As　the　projection　matrix　is　only　generated　in　the　initial　stage　and　kept　constant　in　subsequent　processing,　so　when　the　object　is　occluded　or　its　appearance　changes,　this　will　result　in　drifting　or　tracking　lost.　To　address　this　problem,　this　paper　proposes　a　real-time　compressive　tracking　algorithm　based　on　online　feature　selection.　First,　the　feature　pools　are　constructed.　Then,　features　with　high　confidence　score　are　selected　from　the　feature　pool　by　a　confidence　evaluation　strategy.　These　discriminating　features　and　their　corresponding　confidences　are　integrated　to　construct　a　classifier.　Finally,　tracking　processing　is　carried　on　by　the　classifier.　Tracking　performance　of　our　algorithm　compares　with　that　of　the　original　algorithm　on　several　public　testing　video　sequences.　Our　algorithm　improves　on　the　tracking　accuracy　and　robustness;　furthermore,　the　processing　speed　is　approximately　25　frames　per　second.　It　meets　the　requirements　of　real-time　tracking.