Visual　Simultaneous　Localization　and　Mapping　(vS-LAM)　is　essential　to　autonomous　robot　navigation　and　AR/VR　systems.　As　deep　learning　technologies　improve　rapidly,　researchers　integrate　object　detection　into　traditional　vSLAM　frameworks　to　acquire　objects　from　images.　The　object　detection　modules　allow　the　SLAM　system　to　construct　object-level　maps　with　the　estimated　pose,　location,　and　scale　estimation　of　detected　objects.　Since　pose　estimation　is　a　non-linear　process,　an　accurate　pose　initialization　is　crucial　to　improve　the　optimality　of　the　estimation　result,　which　is　vital　to　constructing　an　object-level　map.　Although　existing　works　have　achieved　remarkable　progress　in　object　pose　initialization,　the　accuracy　still　needs　to　be　improved　to　meet　the　needs　of　real　applications.　In　this　paper,　an　improved　object　pose　initialization　strategy　based　on　weighted　adaptive　sampling　is　proposed,　which　is　robust　to　noise　in　the　extracted　line　features.　In　addition,　a　SLAM　system　that　can　construct　an　object-level　map　with　semi-dense　line　features　based　on　the　proposed　weighted　scoring　of　pose　initialization　is　presented.　The　qualitative　and　quantitative　evaluation　experiments　on　public　datasets　show　that　the　proposed　approach　outperforms　the　state-of-the-art　works　in　accuracy　and　robustness.　©　2023　IEEE.