Road Surface Condition Recognition Based on Visual Information Fusion

To address the problem of lower classification accuracy in current road surface condition recognition methods for both daytime and nighttime, this paper proposes an end-to-end deep learning model. Which integrates RGB and Near Infrared (NIR) dual-channel visual information for autonomous road surface condition classification. The collected image data is first augmented to mitigate the effects of model overfitting. A lightweight network based on ConvNeXt-Tiny is designed to extract basic features from RGB and NIR images, respectively. A feature fusion module based on Feature Pyramid Network (FPN) is designed to fuse hierarchical features, which are then classified together with the basic features using fully connected layers. Comparative experiments with MBFN and ablative studies demonstrate the effectiveness of the proposed network model and feature fusion module.