A Video Based Fire Smoke Detection Using Robust AdaBoost

Xuehui Wu, Xiaobo Lu, Henry Leung2018 · DOI

TL;DR

This paper introduces a novel video-based fire and smoke detection system that leverages a "Robust AdaBoost" (RAB) classifier. By combining static and dynamic visual features and enhancing the AdaBoost algorithm to handle noisy data and improve robustness, the system achieves high accuracy and a low false alarm rate, making it suitable for real-world surveillance applications.

What Problem Does This Paper Address?

Early detection of fire and smoke is crucial for safety and economic reasons. While traditional point detectors are effective indoors, they struggle with large outdoor areas. Video-based detection using existing CCTV infrastructure offers a cost-effective solution for these open spaces. However, existing video-based systems face challenges like high false alarm rates due to variations in smoke density, lighting, and interfering objects. Furthermore, modeling smoke and fire accurately is difficult, and many current image processing methods fail to capture dynamic features effectively. This paper aims to address these limitations by developing a more robust and accurate video-based fire smoke detection system.

Methods

The core of this work lies in two main areas: feature extraction and a novel classification approach.

Feature Extraction: The system extracts a comprehensive set of features from video frames:

  • Static Features: These include color moments (mean, standard deviation, skewness in HSV color space), texture features (using Local Binary Patterns - LBP), wavelet energy (from Gabor wavelets), Edge Orientation Histogram (EOH) for dominant and symmetry features, irregularity (based on edge perimeter and area), and sparsity (counting non-zero pixels in foreground regions).
  • Dynamic Features: These capture motion characteristics: motion direction, change in motion direction, and motion speed.

These features are extracted from motion regions identified using Robust Orthonormal Subspace Learning (ROSL) for foreground segmentation.

Robust AdaBoost (RAB) Classifier: The authors propose a modified AdaBoost algorithm, termed Robust AdaBoost (RAB), to improve classification accuracy and robustness. The key enhancements include:

  • Modified Weight Update: The weight update mechanism for base classifiers is modified to be less sensitive to small errors, allowing for better handling of classifiers with errors slightly above 0.5. This is achieved by using a new weight calculation formula: $a_m^* = \frac{1}{n} \ln \left(\frac{1+e^{-a_m}}{1-e^{-a_m}}\right)$, where $a_m$ is the conventional weight and $n$ is a parameter controlling robustness.
  • RBFSVM as Base Classifier: Radial Basis Function Support Vector Machines (RBFSVM) are used as base classifiers. The authors explore adjusting the Gaussian width ($\sigma$) to improve diversity and reduce errors.
  • Diversity and Sample Denoising: To further enhance performance, the method incorporates diversity measures to ensure component classifiers are uncorrelated and a sample denoising technique to reduce the weights of potentially noisy training data.
  • Improved Initial Weights: For binary classification, the initial weights are adjusted to balance negative and positive samples.

The overall framework involves motion detection, feature extraction, model training using the proposed RAB classifier, and finally, classification.

Datasets

The paper utilizes two types of datasets:

  • Pre-collected Image Datasets: Mentioned as a source for training data, though specific datasets are not detailed beyond their general use for static feature extraction.
  • Public Video Clips: The authors collected smoke and non-smoke videos from public sources (e.g., http://signal.ee.bilkent.edu.tr/VisiFire/Demo/SampleClips and http://imagelab.ing.unimore.it/visor). These clips cover indoor and outdoor scenarios with varying illumination and distances. Seven smoke videos and ten non-smoke videos were used for training and testing.
  • UCID Repository Datasets: Used for comparing the proposed classifier with other methods. Specific datasets mentioned include Parkinsons, Ionosphere, Vote, Sonar, Wdbc, and Wpbc.

Key Insights and Ideas

  • Feature Combination is Key: The study strongly suggests that combining a diverse set of static and dynamic features significantly improves detection performance compared to using either type alone.
  • Robustness over Strict Error Thresholds: The "Robust AdaBoost" approach highlights that it's more beneficial to have a classifier that can handle a wider range of errors (even slightly above 0.5) rather than discarding it, leading to a more stable and effective ensemble.
  • Diversity Enhances Ensemble Learning: Explicitly incorporating diversity among base classifiers is shown to be crucial for boosting algorithms, preventing overfitting and improving generalization.
  • Handling Noisy Data: The denoising of initial weights for samples that are consistently misclassified is a practical step to improve robustness against potentially noisy training data.
  • RBFSVM as a Strong Base Classifier: The paper demonstrates that RBFSVM, when properly tuned and integrated into a boosting framework, can be a powerful base classifier for complex tasks like fire smoke detection.

Key Results

The proposed Robust AdaBoost classifier (RAB-RBFSVM+) demonstrates superior performance:

  • High Accuracy and Low False Alarms: The method achieved an accuracy of 93.31% with a false alarm rate of 0.70% using the TCW+MDC feature combination. Another combination, TCW+MDC+MS, yielded 91.25% accuracy with an even lower false alarm rate of 0.31%.
  • Outperforms Existing Methods: Compared to prior works (Cai et al. [29] and Wu et al. [34]), the proposed method achieved significantly higher accuracy (e.g., 93.31% vs. 73.41%) and lower false alarm rates (0.70% vs. 5.52%).
  • Effective Feature Combinations: The study found that combinations like CES+MD+MS, TCW+MDC, and TCW+MDC+MS provided the lowest errors.
  • Improved Boosting: RAB-RBFSVM+ consistently outperformed conventional AdaBoost across various base classifiers, especially with more complex ones like Linear SVM and RBFSVM.

Limitations and Future Work

  • Computational Complexity: While the proposed method is effective, the extraction of a large number of features and the iterative nature of AdaBoost can lead to significant computational demands, which might be a concern for real-time applications on resource-constrained devices.
  • Parameter Tuning: The performance of RBFSVM as a base classifier is sensitive to parameters like $\sigma$ and $C$. While the paper discusses adjusting $\sigma$, finding optimal parameters for both the base classifier and the robust AdaBoost framework can still be challenging.
  • Dataset Limitations: Although public video clips were used, the paper mentions the difficulty in collecting diverse datasets. Future work could involve testing on larger and more varied datasets to further validate the system's generalization capabilities.
  • Further Improvements: The authors suggest exploring more advanced deep learning techniques in conjunction with their proposed robust boosting framework to potentially achieve even higher performance. Investigating real-time implementation on embedded systems is also a logical next step.

Summary updated

BibTeX
@misc{wu2018video,
  author = {Xuehui Wu and Xiaobo Lu and Henry Leung},
  title = {{A Video Based Fire Smoke Detection Using Robust AdaBoost}},
  year = {2018},
  doi = {10.3390/s18113780},
}
Figure