Object　segmentation　is　one　of　the　key　issues　in　object-based　video　analysis　and　retrieval.　In　this　paper,　a　two-stage　object　segmentation　method　is　proposed.　In　the　starting　frame　a　global　search　algorithm　is　invoked,　in　which　the　candidate　regions　are　extracted　using　motion,　size　and　color,　and　the　candidate　regions　are　confirmed　using　Discriminative　Map　(DM)　and　Adaboost　classifier.　Only　when　the　output　of　global　searching　is　true,　the　local　searching　algorithm　is　invoked　in　the　following　frame,　in　which　object　is　searched　and　segmented　using　DM　and　Adaboost　classifier　in　a　local　area.　If　the　output　of　the　local　search　algorithm　is　positive,　it　will　be　used　in　the　next　frame;　otherwise　the　global　search　algorithm　is　re-invoked.　The　experimental　results　using　basketball　videos　show　that　our　method　is　effective　and　robust　to　complex　back-grounds　and　the　blurred　contour　of　object　in　motion.