Wearing　a　safety　helmet　is　one　of　the　most　important　requirements　of　the　construction　site　and　is　essential　to　the　safety　of　workers.　Computer　vision　can　be　applied　to　identifying　the　helmet　worn　by　the　workers　as　external　supervision.　In　this　paper,　helmet　detection　algorithms　based　on　YOLO　models　with　a　special　data　set　where　the　training　set　consists　of　simple　helmet　pictures　but　the　test　set　holds　complicated　real　construction　sites　are　studied.　In　view　of　actual　situations　of　the　construction　site,　some　pretreatment　methods　for　the　training　set　are　tested　to　enhance　the　performance.　The　result　shows　that　with　proper　pretreatment,　the　YOLOv3　model　with　a　simple　training　set　can　have　good　performance　in　detecting　helmet　in　complicated　construction　sites.　©　2021,　The　Editor(s)　(if　applicable)　and　The　Author(s),　under　exclusive　license　to　Springer　Nature　Singapore　Pte　Ltd.