This　paper　presents　a　new　gait　recognition　approach　(DR-KCDML)　to　enhance　the　recognition　performance　for　randomly　arranged　unlabeled　cross　view　gaits.　Existing　methods　always　require　the　label　information　or　assume　corresponding　gait　samples　under　different　views　originate　from　the　same　person.　Our　method　learns　original　affinity　between　cross　view　samples　by　dictionary　representation.　Moreover,　the　gait　samples　of　each　view　are　mapped　into　a　higher　dimensional　feature　space　by　nonlinear　mapping　determined　by　the　kernel　function.　Theses　make　DR-KCDML　can　preserve　the　obtained　cross-view　affinity　relationship　and　nonlinear　characteristics　of　original　cross-view　gait　data.　Experiments　on　the　largest　multi-view　gait　database,　i.e.,　CASIA-B,　show　that　our　approach　is　effective　against　view　change　in　gait　recognition.　©　2018　IEEE.