Using Elimination Methods to Compute Thermophysical Algebraic Invariants from Infrared Imagery

AAAI 1996

TL;DR

This paper introduces a novel method for computing invariant features from infrared (IR) imagery by leveraging thermophysical properties of objects, going beyond traditional reflection and geometry-based approaches. By applying algebraic elimination techniques to an energy balance equation, they derive features that are robust to changes in scene conditions, enabling more reliable object recognition.

What Problem Does This Paper Address?

Traditional object recognition often relies on features derived from visible light imagery, which are primarily sensitive to surface reflectance and geometry. However, long-wave infrared (LWIR) imagery, which senses thermal radiation, is influenced by a broader range of object properties, including internal composition, thermal state, and environmental factors like ambient temperature and wind speed. This complexity makes it challenging to extract features that remain consistent across different imaging conditions (e.g., time of day, weather). The paper addresses this by developing a method to compute "Thermophysical Algebraic Invariants" (TAIs) that are invariant to these scene variations, thus improving object recognition in LWIR.

Methods

The core of the paper lies in applying algebraic elimination techniques to derive invariant features from an energy balance equation at the object's surface.

  1. Energy Balance Equation: They start with a fundamental equation describing energy exchange at an object's surface, accounting for absorbed energy and energy lost through radiation, convection, conduction, and storage.
  2. Algebraic Formulation: This physical equation is translated into a polynomial equation where object properties (like thermal conductivity, specific heat) are coefficients, and scene parameters (like ambient temperature, wind speed) are variables.
  3. Transformation Equations: To achieve invariance, they model how these variables change between different scenes using a transformation matrix (a subgroup of the general linear group).
  4. Elimination Methods: The key innovation is using algebraic elimination techniques (specifically, referencing Gröbner basis and resultant computations) to remove the scene-dependent transformation parameters from the equations. This process yields "separable invariant relations."
  5. Invariant Extraction: From these relations, they extract absolute invariants, which are rational functions of object parameters that remain constant regardless of scene conditions. The paper outlines a specific algorithm for this extraction, ensuring a unique invariant is identified.

Datasets

The paper uses real LWIR imagery of vehicles (van, two types of trucks, military tank, car) acquired at different times of the day and over multiple days. Specific points were selected on the surfaces of these vehicles to represent various materials and surface normals. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Beyond Reflectance: LWIR imagery offers a richer source of information than visible light by capturing thermal properties. This paper demonstrates how to exploit this by linking image data to thermophysical object characteristics.
  • Physics-Based Invariants: By grounding invariant feature derivation in the physical principle of energy conservation, the resulting features have a strong theoretical justification and are expected to be more robust.
  • Algebraic Elimination Power: The paper highlights the power of algebraic elimination theory as a general framework for computing invariants for complex, non-linear systems, extending its application beyond purely geometric problems.
  • Scene-Condition Robustness: The derived TAIs are shown to be invariant to changes in environmental conditions (ambient temperature, wind speed, solar irradiation), which is crucial for reliable recognition in real-world scenarios.

Key Results

  • The paper demonstrates the derivation of two specific non-trivial thermophysical algebraic invariants, denoted as $I_1$ and $I_2$.
  • Experimental results on real LWIR imagery of vehicles show good intra-class stability (features remain consistent for the same object under varying conditions) and good inter-class separation (features differ significantly between different vehicle classes).
  • Table 1 shows that for a specific feature ($I_1$) and point set on a truck, the quality factor (std/mean) is generally low (e.g., 0.0026 to 0.1984), indicating stability.
  • Table 2, illustrating mistaken hypotheses, shows that feature values for incorrect object class hypotheses are often significantly different from those of correct hypotheses, suggesting good discriminative power.

Limitations and Future Work

  • Parameter Interpretation: The paper acknowledges that one of the derived coefficients ($a_7$) in the algebraic formulation doesn't have a direct physical interpretation but is included to facilitate the algebraic manipulation and achieve the desired invariance.
  • Hypothesize-and-Verify: The proposed feature computation scheme is suitable for a hypothesize-and-verify object recognition system. The paper suggests that further work could explore integrating these TAIs with geometric features for a more comprehensive recognition system.
  • Computational Complexity: While not explicitly detailed, algebraic elimination methods can be computationally intensive. Future work might focus on optimizing these computations for real-time applications.
  • Generalization: The current work focuses on specific thermophysical properties and a particular energy balance model. Extending this to a wider range of materials, thermal phenomena, and more complex object geometries could be a future direction.

Summary updated

BibTeX
@inproceedings{1996using,
  title = {{Using Elimination Methods to Compute Thermophysical Algebraic Invariants from Infrared Imagery}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}

Loading related content...

Figure