Aiming　at　the　existing　deep　learning-based　fall　detection　model　in　multiple　complex　scenes　such　as　multi-person　scenes,　long-distance　scenes,　and　occlusion　scenes,　there　are　problems　such　as　leakage　detection　and　false　detection.　In　this　paper,　we　take　the　falling　person　as　the　research　object,　construct　a　dataset　containing　multiple　complex　scenes　such　as　multi-person　scenes,　long-distance　scenes,　and　occlusion　scenes,　and　propose　a　fall　detection　model　based　on　improved　YOLOv5.　By　improving　the　backbone　network　and　neck　network　structure　of　the　detection　model,　the　problem　of　low　accuracy　of　fall　detection　due　to　incomplete　target　deformation　features　and　target　detail　features　in　complex　scenes　is　solved.　The　proposed　algorithm　in　this　paper　achieves　an　average　accuracy　of　92.78%　for　fall　detection　in　complex　scenes,　and　the　detection　speed　reaches　36　frames/second.　©　2024　Copyright　held　by　the　owner/author(s).