An　improved　Yolov4　algorithm　was　proposed　to　autodetect　indoor　safety　helmet-wearing.　Firsta　data　set　dedicated　to　the　indoor　safety　helmet-wearing　detection　was　self-built　for　testing　and　evaluating　the　algorithm　due　to　a　lack　of　safety　helmet-wearing　detection　experimental　data　in　indoor　scenarios.　Thenan　adaptive　recalibration　multiscale　feature　fusion　module(ARMFFM)　was　designed　and　embedded　into　the　original　Yolov4　network　to　improve　the　detection　accuracy　of　fuzzy　and　tiny　targets　far　away　from　the　surveillance　camera.　In　ARMFFMthe　features　were　fused　top-down　and　bottom-up　at　different　scales　through　depthwise　over-parameterized　convolutional　layers　for　the　fuzzy　and　tiny　objects　to　obtain　the　more　obvious　texture　and　feature　at　first.　Afterwardsthe　feature　recalibration　module　strengthened　or　suppressed　each　pixel　in　the　fused　feature　map　to　make　the　model　precisely　detect　it　to　avoid　a　conflict　among　the　feature　maps　at　different　scales.　Furthermorea　decoupled　detection　head　replaced　the　detection　head　of　the　original　Yolov4　for　the　individual　performances　of　the　location　and　classification　tasks　of　the　indoor　safety　helmet-wearing　detection.　Additionallya　Soft-CIoU-NMS　post-process　algorithm　was　developed　for　detecting　overlapping　targets.　The　experimental　results　demonstrated　that　the　accuracy　of　the　improved　Yolov4　algorithm　in　the　detection　of　safety　helmet-wearing　in　indoor　scenarios　reached　95.1%about　4.7%　higher　than　that　of　the　original　Yolov4.　Besidesthe　detection　precision　of　fuzzytiny　and　overlapping　targets　was　significantly　enhancedproving　the　superiority　of　the　algorithm　for　indoor　safety　helmet-wearing　detection.　©　2023　Tianjin　University.　All　rights　reserved.