Real-time　detection　of　vehicular　volume,　mean　speed　and　vehicle　type　has　important　significance,　but　the　existing　video-based　detection　methods　are　not　satisfactory　at　processing　speed　and　accuracy.　This　paper　proposes　a　high-efficient　method　to　detect　all　the　three　parameters　from　two　foreground　temporal-spatial　images　(TSIs)　directly,　which　are　obtained　from　two　virtual　detection　lines　(VDLs)　in　video　frames.　Such　usage　of　the　TSIs　provides　a　feasible　approach　to　solve　the　problems　of　vehicle　occlusion,　mean-speed　estimation,　and　vehicle　classification　without　using　original　frame　images.　Firstly,　for　improving　the　accuracy　of　detection,　during　generation　of　the　foreground　TSIs,　we　set　a　small-wide　region　of　interest　for　each　VDL　and　propose　a　local　background　subtraction　method　and　an　improved　moving　shadows　elimination　method　to　eliminate　unwanted　interferences.　Then,　in　order　to　reduce　the　calculation　complexity,　during　extraction　of　the　parameters,　we　analyze　the　feasibility　of　vehicle　classification　direct　from　the　foreground　TSIs,　and　propose　a　method　to　extract　shape-feature　vector　from　the　TSIs　directly.　The　dependence　on　original　frame　images　is　minimized,　so　the　pressing　speed　is　improved　obviously.　Experimental　results　prove　the　feasibility　and　efficiency　of　the　proposed　method.　©　2013　IEEE.