For　visual-based　robotic　manipulation,　it　has　always　been　a　challenging　task　to　perform　real-time　and　accurate　pose　estimation　of　target　objects　under　cluttered　background,　illumination　variations,　occlusion,　and　weak　texture,　especially　under　severe　occlusion　conditions.　In　recent　years,　the　RGB-based　methods　based　on　vector　field　prediction　are　proved　to　be　robustness　on　6D　object　pose　estimation　under　occlusion.　At　the　same　time,　network　with　attention　mechanism　has　achieved　outstanding　performance　in　2D　object　detection.　In　this　paper,　we　propose　an　attention-driven　6D　pose　estimation　method　with　multi-constraints　loss　and　pixel-wise　voting.　We　calculate　the　distance　weighted　unit　vector　length　and　included　angle　length　based　on　prediction　results　to　regularize　unit　vectors　prediction.　Moreover,　we　introduce　Dense　Atrous　Spatial　Pyramid　Pooling　(DenseASPP)　and　Channel-wise　Cross　Attention　(CCA)　mechanisms　into　the　network　structure　to　improve　the　accuracy　of　output　prediction.　Experiments　on　LINEMOD　and　Occlusion　LINEMOD　datasets　manifest　that　our　method　outperforms　state-of-the-art　two-stage　sparse　2D　keypoints　prediction　methods　without　pose　refinement.　©　2022,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG.