In　recent　years,　autonomous　driving　has　become　a　hot　topic,　especially　in　the　complex　urban　road　environment.　The　visual　algorithm　is　the　most　used　scheme　for　autonomous　driving.　The　traditional　condition　imitation　learning　adopts　the　end-to-end　deep　learning　network.　But　it　lacks　interpretability,　and　the　ability　of　feature　extraction　and　expression　of　network　is　limited.　There　are　still　some　problems　in　the　local　planning　and　detail　implementation.　To　solve　these　problems,　we　propose　to　use　the　deep　residual　network　architecture　and　add　the　dual　attention　module　to　learn　driving　skills,　which　are　closer　to　human　beings.　To　further　improve　the　detailed　feature　extraction　ability　of　the　network,　the　deeper　residual　network　architecture　is　used.　To　adaptively　integrate　the　global　context　long-range　dependence　of　the　image　in　the　spatial　and　feature　dimensions,　the　dual　attention　module　is　adopted　to　improve　the　ability　of　network　expression.　At　the　same　time,　in　order　to　make　full　use　of　the　multi-period　attribute　information　of　the　camera　image　itself,　we　redesign　the　network　architecture,　extract,　integrate　the　three-way　temporal　information　features　and　the　high-level　semantics,　and　increase　the　interpretability　of　the　temporal　information　of　the　model.　This　method　is　tested　on　the　CARLA　simulator.　The　experimental　results　show　that　compared　with　the　benchmark　algorithm,　it　achieves　better　driving　effect.　Deeper　feature　extraction　and　multi-period　information　fusion　can　effectively　improve　the　driving　ability　and　driving　completion　of　the　agent.　©　2020　ACM.