Indoor accessibility detection for wheelchair users with visual impairments, which aims to identify whether the regions are wide enough for wheelchair to pass, is crucial to ensuring autonomy and safety. Existing smart wheelchair systems equipped with sensors and algorithms have recently shown promise in improving navigation by detecting obstacles and optimizing route planning. However, they generally neglect the physical feasibility of paths, evaluating accessibility only in terms of obstacle-free space rather than actual geometry, such as path width. To fill this gap, this paper proposes a geometry–aware accessibility detection system driven purely by monocular vision. The proposed system consists of three key components: a contour-compensated geometry inference to acquire accurate width information by compensating sparse geometric data, a cross-dimensional accessibility alignment maintains stable estimates despite landmark loss from intermittent 3D point associations and a graph-aware contextual memory to keep temporal consistency under occlusions and viewpoint shifts. Experiments on a self-collected indoor dataset demonstrate that the proposed method achieves a mean Intersection-over-Union (mIoU) of 0.952, an average path-width estimation error of 0.078 m and a successful frame ratio of 92.71% , improving the safety and independence of visually impaired wheelchair users in complex indoor environments.