Abstract
Visual object tracking has long been dominated by Correlation Filter (CF) paradigms, which suffer from inherent boundary effects due to periodic assumptions. This paper introduces the Gaussian Process Regression Tracker (GPRT), which reformulates tracking as a regression task. By leveraging Gaussian Processes, GPRT effectively eliminates boundary effects without sacrificing the performance-enhancing kernel trick. Experimental results on OTB-2013 and OTB-2015 benchmarks demonstrate that GPRT sets a new state-of-the-art for trackers using hand-crafted features.
Introduction
Correlation Filter (CF) trackers are widely used for their computational efficiency and accuracy. However, their reliance on the Fast Fourier Transform (FFT) necessitates a periodic assumption of the image signal, leading to the "boundary effect." This effect introduces inaccurate samples at the search window's edges, degrading tracking performance. While methods like SRDCF attempt to mitigate this, they often lose the ability to use the "kernel trick," which is vital for mapping features into high-dimensional spaces to distinguish targets from complex backgrounds.
The Gaussian Process Regression Tracker (GPRT) is proposed as a conceptually natural solution. It bypasses the boundary effect entirely by moving away from the circular shift assumption of CFs, while fully retaining the power of kernel-based learning.
Methodology
Gaussian Process Regression for Tracking
GPRT treats tracking as a regression problem where the goal is to predict a response map for a target's location in a new frame. Instead of learning a filter in the Fourier domain, GPRT constructs a Gaussian covariance matrix from training samples. This approach allows for a more flexible search area and more precise localization since it does not rely on periodic signal assumptions.
Efficient Online Updates
Standard Gaussian Process Regression is computationally intensive for online tasks. To make it viable for real-time tracking, the authors identified a mathematical equivalence between the GPR posterior mean and Kernel Ridge Regression (KRR) estimates. This allowed for the development of two efficient update schemes:
- GPRTE (Approximate): A high-speed version for rapid appearance model updates without storing all previous frame data.
- GPRT (KRR-based): A robust version utilizing dual-space optimization to adapt to target changes over time efficiently.
GPU Acceleration
To address the bottleneck of large covariance matrix calculations, the authors implemented parallelized matrix construction on GPUs. This optimization provides a theoretical 3000x speedup over standard CPU implementations, enabling the tracker to operate at practical speeds (approximately 5-7 FPS on a TITAN X).
Experimental Evaluation
The GPRT framework was evaluated on the OTB-2013 and OTB-2015 benchmarks, which include challenges such as occlusion, fast motion, and illumination variation.
Quantitative Results
GPRT consistently outperformed existing trackers using hand-crafted features (HOG and Color-Names). On OTB-2013, it achieved a mean overlap precision of 84.1%, a significant gain over previous leaders like ECOHC. On the larger OTB-2015 dataset, it maintained its lead with 79.2% precision.
Qualitative Performance
Qualitative analysis shows that GPRT is particularly robust in scenarios involving blurred foregrounds or similar-looking distractors—cases where traditional CF trackers often fail due to boundary-induced noise. By avoiding periodic assumptions, GPRT maintains a more accurate representation of the target and its immediate background.
Discussion and Conclusion
GPRT demonstrates that a regression-based approach using Gaussian Processes can effectively solve the boundary effect problem in visual tracking. By integrating the kernel trick and optimizing for GPU execution, the authors achieved state-of-the-art results with simple hand-crafted features. Future work includes integrating deep convolutional features and exploring automated hyperparameter optimization to further enhance adaptation and accuracy.
Summary updated