A research team headed by Prof. ZHAO Hui from Xi'an Institute of Optics and Precision Mechanics (XIOPM) of the Chinese Academy of Sciences devised an innovative robust tracker that utilized Discriminative Correlation Filters (DCFs) and hierarchically infers the target's state through adaptive multi-feature fusion. The study was published in Expert Systems with Applications.
Visual tracking is a vital research topic in computer vision, with practical applications spanning human-computer interaction, aerial reconnaissance, etc. DCFs have recently garnered attention for their remarkable efficiency and effectiveness. However, existing trackers often have trouble fully mining the structural complementarity and diversity among various features, resulting in a decline in discriminability in complex scenarios.
In this study, researchers developed a context constraint and sparse learning based on DCFs copied with a novel coarse-to-fine high-confidence learning strategy for robust tracking. Firstly, the discriminative contexts were directly incorporated into the DCFs framework, which enhanced the discriminability under diverse scenes. Meanwhile, unexpected crests caused by frequent changes in the appearance of the target were calmed down by performing sparse responses.
Furthermore, researchers utilized the Least Absolute Shrinkage and Selection Operator regression to reduce the redundancy and irrelevance of high-dimensional multi-channel features, while enhancing the interpretability of the model, thereby achieving a compact target representation. Then, researchers introduced a coarse-to-fine high-confidence heuristic tracking strategy that enhances complementary characteristics between the hierarchical Convolution features and hand-crafted features.
Kalman filter is typically used to track when the state is unreliable. Therefore, researchers employed the Kalman filter to supervise coarse localization to improve the positioning accuracy and introduced a multi-peak detection mechanism to supervise the precise localization.
This study fully explores the inherent relationship between structural among different features and context-aware sparse tracking framework，greatly premoting the progress of visual tracking.
(Available online 20 December 2024)
Fig. Main flow chart of discriminative correlation filter based on context awareness and sparse learning. (Image by SU Yinqiang)
Download：