With　the　continuous　development　of　deep　learning,　the　field　of　repetitive　action　counting　is　gradually　gaining　notice　from　many　researchers.　Extraction　of　pose　keypoints　using　human　pose　estimation　networks　is　proven　to　be　an　effective　pose-level　method.　However,　the　existing　pose-level　methods　have　some　drawbacks,　for　example,　ignoring　the　fact　that　occlusion　and　unfavourable　viewing　angles　in　videos　lead　to　affect　the　accuracy　of　pose　keypoints　extraction.　To　overcome　these　problems,　we　propose　a　simple　but　efficient　Salient-Part　Pose　Keypoints-Based　Dual-Branch　Network　(SPKDB-Net).　Specifically,　we　design　a　dual-branch　input　channel　consisting　of　a　global-based　and　a　salient-part　input　branch.　The　global-based　input　branch　is　used　to　input　the　pose　keypoints　of　the　whole　body　extracted　by　the　human　pose　estimation　network,　and　the　salient-part　input　branch　is　used　to　input　the　salient-part　pose　keypoints　(i.e.,　head,　shoulders,　and　hands).　The　second　branch　acts　as　an　auxiliary　to　the　first　branch,　thus　effectively　addressing　the　influence　of　external　factors.　In　addition,　we　propose　a　DFEPM-Module　that　obtains　long-distance　dependency　between　pose　keypoints　through　the　attention　mechanism,　and　obtains　salient　local　features　fused　by　the　attention　mechanism　through　convolution.　Eventually,　extensive　experiments　on　the　challenging　RepCount-pose,　UCFRep-pose　and　Countix-Fitness-pose　benchmarks　show　that　our　proposed　SPKDB-Net　achieves　state-of-the-art　performance.　©　2025　Elsevier　Inc.