To　extend　the　velocity　measurement　range　for　the　omni-directional　intelligent　wheelchair　and　improve　the　measurement　accuracy　and　computational　efficiency,　the　traditional　optical　flow-based　velocity　estimation　method　is　improved　in　this　paper.　First,　a　TV-L1　model　is　introduced　to　estimate　optical　flow,　the　displacement　of　corresponding　pixels　between　two　consecutive　frames　is　efficiently　predicted,　and　the　searching　area　is　reduced.　Second,　a　planar　surface　optical　flow　model　based　on　random　sample　consensus　(RANSAC)　method　is　presented　to　remove　outlier　vectors　produced　by　non-uniform　ambient　illumination　and　local　motion　blur.　Finally,　the　algorithm　is　accelerated　under　the　framework　of　compute　unified　device　architecture　to　improve　the　real-time　performance　of　the　system.　Experimental　results　show　that　maximum　measurable　velocity　obtained　the　proposed　method　is　1.67　times　as　fast　as　original　method,　and　it　is　more　accurate　than　that　of　wheel　odometry.　The　proposed　method　performs　robustly　in　the　presence　of　non-uniform　ambient　illuminations　and　local　motion　blur,　and　it　can　improve　the　maximum　measurable　velocity　of　the　omni-directional　wheelchair　and　the　measurement　accuracy.　©,　2015,　Beijing　University　of　Technology.　All　right　reserved.