An　improved　multiple　instance　learning　(IMIL)　algorithm　is　proposed　for　person　following　with　a　mobile　robot.　In　the　tracking　process,　radio　frequency　identification　(RFID)　provides　a　searching　area　for　the　IMIL　algorithm,　which　then　successfully　detects　the　person　of　interest.　In　IMIL,　compressed　features　are　extracted　to　describe　instances　of　bags　from　the　low　dimensional　space　so　as　to　reduce　the　time　complexity　in　the　operation.　Then,　the　most　discriminative　weak　classifiers　are　selected　from　the　weak　classifier　pool　by　maximizing　the　inner　product　between　the　weak　classifier　and　the　log-likelihood　function.　The　scheme　avoids　computing　the　bag　probability　and　instance　probability　many　times,　which　further　reduces　the　computational　time.　To　deal　with　the　drift　problem,　the　bag　probability　equally　depends　on　each　instance.　Furthermore,　the　classifiers　are　updated　according　to　the　similarity　between　the　current　tracking　result　and　the　target　model,　thus　they　can　deal　with　appearance　changes　adaptively.　The　method　is　conducted　on　video　sequences　and　a　mobile　robot.　Experimental　results　demonstrate　that　the　presented　method　can　track　the　target　accurately　and　robustly　when　there　are　abrupt　motions　and　appearance　changes,　and　satisfy　the　real-time　requirement.