With　the　increase　of　the　elderly　population,　the　phenomenon　of　the　elderly　falling　at　home　or　out　is　more　and　more　common.　Therefore,　fall　detection　is　of　great　significance　for　the　health　protection　of　the　elderly.　Throughout　the　research　of　fall　detection　at　home　and　abroad,　most　of　the　fall　detection　based　on　video　monitoring　is　complex　and　redundant,　which　affects　the　real-time　and　accuracy　of　detection.　In　view　of　the　above　problems,　this　paper　proposes　a　fall　detection　method　based　on　video　in　complex　environment,　aiming　to　detect　fall　behavior　more　accurately　and　quickly.　The　main　work　of　this　paper　is　as　follows:　firstly,　YOLOv3　network　model　is　proposed　for　detection　algorithm.　Secondly,　the　human　fall　detection　data　set　is　constructed　by　referring　to　Pascal　VOC　data　set　format.　Then,　the　algorithm　model　is　optimized　and　trained　in　GPU　(graphic　processing　unit)　deep　learning　server.　Finally,　comparison　of　test　results　with　our　YOLOv3　network　model　and　other　detection　algorithms　shows　that　our　detection　algorithm　has　a　good　recognition　effect.　©　2020　IEEE.