Abstract
Correlation filter has drawn increasing interest in visual tracking due to its high efficiency, however, it is sensitive to partial occlusion, which may result in tracking failure. To address this problem, we propose a novel local-global correlation filter (LGCF) for object tracking. Our LGCF model utilizes both local-based and global-based strategies, and effectively combines these two strategies by exploiting the relationship of circular shifts among local object parts and global target for their motion models to preserve the structure of object. In specific, our proposed model has two advantages: (1) Owing to the benefits of local-based mechanism, our method is robust to partial occlusion by leveraging visible parts. (2) Taking into account the relationship of motion models among local parts and global target, our LGCF model is able to capture the inner structure of object, which further improves its robustness to occlusion. In addition, to alleviate the issue of drift away from object, we incorporate temporal consistencies of both local parts and global target in our LGCF model. Besides, we adopt an adaptive method to accurately estimate the scale of object. Extensive experiments on OTB15 with 100 videos demonstrate that our tracking algorithm performs favorably against state-of-the-art methods.
Introduction
Correlation Filters (CF) have become a staple in visual tracking because they are incredibly fast, operating in the Fourier domain. However, they have a significant weakness: partial occlusion. When a target is partially blocked—such as a person walking behind a lamp post—a standard CF tracker often begins learning background features, leading to "model drift" and eventual tracking failure.
While part-based trackers have been proposed to mitigate this, they often fail because they ignore the global structure; individual parts may drift away from each other because the model lacks a unified geometric understanding. The Local-Global Correlation Filter (LGCF) bridges this gap by creating a model that simultaneously monitors both the global target and its constituent local parts.
Methodology: The LGCF Framework
The LGCF model treats tracking as a joint optimization problem that preserves the object's structural integrity while remaining flexible enough to handle deformations. Key components include:
- Local-Global Coupling: Instead of training independent filters, LGCF assumes the motion model of a local part is equal to the global motion model plus a small deviation. This is mathematically enforced through a sparse constraint ($L_1$-norm penalty), which acts like an "elastic" bond—keeping parts attached to the main body while allowing enough slack for natural movement and deformation.
- Temporal Consistency: To prevent drastic model changes between frames, a regularization term penalizes large weight fluctuations over time, significantly reducing drift.
- Adaptive Scale Estimation: LGCF estimates scale changes by monitoring the relative distances between tracked local parts. If the parts spread out, the object is likely getting larger.
Experimental Results
The model was evaluated on the OTB15 (Object Tracking Benchmark 2015), consisting of 100 challenging video sequences featuring illumination variation, rotation, and heavy occlusion.
LGCF achieved a precision score of 0.782 and a success rate of 0.585, outperforming 34 state-of-the-art trackers. Notably, it showed a 9% improvement in precision and an 11% improvement in success rate over the standard Kernelized Correlation Filter (KCF). Qualitative results demonstrate exceptional performance in scenarios involving in-plane rotation and illumination changes where global-only trackers typically fail.
Conclusion and Future Work
While LGCF provides robust tracking, it faces a computational trade-off, running at approximately 8 FPS on a high-end CPU due to the iterative ADMM optimization required. Additionally, its performance in "Low Resolution" scenarios is more modest, as there is often insufficient texture to define distinct local parts. Future work may involve integrating deep features (CNNs) to further enhance accuracy, potentially at the cost of additional computational complexity.
Summary updated