Abstract: Video target tracking is one of the research topics in computer vision.Convolution neural network(CNN)based tracking methods have achieved significant improvement in recent years.However,dimension disaster,information redundancy,noise interference,and different representation abilities of different layers of the convolution feature are still challenges.To mitigate the above-mentioned issues,we propose to select convolutional feature centers adaptively and employ them to perform object tracking.Firstly,we utilize distance matrix and affinity propagation theory to iteratively select feature centers to compress the feature dimension and reduce the computational load in model training process.Secondly,convolutional features from multiple layers are used to train multiple trackers to determine the target state jointly,and the weight of each tracker is updated based on its real-time decision loss.By doing so,the proposed method can effectively eliminate the feature redundancy and noise interference between convolution features.Moreover,it also can improve the robustness and target discrimination ability of convolution features.Finally,extensive experiments conducted on publicly available datasets demonstrate the effectiveness of the proposed tracking method.Moreover,the efficiency is greatly improved while ensuring the tracking success rate and precision rate.