Correlation Filters with Pre-position by Reconstruction Error for Visual Tracking

Correlation filter based on deep neural network is a kind of mainstream method for real-time object tracking. It combines the high efficiency of correlation filtering and the great representation ability of convolutional neural network. However, this method inherits most shortcomings of correlation filter such as boundary effects. If an object is close to the boundary of a search area due to a large displacement, the useful information will be filtered out by cosine window and padding. In order to alleviate boundary effects, we propose a coarse positioning module to fine tune the search area before cosine window and padding. The core of the proposed module is saliency detection based on reconstruction error. This enables the improved trackers to retain more object information than the prototypes. Experimental results show that our method obviously promotes the baseline model, namely DCFNet, in the case of fast motion. Due to the low computational cost of our coarse positioning module, the improved trackers still have real-time rate.