A dual-channel correlation filtering tracker for real-time tracking based on deep features of improved CaffeNet and integrated manual features

This study proposed an adaptive dual-channel correlation filtering visual tracking approach based on deep features of improved CaffeNet and integrated manual features (ADCF for short). Firstly, to deal with the issue that the fixed learning rate in the update process of correlation filter can easily mix in too many invalid features and increase the burden of computation, we present a modified correlation filter learning rate, which can adjust its value according to the relationship between the CF-response value of the current candidate region and the given threshold value. Secondly, we integrated multiple manual features with fuzzy logic rule to combine their strengths including strong resistance to size change, rotation interference and illumination variation. Thirdly, we promoted the standard CaffeNet framework by introducing batch normalization, optimizing the nonlinearity of the activation function and accelerating the efficiency of the dropout layer with SVM. Finally, we parallel the depth features extracted by the improved CaffeNet with the proposed manual fusion features to form a dual-feature channel, thus obtaining a strategy for parallel extraction of visual features based on depth features and fusion of manual features. This dual-channel feature processing strategy combines the advantages of manual features and deep features, which can balance the details and semantic information of the target, making the tracker more robust in tracking diverse interference scenarios such as target rotation, lighting changes, scale changes and complex backgrounds. We visualized and numerically compared the tracking performance of the proposed method with other cutting-edge visual trackers using multiple classic tracking evaluation benchmarks, including OTB2015, VOT2018, TrackingNet and LaSOT. The tracking performance under diverse interference factors demonstrated that our proposed tracker has good performance in terms of accuracy, robustness and efficiency.