This study presents a multimodal feature fusion detection algorithm aimed at improving nighttime vehicle detection for intelligent transportation systems. The proposed method integrates data from multiple modalities, including thermal imaging, visible light cameras, and LiDAR sensors, to address the challenges of low visibility and environmental variability at night. Central to this approach is the FusionNet-NV model, which utilizes a dual-stream architecture to independently process and extract modality-specific features before combining them into a unified feature representation. This fusion process is optimized through an attention mechanism and a multimodal alignment module, ensuring that complementary strengths of each modality are fully leveraged while mitigating their respective limitations. An adaptive multimodal fusion strategy (AMFS) is introduced to dynamically adjust the fusion weights based on environmental and contextual factors. This strategy ensures optimal fusion of thermal and visible light data, improving detection accuracy under varying nighttime conditions. By incorporating spatial-temporal modeling and feedback mechanisms, the detection system maintains stability and precision despite changing scenarios. Experimental evaluations demonstrate that the proposed system significantly outperforms existing methods, achieving higher detection precision and robustness under low-light conditions. These results highlight the potential of the FusionNet-NV and AMFS for real-time, accurate vehicle detection in intelligent transportation applications. The study also emphasizes the importance of optimizing computational efficiency for real-world deployment and suggests further research into sensor calibration and fusion strategies to enhance adaptability across diverse environments.
Luo Peng (Wed,) studied this question.