Visible　light　positioning　systems　(VLPs)　that　are　based　on　line-of-sight　(LOS)　are　highly　accurate;　however,　they　cannot　operate　under　shadowing/blocking　conditions　commonly　encountered　in　indoor　environments.　Relaying　on　non-line-of-sight　(NLOS)　transmission　paths　can　resolve　this　problem　at　the　cost　of　poor　performance.　In　this　paper,　we　propose　an　indoor　LOS　and　NLOS　adaptive　VLP　system　that　overcomes　shadowing/blocking　and　maintain　a　high　level　of　performance.　To　achieve　dynamic　algorithm　switching　and　position　estimation,　the　proposed　system　uses　an　improved　perspective-4-points　(P4P)　algorithm.　Further,　a　deep　learning　(DL)　model,　YOLOv5,　is　designed　to　detect　the　highlight　spots　formed　by　light　emitting　diodes　(LEDs)　over　LOS　or　NLOS　paths　due　to　its　improved　accuracy　and　generalization.　Experimental　results　show　that　the　proposed　system　can　provide　a　good　three-dimensional　positioning　accuracy,　with　mean　square　errors　of　13.2　cm　for　LOS　and　28.1　cm　for　NLOS,　respectively.