A mobility aid for the visually impaired should be able to detect and warn about nearby obstacles. Reliable detection of curbs and steps is critical to meet this goal. This paper describes a stereo-vision based algorithm that estimates the underlying planar geometry of the 3D scene to generate hypotheses for the presence of steps. Tensor voting is used to calculate globally consistent normals at each data point and a clustering algorithm is described to generate a piecewise planar model of the scene. Results demonstrate the improvement in plane clustering using tensor voting and the ability of the algorithm to find sufficient evidence for the presence of curbs and steps.