Abstract
Discriminative correlation filter (DCF) based trackers have recently achieved excellent performance with great computational efficiency. However, DCF based trackers suffer boundary effects, which result in unstable performance in challenging situations exhibiting fast motion. In this paper, we propose a novel method to mitigate this side-effect in DCF based trackers. We change the search area according to the prediction of target motion. When the object moves fast, broad search area could alleviate boundary effects and reserve the probability of locating the object. When the object moves slowly, narrow search area could prevent effect of useless background information and improve computational efficiency to attain real-time performance. This strategy can impressively soothe boundary effects in situations exhibiting fast motion and motion blur, and it can be used in almost all DCF based trackers. The experiments on OTB benchmark show that the proposed framework improves the performance compared with the baseline trackers.
Introduction
Visual object tracking is a fundamental task in computer vision, requiring a model to follow a target across video frames despite challenges like occlusion, illumination variation, and fast motion. Discriminative Correlation Filter (DCF) trackers have emerged as a dominant paradigm due to their exceptional speed, achieved by performing heavy computations in the frequency domain using Fast Fourier Transforms (FFT).
Despite their efficiency, DCF trackers are plagued by "boundary effects." Because FFT-based correlation assumes periodic repetition of image patches, the edges of the search window contain artificial transitions. This leads to model degradation as the tracker learns from "fake" negative samples at the borders and a limited search range that fails when objects move quickly toward or beyond the window's edge. Existing solutions often penalize these boundaries in the learning objective, but this typically breaks the closed-form mathematical solution, significantly reducing tracking speed. This paper introduces a self-adaptive search area (SASA) framework that dynamically adjusts the search window size to mitigate these effects without sacrificing real-time performance.
Proposed Method: Self-Adaptive Search Area (SASA)
The core innovation of the SASA framework is the dynamic adjustment of the search area based on predicted target motion. By expanding the search area during fast motion and shrinking it during slow motion, the tracker balances the need for a wide search range with the need to exclude background noise.
Motion Estimation and Movement Criterion
To predict motion, the system employs polynomial curve fitting. Specifically, it uses a second-order quadratic curve to model the object's position over the preceding five frames. This allows the tracker to estimate the current velocity of the target. A movement criterion, denoted as $\zeta$, is then calculated to represent the object's velocity relative to its initial size. This normalized metric provides a robust indicator of whether the current search area is sufficient.
Hysteresis Thresholding for Stability
A naive approach to resizing would be to switch sizes based on a single velocity threshold. however, this would cause the filter size to jitter frequently, leading to model instability. To prevent this, the authors implement a hysteresis function with "entangled" thresholds.
As shown in the threshold evolution diagrams, the search area only expands if the motion exceeds a high threshold and only shrinks if it falls below a significantly lower threshold for a sustained period. This "memory" effect ensures that the tracker remains stable and does not prematurely discard learned appearance information due to minor fluctuations in speed.
Frequency Domain Resizing
To maintain computational efficiency, resizing is not performed by simple spatial cropping. Instead, the framework utilizes trigonometric polynomial interpolation to resize the filter directly in the frequency domain. This approach preserves historical information within the filter and ensures that the tracker retains its closed-form solution, allowing it to operate at high frame rates.
Experimental Results
The proposed SASA framework was evaluated on the Online Tracking Benchmark (OTB-100), which includes 100 video sequences with various challenging attributes. The framework was integrated into three baseline trackers: DCF, DSST, and SAMF.
Performance Gains
The SASA-enhanced trackers showed consistent improvements across the board:
- Accuracy: The DSST tracker saw an absolute improvement of 4.4% in precision and 2.5% in success rate.
- Fast Motion Scenarios: The gains were most pronounced in sequences characterized by fast motion, where the SASA versions of SAMF, DSST, and DCF achieved precision gains of 6.7%, 8%, and 5.5% respectively.
- Computational Efficiency: The method remains remarkably fast. The DCF_SASA variant achieved a speed of 363 frames per second (fps), while the more complex DSST_SASA maintained a real-time rate of 28.9 fps.
Conclusion
The Self-Adaptive Search Area (SASA) framework provides a generic, plug-and-play solution to the boundary effect problem in DCF trackers. By intelligently scaling the search window based on target dynamics and ensuring stability through hysteresis, the method significantly improves tracking robustness in fast-motion scenarios while preserving the high-speed advantages of correlation filters. Future work may involve integrating these adaptive strategies into modern deep-learning-based tracking architectures.
Summary updated