Improved CenterNet-Based Multimodal Object Detection for Low-Light and Complex Environments

To address insufficient detail representation, inadequate cross-modal fusion, and limited localization accuracy in object detection under low-light and complex background conditions, this study proposes an improved CenterNet-based multimodal object detection method. The model uses fused images and infrared images as dual-source inputs, where infrared wavelet priors are introduced to enhance texture and structural representation. A Feature Fusion Attention (FFA) module is designed to improve cross-modal feature interaction, while a Heatmap-Guided Detection Head (HGDH) is introduced to explicitly enhance target-related regions during detection. In addition, a two-stack Hourglass backbone is adopted for multi-scale modeling of global semantics and local details. Based on the public LLVIP dataset, a precisely paired fused-image/infrared-image dataset, termed RH-25, was constructed for experiments. On the RH-25 dataset, the proposed method achieves a 3.51% improvement in mAP@0.5 relative to the baseline, as demonstrated by comparative and ablation experiments. Moreover, supplementary experiments on the MFAD dataset indicate potential cross-dataset adaptability under different scenes and multi-class conditions. These results indicate that the proposed method can improve detection performance in low-light and complex environments.