Robust Visual Tracking via Multilayer CaffeNet Features and Improved Correlation Filtering

Abstract

For problems related to the robust tracking of visual objects in various challenging tracking conditions, a robust visual tracking method based on multilayer convolutional features and correlation filtering is proposed. To solve the problems of mean deviation and insufficient discrimination ability in traditional convolutional neural networks (CNN), this study proposes randomized parametric rectified linear units (RPReLU) as the activation function. Meanwhile, the zero-setting operation of weights in the traditional dropout process occurs randomly and fails to discriminate the features with different weights, which leads to a low learning efficiency. Therefore, this study proposes an improved dropout method based on a support vector machine (SVM), which provides a selective dropout rate to increase the manual orientation and improve the learning efficiency of the dropout process. In addition, traditional CNN trackers only employ the output of the last layer, which can effectively capture semantic features but not spatial features. To solve this problem, we propose to use the rich features of the multiple convolution layers of CaffeNet as the target representation. Furthermore, we propose an improved correlation filter to further improve the tracking performance and improve the tracker's capability of dealing with scale changes, which effectively solves the problem of adaptive estimating of target size. The extensive experimental evaluations have been carried out through the OTB2015, VOT2016 and VOT2018 datasets, proving that the proposed method is very effective in dealing with a variety of challenging factors.