Speeding up the tracking and updating of the convolutional residual tracking networks

Abstract

The convolutional residual tracking networks (CREST) uses a single-layer convolutional network as the implicit correlation filter, which can perform end-to-end training and has the advantages of simple model structure and high tracking accuracy. However, the forward and backward convolution computation in its base layer is too excessive, resulting in its slow tracking speed. In addition, the back-propagation algorithm is adopted as the fine-tuning method during online updating, thus the parameter updating efficiency is quite slow. When the object appearance changes rapidly, there is a large risk of losing the tracked object. In this paper, we improve the residual convolutional network model and its training method. The improved model explicitly uses the base layer as a correlation filter and uses the frequency domain parameters to describe the base layer. When updating online, the base layer is pre-trained using the standard learning method of the correlation filter to efficiently learn the main information of the object appearance, and then the back-propagation algorithm is utilized to optimize the entire model. Thus, both the forward and backward convolution computation in the improved base layer can be conducted efficiently in the frequency domain, which significantly reduces the amount of computation and improves the tracking speed and online update efficiency. Extensive experiments on the evaluation data sets validate the effectiveness of the improved method.