Vision-Based Stair Environment Perception for Lower-Limb Exoskeletons Using Deep Semantic Segmentation

Stair environment perception is a critical component for enhancing the mobility and safety of lower-limb exoskeletons, especially in unstructured or dynamic environments. In order to solve the problems of insufficient accuracy and weak generalization ability of existing visual perception models in dealing with diverse stair structures and irregular geometries, this paper proposes an innovative deep semantic segmentation network architecture named StairFormerNet. Based on the StairNet backbone network, the model designs a double-branch encoder structure: a convolutional neural network ased on the Convolutional Neural Network for extracting fine local texture features. The other uses a hierarchical Vision Transformer to model long-distance contextual information and spatial dependencies. The two-way features are deeply fused through the introduced Cross-Attention Fusion Module, and the high-precision segmentation results are output by the multi-scale decoder and the Stair-Aware Refinement Blocks, which especially strengthens the recognition ability of the stair edge and tread boundary. Furthermore, a double-loop intelligent optimizer is constructed, which starts from the standard deviation propagation formula and combines the triple constraints of time-cost-accuracy to form a multi-objective function that minimizes measurement error and optimization efficiency, and proposes a set of interpretable recognition algorithms. The cross-domain joint training is performed on the StairData dataset and the real staircase images collected from the first-person perspective.