Visual　tracking　is　one　of　hot　researches　in　computer　vision　in　recent　years.　C-COT　[8]　has　obtained　excellent　results　on　many　visual　tracking　benchmarks.　However,　it　cannot　exploit　CNN　features　effectively　because　it　gave　the　same　weight　for　different　CNN　features.　Furthermore,　it　updated　model　frame　by　frame,　it　possibly　results　in　model　drift.　To　address　these　problems,　we　propose　an　improved　C-COT　based　visual　tracking　scheme　to　weighted　fusion　of　diverse　features.　We　present　a　weighted　sum　model　that　convolutional　responses　from　different　convolutional　layers　are　weighted　and　summed　to　obtain　the　final　response　score.　Secondly,　we　introduce　a　context　based　updating　strategy　for　high　confidence　model　update　to　avoid　samples　corruption　and　model　drift.　The　experimental　results　on　the　challenging　OTB　dataset　demonstrate　that　the　proposed　method　is　more　competitive　than　state-of-the-art　trackers.