In　the　monocular　vision　inertial　system,　an　accurate　real-time　initialization　state　estimator　is　the　prerequisite　for　the　system　to　maintain　steady　operation.　However,　in　the　initialization　of　most　current　monocular　visual　inertial　systems,　the　impact　of　rotation　matrix　on　the　other　external　parameters,　such　as　gyroscope　bias,　accelerometer　bias　and　velocity,　is　often　ignored.　Moreover,　in　global　optimization,　the　keyframe　constraints　in　the　sliding　window　are　not　fully　utilized,　resulting　in　the　continuous　accumulation　of　estimated　path　errors.　In　order　to　solve　the　above　problems,　we　propose　our　monocular　visual　inertial　system　based　on　the　initialization　state　estimator　with　online　external　calibration　and　the　relative　pose　based　marginalization.　In　this　system,　the　external　parameters　of　the　camera　and　IMU　are　calibrated　during　the　initialization　phase,　the　estimations　of　speed,　gyroscope　bias,　accelerometer　bias,　scale　factor　and　gravity　are　performed,　and　the　relative　pose　optimization　method　is　adopted　to　retain　the　constraint　relation　of　keyframes　to　be　marginalized　in　the　optimization　window　and　then　input　into　the　global　optimization.　Experiments　show　that　our　system　can　give　real-time　and　accurate　estimation　in　initialization,　and　can　realize　more　stable　and　accurate　pose　estimations,　even　with　continuous　error　accumulation.