AdaVIT: an Adaptive Visual-Tabular Fusion Multi-modal for road surface snow detection

Snow detection on road surfaces, especially in bare pavement lost and regain periods, is crucial for traffic safety. Traditional methods like manual observation and computer vision, encounter difficulties under complex weather and lighting conditions, with factors like snowfall intensity, temperature fluctuations, and road surfaces complicating detection. To address these limitations, we propose AdaVIT (Adaptive Visual-Tabular Fusion Multi-modal), an innovative framework for road surface snow detection that integrates real-time monitoring images with environmental data. It employs ConvNeXt for image recognition and an enhanced random forest for analysis environmental data. A feature-level multi-modal fusion strategy is designed to integrate environmental data into the image recognition process, while an adaptive feature fusion mechanism ensures an intelligent balance between different data types. Experimental results demonstrate that AdaVIT delivers optimal performance, with significant improvements in both robustness and accuracy, particularly as the dataset size increases. An ablation study further confirms the effectiveness of the proposed adaptive fusion mechanism.