Abstract
<title>Abstract</title> <p>Urban road traffic safety is confronted with severe challenges, especially for the detection of small and occluded objects in complex traffic scenarios. Traditional multispectral feature fusion methods for fusing RGB and infrared images suffer from such issues as feature imbalance and sensitivity to lighting variations. This paper introduces MEFP-Net, a multimodal feature enhancement algorithm employing hierarchical mid-stage fusion. It features a dual-path backbone network and a multiparametric feature pyramid module aimed at improving detection of small objects. Experimental results from the M3FD dataset indicate the model achieves 85.5% mAP50% and 57.8% mAP50-95%, outperforming state-of-the-art models and effectively reducing false detection rates under challenging conditions (e.g., nighttime and foggy weather), thereby enhancing the accuracy of detecting difficult targets.</p>