracy. The method uses depth-to-RGB diffusion models as a cross-modal bridge to extract features that enhance thermal representations; critically, single-step LCM denoising preserves structural fidelity better than multi-step approaches. Tested on RGBT-CC and DroneRGBT datasets, it matches state-of-the-art RGB-thermal fusion performance while requiring only thermal input, directly addressing privacy constraints in real surveillance systems. Code will be released.