It　is　difficult　to　realize　the　exact　self-localization　of　mobile　robot　by　using　a　single　type　sensor.　The　heterogeneous　sensor　information　fusion　is　utilized　to　improve　the　self-localization　precision.　First,　the　motion　model　of　the　mobile　robot　and　observed　model　of　CCD　vidicon　are　established.　The　optimal　state　estimation　is　derived,　model　disturbances　and　measurement　noises　are　restrained　by　the　Q,　R　matrices,　and　the　self-localization　is　realized　by　the　extended　Kalman　filter.　Then,　the　observed　model　of　the　ultrasonic　sensor　is　established,　and　the　self-localization　information　is　obtained.　Finally,　the　data　from　CCD　vidicon　and　the　ultrasonic　sensor　are　fused　by　BP　neural　network.　The　cooperation　of　the　two　types　of　sensors　is　realized.　The　simulation　results　show　that　the　self-localization　precision　of　the　mobile　robot　is　obviously　improved　by　the　heterogeneous　sensor　information　fusion.