LSW-Net: A complex wilderness scenarios segmentation technique based on learnable spherical window multi-modal feature fusion

Wilderness scenarios are characterized by their unstructured and complex diversity, which makes segmentation in such environments more challenging. Current research mainly focuses on perception methods for structured environments (such as urban roads), with relatively less attention given to unstructured wilderness scenarios. Therefore, this paper investigates segmentation techniques for complex wilderness scenarios. Firstly, we propose a multi-channel point cloud mapping method specifically designed for wilderness environments, which extracts both geometric distribution features of terrain structures and ground texture characteristics directly from 3D point cloud data. Secondly, we propose a learnable spherical window mechanism for multi-modal feature fusion, enabling geometric-aware cross-modal interaction. By constructing point cloud spherical windows, the most relevant image context features to the point cloud features are filtered out, enabling the registration and fusion of complementary multi-modal features. Finally, a multi-head fusion classifier is employed to achieve effective segmentation of complex wilderness scenarios under multi-modal data fusion. An experimental platform for ground unmanned intelligent agent perception technology research was built, and the proposed model was subjected to simulation and experimental analysis. The results show that the model has high segmentation accuracy, with mIoU precision reaching 82.01% in simulated environments and 79.30% in real environments, representing an improvement of 7.63% and 5.99% respectively over traditional methods. This model is suitable for segmentation tasks in complex wilderness scenarios, providing a new solution to enhance the perception capabilities of wilderness scenarios.