Leveraging Attention for Improved Discriminative Correlation Filters in Single Object Tracking

Abstract

Correlation Filter-based Trackers have shown impressive results in the object tracking research, outperforming classical trackers in several benchmarks. However, accurately tracking objects with deformation, fast motion, or occlusion remains a main challenge in the process of tracking. The cyclic suggestion of training samples used in correlation filter tracking usually lead to undesirable boundary effects, that significantly reduce the tracking efficiency. To address this issues, a hybrid attention-based correlation filter method (Att-DCF) is proposed for robust object tracking. By developing a discrimination filter and reducing the limitations of irrelevant features, the proposed approach provides more accurate and consistent estimated target locations. The proposed Att-DCF tracker is evaluated using the recent and standard datasets, OTB-100, Temple-Colour128, and UAV123. Compared to the baseline DCF tracker, the proposed tracker achieves an improvement of 2.71$\%$ and 1.38$\%$ in terms of area under the curve and precision measures using the OTB-100 dataset, 3.99$\%$ and 3.43$\%$ using the Temple-Colour128, and 4.72$\%$ and 1.76$\%$ using the UAV123, respectively.