Terrain Segmentation Network in Wild Environments With Hybrid Plus Downsampling

Existing segmentation networks primarily use single downsampling to extract low-resolution semantic information, which may not adapt well to features of different scales, leading to information imbalance and distortion. Here, we propose a hybrid plus downsampling method to address this issue. Concretely, we first introduce a linear dilated convolutional unit block to capture long-range dependencies; second, we joint nonlinear pooling to construct comprehensive downsampling features; we then utilize carefully designed super-resolution reconstruction module and similarity structural loss to ensure the completeness of the downsampling features. Furthermore, considering the significance of semantic information, we propose a residual semantic encoding module to gather rich semantic information from a local and global perspective. Based on the aforementioned efforts, we propose a terrain segmentation network (TSNet) for safe navigation of mobile robots in wild environments. Extensive experimental results on the wild datasets demonstrate that TSNet outperforms other state-of-the-art methods in recognizing wild unstructured terrain.