In　order　to　implement　high-accuracy　self-localization,　kidnapping　and　tracking　of　soccer　robots　during　the　matches　among　medium-sized　teams,　a　self-localization　method　based　on　the　improved　genetic　algorithm　is　proposed.　In　this　method,　first,　a　mathematical　model　of　genetic　algorithm　is　established,　in　which　the　minimum　sum　of　the　white　line　points　in　the　image　and　the　corresponding　points　in　the　model　map　is　used　to　evaluate　the　target　function.　Then,　based　on　the　global　self-localization　of　the　genetic　algorithm,　the　gradient　optimum　algorithm　is　used　to　partially　modify　the　major　pose　for　the　purpose　of　improving　the　self-localization　precision　and　the　algorithm　robustness.　Finally,　with　regard　to　the　kidnapping　and　tracking　of　the　robot,　the　author　points　out　that　the　error　of　the　distance　between　the　observation　points　and　the　actual　points　should　accord　with　the　Gaussian　distribution　for　the　purpose　of　updating　the　population　status　and　realizing　the　tracking　of　robot,　and　that,　when　the　individual　adaption　degree　of　population　sharply　declines,　the　dynamic　self-adaptive　tuning　of　mutation　probability　helps　to　reduce　the　population　deficiency　effect　and　realize　the　recovered　self-localization　of　kidnapping.　Simulated　and　experimental　results　indicate　that　the　proposed　self-localization　method　is　superior　to　those　based　on　the　traditional　genetic　algorithm　and　on　the　Monte　Carlo　algorithm,　with　its　average　self-localization　tracking　error　being　(0.046　m,　0.22°).