Abstract
Visual tracking in non-stationary environments remains a significant challenge due to rapid changes in object appearance and frequent occlusions. This paper presents Randomized Ensemble Tracking (RET), a novel framework that treats classifier weights as random variables within a Bayesian estimation framework. By modeling weight distributions using a Dirichlet distribution and employing randomized voting, the proposed tracker dynamically adapts to changing conditions, effectively "forgetting" unreliable features while maintaining focus on visible target components.
Introduction: The Challenge of Non-Stationarity
Visual tracking-by-detection involves training an online classifier to distinguish a target from its background. However, traditional methods often struggle with non-stationarity—the phenomenon where an object's appearance evolves due to lighting shifts, rotations, or partial occlusions. Standard ensemble methods, such as online boosting, typically assume a fixed data distribution or rely on deterministic weight updates. These approaches can fail when a previously learned "positive" feature becomes misleading due to environmental changes. There is a critical need for tracking systems that can gracefully adapt their internal models by prioritizing currently reliable features over historical data.
Methodology: Randomized Ensemble Tracking (RET)
The core innovation of RET lies in its treatment of weak classifier reliability as a hidden state in a non-stationary system. The tracker decomposes the target into an 8x8 grid of local patches, each monitored by a weak classifier (a linear SVM).
Bayesian Weighting and Randomized Voting
Unlike deterministic models, RET treats the weight vector of these weak classifiers as a random variable following a Dirichlet distribution. At each frame, the system draws multiple samples (instantiations) from this distribution. Each sample generates a unique "strong classifier," and the final tracking decision is reached through a consensus vote among these randomized versions. This sampling process allows the tracker to explore model uncertainty, preventing it from prematurely converging on false positives or distractors.
Consistency-Based Updates
The model evolves through a recursive Bayesian update based on "consistency." If a weak classifier's prediction aligns with the final ensemble's consensus, its reliability score—and consequently its weight in the Dirichlet distribution—is increased. Conversely, if a classifier disagrees (e.g., when its corresponding patch is occluded), its influence is diminished.
Key Insights and Results
A primary advantage of RET is its inherent ability to handle partial occlusions without an explicit detection module. When a portion of the target is hidden, the classifiers for those specific patches begin to disagree with the visible majority, leading to a natural down-weighting of the occluded regions.
Performance Evaluation
The researchers evaluated RET across 28 challenging benchmark sequences, including the Babenko, PROST, and VTD datasets. Key findings include:
- State-of-the-Art Accuracy: RET achieved top-tier performance on 50% of the tested sequences, showing particular resilience in "distractor-heavy" environments where standard trackers often fail.
- Robustness: The Bayesian weight evolution proved significantly more effective than simple error-rate updates used in standard online boosting.
- Efficiency: The implementation maintains practical processing speeds, handling multiple scales in approximately 3 seconds per frame.
Limitations and Future Directions
While RET excels at feature adaptation, it currently lacks a formal motion model (e.g., a Kalman filter), meaning it does not utilize temporal dynamics like velocity or direction. Additionally, extremely similar distractors can still cause tracking drift due to a lack of global spatial context. Future work may involve integrating motion priors and exploring dynamic adaptation of the learning rate to further enhance robustness in highly complex scenes.
Summary updated