MFEN: A Multi-Layer Fractal Encoding Network Based on Color Distribution for Road Surface Condition Recognition

The recognition of road surface conditions significantly influences the active safety control of autonomous driving vehicles and overall driving performance. Due to the improvement of vehicle perception capabilities, vision-based road surface condition recognition has gained more attention in recent years. Nevertheless, the existing methods are mainly based on CNN, and perform poorly on texture-like data such as road surfaces due to their failure to capture spatial layouts accurately or leverage shallow low-level visual features. To address this problem, we propose a novel network based on fractal geometric dimensions, which primarily emphasizes the analysis of fractal dimensional statistical information extracted from road surface texture features within the color space rather than the coordinate space. Specifically, we propose a Point Repackage Block (PRB) module to effectively segment the feature maps into sub-feature groups based on statistical similarity, which enables the exchange of texture features information on the road surfaces, as well as the reorganization of the feature maps with similar attributes across all channels. In addition, a FractalHead module is employed to conduct fractal geometric statistics on the sub-feature groups in the color space. Finally, we combine the local orderless descriptors generated by FractalHead and the spatial information accumulated by GAP to achieve road surface condition recognition. Through extensive experiments conducted on a dedicated road surface dataset against the other methods, the superiority of the proposed method is demonstrated.