This　paper　describes　a　new　scheme　based　on　Latent　Dirichlet　Allocation　for　place　recognition　of　mobile　robot　system.　It　firstly　extracts　the　local　features　from　the　training　images,　forms　a　discrete　set　of　＇image　words＇　which　are　commonly　known　as　vocabulary　or　codebook,　and　each　image　is　represented　as　a　frequency　vector　based　on　this　vocabulary.　Then　the　model　based　on　Latent　Dirichlet　Allocation　is　used　to　learn　themes　distribution　in　the　training　set　and　testing　images.　Finally　the　unknown　test　images　are　recognized　according　to　the　similarity　of　themes　distribution.　In　order　to　evaluate　the　method,　we　perform　it　on　the　IDOL2　Database　and　our　own　pictures.　Experimental　results　show　that　the　method　has　good　robustness　to　different　types　of　variations,　including　different　illumination　conditions,　different　perspective　and　other　changes　over　long　periods　in　real-world　environments.　©　2011　IEEE.