Road　scene　segmentation　is　mainly　used　in　the　field　of　autonomous　driving　today.　However,　the　complexity　of　road　scenes　brings　great　challenges　and　difficulties　to　the　perception　and　understanding　of　the　vehicle　environment,　so　the　semantic　segmentation　of　complex　traffic　scenes　is　a　challenging　research　topic　in　the　field　of　computer　vision.　When　facing　complex　background　and　variable　scale　real　scenes,　the　accuracy　of　segmentation　method　based　on　full　convolution　neural　network　needs　to　be　further　improved.　In　response　to　this　problem,　this　research　proposed　a　model　based　on　a　multi-scale　attention　to　improve　the　performance　of　semantic　segmentation　algorithms　from　multiple　perspectives.　First,　we　design　a　multi-scale　attention　module,　which　can　combine　multi-scale　information　and　attention　to　obtain　semantic　correlation　between　spatial　dimensions　and　channel　dimensions　at　different　scales　in　encoder.　Secondly,　the　number　of　fusions　of　low-level　features　and　high-level　features　is　increased　to　alleviate　the　problem　of　low-level　information　loss　caused　by　multiple　downsampling　in　the　decoder.　Finally,　a　better　feature　fusion　is　achieved　by　adding　an　adaptive　feature　fusion　module　after　concatenating　the　decoder　feature　maps.　The　experimental　results　show　that　on　the　Cityscapes,　compared　with　the　baseline　DeepLabV3+,　the　MIoU　of　the　model　is　improved　by　1.3%　on　the　validation　set　and　1.2%　on　the　test　set.　　©　2022　IEEE.