In　omnidirectional　images　or　videos,　the　viewer　receives　an　interactive　and　immersive　experience　from　the　viewport　by　changing　the　viewing　angle.　Due　to　the　wide　application　of　omnidirectional　videos,　the　visual　quality　assessment　for　omnidirectional　videos　is　becoming　an　urgent　issue.　Due　to　the　large　resolution　of　an　omnidirectional　video,　regions　with　object　motions　usually　catch　the　viewers＇　attention,　so　the　motion　regions　have　great　influences　on　the　visual　quality　perception.　Since　the　number　of　potential　viewports　is　huge　and　the　viewer　spends　varying　amounts　of　time　for　different　viewports,　viewport　selection　is　a　critical　yet　not　resolved　problem　for　omnidirectional　video　quality　assessment　(VQA).　In　this　paper,　we　propose　a　two-stream　network　with　viewport　selection　for　blind　omnidirectional　VQA　to　incorporate　the　influences　of　motion　regions　and　viewport　selection.　Firstly,　we　propose　a　two-stream　multi-task　convolutional　neural　network　(TSMT)　for　VQA　at　any　viewport,　which　uses　video　frame　sequences　and　motion　sequences　as　inputs.　The　motion　sequences　are　represented　as　horizontal　and　vertical　optical　flows.　Based　on　the　observation　that　the　low　latitude　regions,　the　front　view,　and　the　moving　objects　have　higher　possibilities　that　appearing　in　the　viewport,　we　propose　a　viewport　selection　method　based　on　a　fusion-based　saliency　map　that　considers　those　regions.　Experimental　results　on　two　datasets　demonstrated　that　the　proposed　model　outperforms　state-of-the-art　omnidirectional　VQA　methods.