With　the　development　of　UAV,　UAV　has　been　applied　to　various　projects　with　its　advantages　of　low　construction　cost,　low　safety　risk　coefficient　and　convenient　operation.In　terms　of　UAV　platform,　currently　composite　wing　and　multi-rotor　UAV　are　typically　adopted,　which　can　realize　basic　flight　route.　In　terms　of　image　detection,　neural　network　is　mainly　used　to　classify　and　recognize　the　target　in　the　image.　In　this　paper,　the　YOLOV4　algorithm　is　improved　to　make　it　more　suitable　for　UAV　detection　of　ground　targets.In　the　ground　detection　of　UAV,　most　of　them　are　small　targets,　so　clustering　method　is　used　to　redesign　anchor　for　small　targets.　Because　the　features　of　small　targets　have　more　details　in　the　shallow　feature　layer,　the　shallow　feature　is　superimposed　into　the　feature　extraction　layer,　and　the　shallow　feature　and　the　deep　feature　are　fused.In　the　data　processing,　data　enhancement,　color　dithering,　flipping,　cutting　of　the　data　set　for　expansion.　Through　the　test　of　the　modified　network,　the　following　results　are　obtained:　the　overall　mAP　is　improved　by　9.3%,　the　detection　mAP　for　small　targets　such　as　people　is　improved　by　23.75%,　and　the　detection　mAP　for　working　vehicles　is　improved　by　15.4%.　The　detection　efficiency　of　small　targets　is　improved,　and　the　speed　can　meet　the　real-time　requirements,　and　it　can　be　deployed　in　the　UAV　for　UAV　detection.　©　2021　ACM.