In　order　to　solve　problems　in　the　process　of　traditional　three-lane　detection,　such　as　low　anti-interference　capacity,　inaccuracy　in　lane　fitting,　and　error　identification　of　side　lane,　a　three-lane　detection　method　was　proposed　in　this　paper　based　on　TopHat　segmentation　and　curve　models.　A　lane　segmentation　algorithm　using　variable-kernel　TopHat　was　proposed　as　image　pre-processing,　by　using　shape　features　and　color　features　of　lane　markings.　For　lane　detection,　firstly,　a　vanish-point　fitting　method　based　on　WLS　(weighted　least　squares)　was　proposed　as　a　constraint　of　Hough　transform.　Secondly,　straight　lines　were　clustered　in　polar　coordinates　by　using　DBSCAN　(density-based　spatial　clustering　of　applications　with　noise),　matching　to　a　template.　Then　lane　ROI　(region　of　interest)　was　made　according　to　the　straight-line　template,　and　the　lane　was　searched　and　fitted　by　using　cubic　curve.　Finally,　for　uncertain　side-lanes,　a　side-lane　driving　judgement　method　was　proposed　by　using　random　seed.　The　algorithm　has　better　performance　in　both　detection　rate　and　lane　miss　rate　than　traditional　three-lane　detection　algorithms.　Experiment　verifies　that　the　method　has　high　accuracy　and　stability,　and　is　useful　for　three-lane　detection.　©　2016,　Editorial　Department　of　Journal　of　Beijing　University　of　Technology.　All　right　reserved.