Towards a solid solution of real-time fire and flame detection

Bo Jiang, Yongyi Lu, Xiying Li, Liang LinMultimedia Tools and Applications 2015 · DOI

TL;DR

This paper presents a robust, real-time fire and flame detection system for video surveillance and event retrieval. By conducting a comprehensive empirical study on a novel, large dataset, the authors identify optimal feature combinations and classifier settings, achieving state-of-the-art performance with 82% recall and 93% precision.

What Problem Does This Paper Address?

Fire detection is crucial for public safety and property protection, yet existing methods often struggle with robustness, real-time performance, and generalization across diverse scenarios. Many approaches rely on limited features, heuristic thresholds, or small, unrepresentative datasets, leading to high false alarm rates or missed detections. The authors aim to bridge this gap by developing a solid, empirically-validated solution that can reliably detect fires in real-world video streams.

Methods

The proposed system employs a three-stage pipeline:

  1. Candidate Region Proposing: A background model is used to identify potential fire regions. For static backgrounds, background subtraction is applied. For moving backgrounds, a multi-level thresholding approach, automatically adjusted based on background intensity, is used to propose candidate regions.
  2. Fire Region Classifying: This stage leverages a combination of features and a Support Vector Machine (SVM) classifier.
  3. Feature Extraction: The authors conducted an empirical study to select the most effective features. They found that a combination of LAB color histogram (global feature) and LAB+SURF (local color-texture feature) provided the best results. SURF (Speeded Up Robust Features) was chosen for its speed and parallelizability, crucial for real-time applications.
  4. Codebook and Classification: Local SURF descriptors are quantized using a codebook (generated via k-means clustering) to create a fixed-length representation. This, along with the global LAB histogram, forms the input for an SVM classifier. The RBF kernel was found to be the most effective for this task.
  5. Temporal Verifying: To further reduce false alarms from objects with similar visual characteristics (e.g., street lamps, car lights), a temporal verification step is employed. This step analyzes the stability of candidate regions over 25 frames using parameters like perimeter, area, and spatial distribution.

Datasets

The authors introduce a new dataset for fire detection, consisting of 64 video clips. This dataset was compiled by randomly downloading clips from public video-sharing websites (like YouTube and Youku) and also includes self-captured videos. It aims to cover diverse scenarios (indoor, outdoor, forest) and conditions (static/moving backgrounds, illumination changes). The dataset is noted as being significantly larger than those used in previous studies.

Key Insights and Ideas

  • Empirical Study is Key: The paper strongly emphasizes the importance of empirical study for selecting optimal features and classifiers, rather than relying solely on theoretical analysis or common assumptions.
  • LAB Color and SURF Synergy: The combination of global LAB color histograms and local LAB+SURF descriptors proved to be highly effective, outperforming individual features or other color spaces. This highlights the benefit of integrating both global context and local texture/color details.
  • Dense Sampling Advantage: Dense sampling of features over the entire image was found to be more robust than keypoint sampling, especially in scenarios with smooth surfaces or less distinct keypoints.
  • Temporal Verification Crucial for False Alarm Reduction: While initial classification is important, the temporal verification step is vital for distinguishing true fires from transient, fire-like objects, significantly improving the system's reliability.
  • Dataset Size Matters: The authors demonstrate that a larger, more diverse dataset leads to more robust models and better generalization, as evidenced by their performance compared to prior work.

Key Results

The proposed system achieved an impressive 82% Recall and 93% Precision on their newly collected dataset. This performance significantly outperforms existing state-of-the-art methods, particularly in terms of recall rate, which is critical for fire detection. The paper also provides detailed comparisons with Toreyin et al.'s method [21], showing superior performance across various metrics and specific instances.

Limitations and Future Work

The authors acknowledge a few limitations:

  • Initial Combustion Detection: The system struggles to detect fires during the very initial stages of combustion when the flame size is extremely small.
  • Nighttime False Alarms: Car lights in night videos can still lead to false alarms due to their bright yellow tint and movement, which share characteristics with fire.
  • Naïve Thresholding: While improved, the reliance on pixel color thresholding can still be a limitation, especially with very bright intensities or reflections.

For future work, the authors suggest exploring more detailed fire characteristics to improve initial combustion detection and developing methods to effectively exclude fire-like objects such as car lights in night videos.

Summary updated

BibTeX
@article{jiang2015towards,
  author = {Bo Jiang and Yongyi Lu and Xiying Li and Liang Lin},
  title = {{Towards a solid solution of real-time fire and flame detection}},
  journal = {Multimedia Tools and Applications},
  year = {2015},
  doi = {10.1007/s11042-014-2106-z},
}
Figure