Video　broadcasting　over　wireless　network　has　become　a　very　popular　application.　However,　the　conventional　digital　video　broadcasting　framework　can　hardly　accommodate　heterogeneous　users　with　diverse　channel　conditions,　which　is　called　the　cliff　effects.　To　overcome　this　cliff　effects　and　provide　a　graceful　degradation　to　multi-receivers,　in　this　paper,　we　use　the　nonlocal　sparsity　and　hierarchical　GOP　structure　to　propose　a　novel　CS　based　soft　video　broadcast　scheme.　CS　has　properties　of　minimizing　bandwidth　consumption　and　generating　measurements　with　equal　importance　which　are　exactly　needed　by　video　soft　broadcast.　In　the　proposed　scheme,　the　measurement　data　are　generated　by　block-wise　compressive　sensing　(BCS),　and　then　the　measurement　data　packets　are　sent　over　a　highly　dense　constellation　though　OFDM　channel　to　achieve　a　simple　encoder.　Ideally,　with　the　GOP　structure,　inter　frame　has　lower　sampling　rate　than　intra　frame　to　achieve　better　compression　efficiency.　At　the　decoder　side,　due　to　equally-important　packets　and　property　of　soft　broadcast,　each　user　can　receive　the　noise-corrupted　measurements　matching　its　channel　condition　and　reconstruct　video.　The　hierarchical　GOP　structure　is　presented　to　explode　the　correlation　and　non-local　sparsity　among　video　frames　during　the　recover　process.　Additionally,　using　non-local　sparsity,　group　based　CS　reconstruction　with　adaptive　dictionaries　is　proposed　to　improve　decoding　quality.　The　experimental　results　show　that　the　proposed　scheme　provides　better　performance　compared　with　the　traditional　SoftCast　with　up　to　8　dB　coding　gain　for　some　channel　conditions.