Crowd　counting　algorithms　play　an　important　role　in　the　field　of　public　safety　management.　Most　of　the　current　mainstream　crowd　counting　methods　are　based　on　deep　convolutional　neural　networks　(CNNs),　which　use　multi-column　or　multi-scale　convolutional　structures　to　obtain　contextual　information　in　images　to　compensate　for　the　impact　of　perspective　distortion　on　counting　results.　However,　due　to　the　locally　connected　nature　of　convolution,　this　method　cannot　obtain　enough　global　context,　which　often　leads　to　misidentification　in　complex　background　regions,　which　affects　the　accuracy　of　counting.　To　solve　this　problem.　First,　we　design　a　double　recursive　sparse　self-attention　module,　which　can　better　obtain　long-distance　dependency　information　and　improve　the　problem　of　background　false　detection　on　the　basis　of　reducing　the　amount　of　computation　and　parameters.　Secondly,　we　design　a　Transformer　structure　based　on　feature　pyramid　as　the　feature　extraction　module　of　the　crowd　counting　algorithm,　which　effectively　improves　the　algorithm’s　ability　to　extract　global　information.　The　experimental　results　on　public　datasets　show　that　our　proposed　algorithm　outperforms　the　current　mainstream　crowd　counting　methods,　and　effectively　improves　the　background　false　detection　problem　of　complex　scene　images.　©　2022,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG.