Correlation filters with adaptive convolution response fusion for object tracking

Abstract

Recently, the rich features extracted by deep learning models have been widely used under the correlation filter tracking framework and achieved great success. However, the features in different layers are often combined with fixed weights, which do not consider the different importance of different layers. In this paper, we propose a novel tracking method which can adaptively tune the weights of the convolutional responses obtained by features in different layers. We propose two adaptive weighting strategies, i.e. the cosine weighting and quadratic optimization weighting, adaptively assigning weights to each submodel, and combining multiple view submodels. Moreover, Normalized Peak Value is used to estimate the tracking reliability. Experimental results demonstrate that the proposed adaptive fusion based method can achieve comparable performance to several state-of-the-art approaches on public dataset.