Pavement　pothole　detection　is　very　important　for　highway　maintenance　and　safety　management.　Aiming　at　the　problems　of　low　detection　accuracy　and　low　efficiency　due　to　the　characteristics　of　various　shapes,　different　scales,　and　complex　background　environment　of　current　pavement　potholes,　this　paper　proposed　a　highway　pavement　pothole　detection　method　based　on　YOLOv5.　First,　a　pothole　image　dataset　containing　multiple　scene　types　was　established.　Then,　based　on　the　adaptive　anchor　frame　adjustment　strategy,　the　YOLOv5　model　was　improved　to　complete　the　design　of　the　pavement　pothole　detection　network,　which　solved　the　problem　of　low　detection　accuracy　caused　by　large　changes　in　the　target　scale　during　the　network　training　process.　By　modifying　the　classification　network,　the　output　dimension　is　reduced,　the　amount　of　calculation　is　reduced,　and　the　detection　speed　is　accelerated.　Finally,　with　the　help　of　transfer　learning　ideas,　the　model　is　trained　and　learned,　and　the　method　of　rotation,　cropping　and　other　methods　are　used　for　data　enhancement,　which　solved　the　problem　of　training　overfitting　caused　by　insufficient　dataset.　The　experimental　results　show　that　the　proposed　method　has　a　detection　accuracy　of　93.99%　for　pavement　potholes,　which　is　3.6%　higher　than　the　original　YOLOv5　model,　and　the　average　detection　time　on　a　single　test　image　is　12.78　ms,　which　is　reduced　by　4.14　m.　©　2022,　The　Author(s),　under　exclusive　license　to　Springer　Nature　Switzerland　AG.