FusionFormer: An Off-Road Sence Semantic Segmentation Network Based on Data Fusion and Hierarchical Transformer

The off-road environment poses significant challenges and obstacles to the further development of environmental perception due to the irregularity of its objects and the randomness of their distribution. In order to pursue higher precision of semantic segmentation in complex and unordered environments with irregular objects and uneven quantities, the Fusion Former is raised, which is based on image data fusion, hierarchical Transformer and Focal Loss. The network has strong learning capabilities by fusing depth and image information, using Transformer hierarchical to obtain multi-scale features, adopting Focal Loss to address class imbalance issues. The experiment corroborate that FusionFormer is Extremely capable to improve the precision and multi-class semantic segmentation capabilities in off-road scene semantic segmentation tasks.