Eye-gaze　tracking　technology　provides　us　with　an　unconventional　way　of　Human　Computer　Interaction,　and　brings　a　lot　of　convenience　with　many　practical　applications　and　industrial　products.　The　main　focus　of　eye-gaze　tracking　is　to　calculate　the　model　of　gaze　directions　and　the　gaze　coordinates.　Based　on　computer　vision,　the　key　problem　in　the　modeling　process　is　the　feature　description.　In　this　paper,　a　new　descriptor,　Local　and　Scale　Integrated　Feature　(LoSIF)　is　proposed　to　extract　eye-gaze　movement　features　based　on　non-intrusive　system,　allowing　slight　head　movements.　The　feature　descriptor　depends　on　two-level　Haar　wavelet　transform,　and　carries　on　a　combination　of　multi-resolution　characteristics　and　effective　dimension　reduction　algorithm,　to　achieve　local　and　scale　eye　movement　characterization.　We　use　Support　Vector　Regression　to　estimate　the　mapping　function　between　the　appearance　of　the　eyes　and　the　corresponding　gaze　directions.　As　the　experiment　results　show,　the　accuracy　is　within　an　acceptable　range　of　1　degree.　©　2011　IEEE.