Aiming　at　application　requirements　of　our　spherical　amphibious　robot,　which　has　one　type　of　propulsion　system　to　be　used　both　on　land　and　in　water,　an　improved　tracking　learning　detection　(TLD)　algorithm　is　designed　and　implemented,　so　as　to　improve　the　robustness　and　environmental　adaptability　of　its　visual　detection　system.　First,　binary　normalized　gradient　method　is　implemented　to　reduce　the　number　of　image　candidate　samples.　Second,　the　compressive　sensing　algorithm　is　introduced　to　speed　up　the　image　feature　calculation.　Then,　some　common　videos　are　applied　to　test　the　performance　of　the　improved　TLD　algorithm,　and　the　results　show　that　the　average　frame　rate　of　the　improved　TLD　algorithm　is　increased.　Finally,　to　evaluate　the　feasibility　and　effectiveness　of　the　improved　TLD　algorithm　in　the　visual　system　of　our　amphibious　spherical　robot,　a　series　of　experiments　are　carried　out　with　illumination　changes　and　in　underwater　environment.　Comparison　experimental　results　demonstrated　that　the　improved　TLD　algorithm　has　better　effectiveness,　robustness　and　real-time　performance　than　the　traditional　TLD　algorithm.