Object Identification in a Bayesian Context

Timothy Huang, Stuart RussellIJCAI 1997

Introduction

In the field of artificial intelligence, a critical distinction exists between recognition (categorizing an object, e.g., "that is a car") and identification (determining if an observation represents a specific individual, e.g., "that is my car"). While classical logic often treats identity as a simple binary equality ($A=B$), real-world sensor data is inherently noisy and propositional. This paper presents a formal Bayesian framework to address the challenge of maintaining object identity across space and time without the aid of unique identifiers like license plates.

A Bayesian Framework for Identity

The authors propose that identity should be defined through a probability space over "life histories" or trajectories. An observation is considered identical to another if both are most likely to belong to the same underlying life history. This approach bridges the gap between high-level logical identity and low-level probabilistic sensing.

Conceptual representation of object trajectories (life histories) in space-time. Observations are matched based on their likelihood of belonging to the same trajectory.

Modeling Appearance and Noise

Rather than relying on perfectly calibrated sensor models, the system utilizes a unified "appearance probability." This model calculates the likelihood of observing specific features (such as vehicle dimensions, color, and lane position) at a downstream camera given an upstream observation.

One of the paper's key insights is the limited information content of these features. Due to sensor noise and environmental factors like lighting, the combined data from 11 features (including hue, saturation, and size) only provides approximately 3 to 4 bits of information. The system accounts for this by updating its appearance models online, using exponential forgetting to adapt to changing conditions.

Correlation of visual feature data (specifically hue) between the upstream and downstream cameras, illustrating the noise and limited information content available for matching.

The Matching Algorithm

To resolve the global problem of matching $N$ upstream vehicles to $M$ downstream vehicles, the authors employ a weighted bipartite matching algorithm ($O(n^3)$). To ensure the reliability of these matches, they introduce a "Leave-One-Out" heuristic: the system temporarily forbids a specific match and recalculates the global assignment. If the total likelihood of the alternative assignment is significantly lower, the original match is deemed reliable.

This mathematical approach aligns with human intuition: it is far easier to confidently identify a "rare" vehicle (e.g., an orange truck) than a common one (e.g., a white sedan), as the prior probability of a similar "replacement" vehicle appearing at the exact right time is extremely low.

Experimental Results and Implications

The framework was tested using real-world data from the Roadwatch project on Highway 99 in Sacramento, California. Human operators established ground truth by manually matching vehicles across two miles of freeway.

The system demonstrated a robust trade-off between coverage and accuracy:

  • Precision: Achieved 100% accuracy at 14% coverage.
  • Recall: Reached 80% coverage at 50% accuracy.
  • Aggregate Metrics: Despite individual matching noise, the system calculated average link travel times to within 1% accuracy.

While currently applied to traffic surveillance, the authors suggest this Bayesian identification framework has broad potential applications, including database deduplication and tracking individuals across disparate records in large-scale administrative systems.

Summary updated

BibTeX
@inproceedings{huang1997object,
  author = {Timothy Huang and Stuart Russell},
  title = {{Object Identification in a Bayesian Context}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1997},
}
Figure