Abstract
Object tracking in real-world environments is often hampered by model mismatch, where the underlying generative assumptions of filters (e.g., Kalman or Particle Filters) fail to account for outliers or non-Gaussian noise. This paper proposes an explanatory framework that redefines tracking as an online learning task. By utilizing the NormalHedge algorithm, the framework minimizes cumulative regret rather than maximizing likelihood. Experimental results on 1D tracking data demonstrate that this approach is significantly more robust than traditional Bayesian methods, maintaining stable performance even as outlier probability increases.
Introduction
In classical tracking, a generative framework is typically assumed. Researchers define rules for how an object moves and how sensors produce observations. Algorithms like the Bayesian filter then attempt to find the most likely state given these rules. However, generative models are notoriously fragile. If sensor noise differs even slightly from the programmed model, or if unexpected outliers occur, performance can degrade rapidly. This paper identifies this weakness and proposes a shift toward an explanatory framework that prioritizes how well a path explains observations over its theoretical probability.
The Explanatory Framework
Instead of asking "What is the probability this path generated these observations?", the explanatory framework asks "How well does this path explain the observations?" This approach is built on several key components:
- Loss Functions: The framework uses loss functions to penalize paths that deviate from data or physical expectations, rather than relying on probability distributions.
- NormalHedge Algorithm: At its core is NormalHedge, an online learning algorithm that maintains a distribution over many possible paths. It assigns weights based on regret—the difference between the algorithm's performance and that of a specific path.
- Sequential Monte Carlo (SMC) Approximation: To handle the computational complexity of tracking numerous paths, an efficient SMC version is used. Unlike standard particle filters that weight by likelihood, this version prunes paths with zero or negative regret and resamples in high-density regions.
- Discounted Regret: A discount factor is applied to prioritize recent observations, allowing the tracker to recover quickly from errors or changes in object behavior.
Experimental Evaluation
The algorithm was evaluated using simulated 1D tracking data. The task involved tracking an object moving in a 1D space where the signal was a "square pulse" (mimicking a face detector score). To test robustness against model mismatch, the researchers introduced a mixture of Gaussian noise and heavy-tailed outliers ($\rho$). This setup simulates real-world scenarios where detectors might intermittently produce high scores for background objects.
Results and Analysis
The performance of the NormalHedge (NH) approach was compared against a standard Bayesian filter and a Particle Filter (PF).
As shown in Figure 3, when the model is perfectly matched ($\rho = 0$), the Bayesian filter is slightly superior, which is expected as it is mathematically optimal under perfect conditions. However, as outliers are introduced ($\rho = 0.2$), the Bayesian and Particle Filters fail significantly. While the Bayesian filter's trajectories become highly erratic, the NormalHedge algorithm remains remarkably stable.
Quantitatively, the results in Figure 5 highlight this robustness. At $\rho = 0.2$ and $\sigma_o = 8$, the Bayesian RMSE jumps to 43.99, and the Particle Filter often diverges completely (RMSE ~81). In contrast, the NormalHedge algorithm maintains a stable RMSE of approximately 14.15. This "parameter-free" weighting logic makes the explanatory framework much harder to "trick" with outliers compared to traditional likelihood-based methods.
Conclusion
The explanatory framework offers a robust alternative to generative tracking models by focusing on regret minimization. While the current study focused on 1D state spaces, the underlying theory is extensible. Future work includes testing the computational efficiency of the SMC approximation in higher-dimensional environments and developing a formal theory for optimal loss function design. Despite being a pre-deep learning approach, the regret-based logic provides a mathematically grounded and distinct alternative to modern data-driven tracking methods.
Summary updated