Learning feature-weighted regularization discriminative correlation filters for real-time UAV tracking

Traditional discriminative correlation filters used for tracking unmanned aerial vehicle objects are often disrupted by concealed noise, resulting in unstable tracking results. Various methods have been developed to search for optimal feature combinations and construct feature weight pools. However, these methods often overlook the significance of different feature channels in tracking frames. Irrespective of the availability of the effective target information, a tracker regards all feature channels similarly. This makes it challenging for the tracker to avoid learning the background noise from such feature combinations. This study proposes a channel-level feature-weighting method called learning feature-weighted regularization discriminative correlation filters (FWRDCF). By introducing feature-weighted regularization (FWR) that automatically adjusts the weights of the feature channels into each frame, the FWRDCF tracker can significantly suppress background noise. Furthermore, the alternating direction method of multipliers is used to obtain the closed-form solution of the model, thereby establishing a robust correlation filter-tracking architecture. Experiments on UAV123@10fps, UAV123, DTB70, and UAVDT demonstrated that the FWRDCF tracker achieved better tracking performance than 15 other state-of-the-art trackers. An integration study of three baselines (AutoTrack, STRCF, and BACF) reveals that the proposed FWR can be integrated with trackers with multi-channel features.