Light-sensitive and adaptive fusion network for RGB-T crowd counting

Mainstream RGB-T crowd counting methods use cross-modal complementary information to improve the counting accuracy. However, most of them neglect the effect of lighting variation on cross-modal data fusion. In this paper, we propose a Light-sensitive and Adaptive Fusion Network (LAFNet) for RGB-T crowd counting. Specifically, we present a Modality-specific Feature Extraction Module (MFEM) that fuses the lighting information, and a Light-sensitive and Adaptive Fusion Module (LAFM) that adjusts the fusion strategies of different modalities according to the lighting conditions of the input crowd images. Moreover, we propose an Improved Multi-scale Extraction Module (IMEM) to extract and fuse multi-modal at different scales. We evaluate our method on the RGBT-CC dataset and the experiment results show the validity of the model and its effectiveness in various scenarios.