The　difficulty　of　data　transmission　of　panoramic　video　has　always　limited　the　wide　application　of　panoramic　video.　In　order　to　solve　this　problem,　a　panoramic　distribution　algorithm　based　on　the　region　of　interest　is　proposed.　Since　the　user＇s　focus　is　often　on　the　person　appearing　in　the　video,　we　consider　the　area　around　the　person　as　a　potential　area　of　interest.　The　acquisition　of　the　region　of　interest　is　based　on　a　deep　learning　method.　YOLOv4　is　used　to　perform　target　detection　on　the　panoramic　video,　and　the　detected　pedestrian　target　is　input　into　the　multi-target　tracking　module　based　on　the　Deepsort　algorithm.　We　use　the　center　of　the　bounding　box　output　by　the　Deepsort　algorithm　as　the　center　of　the　region　of　interest　and　distribute　the　region　of　interest　in　the　panoramic　video.　This　greatly　reduces　the　amount　of　data　transmitted　by　panoramic　video,　and　basically　does　not　affect　the　user＇s　viewing　experience.　In　order　to　reduce　the　switching　of　target　id,　this　paper　improves　the　appearance　extraction　model　in　Deepsort,　and　proposes　an　appearance　extraction　network　that　focuses　on　both　local　and　global　information,　so　that　the　appearance　extraction　model　can　extract　more　representative　appearance　features　of　pedestrians.　The　validity　of　the　model　is　verified　on　multiple　pedestrian　re-identification　data　sets.