First,　the　outer　orientation　of　pedestrian　target　was　obtained　through　head-shoulder　or　color　features.　Then　the　improved　SIFT　features　of　the　pixels　were　extracted　to　locate　the　target　accurately　in　the　coarse　location　area.　The　pedestrian　scale　change　problem　was　solved　based　on　SIFT　features　to　figure　out　the　scale　variation　of　the　pedestrian.　Simultaneously,　the　updating　mechanism　of　SIFT　features　template　library　was　introduced　into　feature　reservation　priority　level　to　solve　the　problem　of　temporary　occlusion　and　deformation　of　the　target.　Aiming　at　the　adaptive　elliptic　kernel　function　of　the　traditional　Cam-Shift　algorithm,　the　variable-bandwidth　and　orientation　ellipse　kernel　was　combined　with　scale　variations　of　the　human　and　the　Epanechnikov　function,　which　reduced　the　interfere　of　background.　Furthermore,　the　searching　area　for　Harris　operator　was　limited　by　the　outler　coarse　location　results,　which　improved　the　real-time　performance　of　SIFT　feature　matching.　The　experimental　results　indicate　that　the　proposed　mobie　robot　human　tracking　algorithm　can　accomplish　human　tracking　under　conditions　of　target　scale　variations,　temporary　occlusion　and　deformation.　©　2016,　Zhejiang　University　Press.　All　right　reserved.