Low-brightness feature Redistribution in fusion process of Cross-Modal Crowd Counting
Chengbing Zhang, Ying Liu, Yu Hao · 2024
Crowd counting techniques are devised to estimate the number of people within the footage. In the low-brightness environment, the majority of the-state-of-the-art counting approaches based on RGB images perform poorly because the insufficiency of information. To address this issue, researchers have begun to investigate cross-modal approaches exploiting both RGB and thermal images obtained from infrared cameras. During the feature fusion process, the multi-head attention mechanisms are widely adopted to address feature misalignment and enhance the quality of features. However, main-stream attention modules don't consider the impact of low-brightness scenario on the RGB branch. In low-brightness conditions, RGB images can't provide the sufficient features required by the feature extraction network, which will affect the effectiveness of fused features for counting. In this study, we propose the Brightness-Dependent Feature Recombination (LBFR) module to handle low-brightness RGB data during feature fusion. With comparative experiments on benchmark cross-modal datasets, we found that the proposed module significantly enhances the performance of baseline cross-modal counting networks.