Fuzzy foreground detection for infrared videos

Fida El Baf, Thierry Bouwmans, Bertrand VachonWorkshops @ CVPR 2008 · DOI

Introduction

Foreground detection—the process of identifying moving objects against a static background—is a critical component of video surveillance. In the infrared (IR) spectrum, this task is complicated by environmental factors such as thermal noise, rain, and moving clouds, which can trigger false positives. While statistical models like the Mixture of Gaussians (MOG) are widely used, they are often computationally demanding and memory-intensive. This paper presents a robust alternative that utilizes fuzzy logic and the Choquet integral to fuse multiple image features, providing high accuracy with significantly lower overhead.

System Architecture

The proposed system follows a structured pipeline consisting of background initialization, foreground detection, and background maintenance. Unlike traditional "crisp" systems that make binary decisions (foreground vs. background), this approach maintains a degree of uncertainty throughout the process.

General architecture of the proposed foreground detection system, illustrating the flow from video frames to the final foreground mask.

The system begins by initializing a background model from the first $N$ frames. For subsequent frames, it calculates similarity measures between the current frame and the model to detect foreground pixels. A key innovation is the fuzzy background maintenance step, which updates the model based on the confidence level of the detection, allowing the system to adapt smoothly to environmental changes.

Feature Extraction and Fusion

To improve robustness against noise, the algorithm considers two distinct features for every pixel:

  1. IR Intensity: A measure of thermal brightness based on pixel ratios.
  2. Texture (LBP): Local Binary Patterns are used to capture the spatial structure of the scene. LBP is particularly effective because it is invariant to monotonic gray-scale changes, making the system resilient to sudden illumination shifts caused by weather.
The weighting matrix used in the calculation of Local Binary Patterns (LBP) to encode local texture information.

These features are fused using the Choquet fuzzy integral. While previous research has explored the Sugeno integral, the authors argue that the Choquet integral is superior for this application. Because similarity scores represent cardinal values (numerical magnitude) rather than just ordinal rankings, the Choquet integral's ability to perform standard arithmetic operations during aggregation leads to more precise fusion.

Experimental Results

The method was evaluated using the OTCBVS benchmark dataset, which includes challenging IR sequences with rain and heavy cloud cover.

  • Performance: The proposed fuzzy system achieved a similarity score of 60%, surpassing an improved MOG implementation which scored 58%.
  • Efficiency: The algorithm is characterized by low memory requirements and high processing speed. While MOG is often categorized as "intermediate" in terms of resource consumption, this fuzzy approach offers performance comparable to a simple running average model.
  • Robustness: Qualitative analysis showed that the system effectively filters out "ghost" movements in rainy conditions that typically cause high false-alarm rates in standard models.

Conclusion

By moving away from binary classification and leveraging the mathematical strengths of the Choquet integral, this research demonstrates that fuzzy logic can provide a powerful framework for infrared video analysis. The resulting system is not only more accurate than traditional statistical models in noisy environments but also significantly more efficient, making it suitable for real-time surveillance applications on resource-constrained hardware. Future work may involve integrating additional features like motion vectors or edge detection to further enhance detection in highly dynamic scenes.

Summary updated

BibTeX
@inproceedings{baf2008fuzzy,
  author = {Fida El Baf and Thierry Bouwmans and Bertrand Vachon},
  title = {{Fuzzy foreground detection for infrared videos}},
  booktitle = {Workshops @ CVPR},
  year = {2008},
  doi = {10.1109/cvprw.2008.4563057},
}
Figure