The　quality　of　Received　radio　signal　strength　(RSS)　is　essential　for　obtaining　higher　accuracy　in　the　fingerprint-based　positioning　system.　However,　due　to　the　variation　of　the　RSS　which　is　caused　by　multi-path　propagation　or　moving　objects,　such　a　precise　positioning　system　is　difficult　to　achieve.　To　address　the　above　issues,　a　new　calibration　method　is　presented.　The　method　consists　of　two　parts:　offline　training　part　and　online　positioning　part.　A　high　accurate　model　is　obtained　in　the　offline　training　part　by　a　neural　network,　and　the　unknown　position　is　determined　in　the　online　part　using　the　model　which　is　trained　in　the　previous　part.　In　the　offline　part,　a　Gaussian　fitting　is　first　applied　to　the　distribution　of　the　obtained　RSS　data,　and　mixture　Gaussian　calibration　method　is　adopted　when　more　than　one　significant　peak　is　occurred　in　the　Gaussian　model.　In　the　online　part,　k-times　measurement　is　adopted　to　enhance　the　quality　of　the　RSS　data.　A　set　of　experiments　in　an　indoor　positioning　scenario　is　conducted.　The　experimental　results　show　that　in　the　offline　part,　the　proposed　mixture　Gaussian　calibration　method　can　significantly　improve　the　accuracy　from　3.2　m　to　2.5　m　for　80%　test　points,　and　in　the　online　part,　the　proposed　k-times　measurement　narrows　down　the　localization　error　from　4.3　m　to　2.4　for　80%　test　points.　For　that,　no　hardware　modifications　are　required.