Aiming at the problem that early convolutional neural networks using the output of full convolution layer which have lower spatial resolution as the layer of feature extraction

that can’t locate target accurately

taking the output of the first layer and the fifth layer as the layer of feature extraction is proposed. Many convolutional layers in VGG-19 network are analyzed. Templates can get

through learning of the correlated filters and the weight of response map can be calculated according to the average peak to correlation energy（ APCE）.After that

locate the final position of the target

by adaptively weighting and fusing response map. Histogram of oriented gradient（ HOG） features are used to estimate the target scale. The experiment chooses 20 video sequences in the set of public test video and compared with other object tracking algorithms.