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

Abstract: Although the object detection and recognition has received growing attention for decades, a robust fire and flame detection method is rarely explored. This paper presents an empirical study, towards a general and solid approach to fast detect fire and flame in videos, with the applications in video surveillance and event retrieval. Our system consists of three cascaded steps: (1) candidate regions proposing by a background model, (2) fire region classifying with color-texture features and a dictionary of visual words, and (3) temporal verifying. The experimental evaluation and analysis are done for each step. We believe that it is a useful service to both academic research and real-world application. In addition, we release the software of the proposed system with the source code, as well as a public benchmark and data set, including 64 video clips covered both indoor and outdoor scenes under different conditions. We achieve an 82% Recall with 93% Precision on the data set, and greatly improve the performance by state-of-the-arts methods.

• The paper presents a real-time fire detection system combining multi-level thresholding, LAB color, and SURF texture features to achieve 93% precision and 82% recall.
• It employs a cascaded framework including candidate proposals, SVM-based region classification with an RBF kernel, and temporal verification to significantly reduce false positives.
• Empirical studies with a newly compiled dataset demonstrate the system’s adaptability to various lighting and background conditions, enhancing video surveillance efficiency.

Towards a Solid Solution of Real-Time Fire and Flame Detection

Introduction

The paper "Towards a Solid Solution of Real-Time Fire and Flame Detection" focuses on developing a robust, real-time method for detecting fire and flames in videos. This system is designed to be integrated into video surveillance and event retrieval applications. The necessity for such a technology is underlined by numerous fire catastrophes, as mentioned in the paper, and the need for early detection systems. The approach taken by this paper is significant as it leverages vision-based methods, using RGB cameras, which provides various advantages over sensor-based methods, including cost-effectiveness and the ability to cover broader areas in real time.

Figure 1: Various appearances of fire.

System Framework

The proposed fire detection system is structured in three cascaded steps: candidate region proposals, fire region classification, and temporal verification. The system's architecture is key to processing video efficiently and in real time.

Figure 2: Framework of proposed fire detection system.

1. Candidate Region Proposals: This step involves proposing potential fire regions by modeling the background and using a multi-level thresholding approach. This method dynamically adjusts thresholds based on background intensity statistics, addressing challenges posed by varying lighting conditions and reflective surfaces.
2. Region Classification: The classification process utilizes a combination of color-texture features, including a CIE LAB color descriptor and SURF texture descriptors, with an SVM classifier to detect fire regions. Importantly, the RBF SVM kernel is identified as the optimal choice, addressing issues with high false alarms associated with purely color-based detection methods.
3. Temporal Verification: The system further reduces false positives through temporal verification, employing statistical measures like perimeter and area stability over time to differentiate transient light sources from actual fires.

Empirical Study and Feature Selection

The system's feature selection is based on extensive empirical studies using a newly compiled dataset. The researchers tested different color spaces and found LAB histograms to outperform others when using an RBF SVM kernel.

Figure 3: Precision-recall curves of different features with 3 kinds of SVM kernels.

Color and Texture Features

• Global Color Feature: Among RGB, YUV, HSV, and LAB spaces, LAB provided the best precision-recall performance, particularly when paired with the RBF SVM kernel.
• Local Color-Texture Feature: The study highlights the efficiency of combining SURF descriptors with a local color histogram, noting that dense sampling is superior to keypoint-based approaches for robust region classification.

  Figure 4: The left image illustrates keypoint sampling and the right image illustrates dense sampling. Blue dot is the location of sampling point and the diameter of cyan circle represents the size of SURF kernel.

These features significantly improved the robustness of fire detection, allowing the system to operate effectively in diverse visual contexts.

System Implementation

The implementation leverages a hierarchical structure of feature extraction and classification to enable real-time processing. The proposed system's capability to adapt to different environmental settings through dynamic threshold adjustments makes it versatile for various surveillance contexts.

Figure 5: Two examples of candidate fire region proposing. The first row shows a surveillance video example with still background, the background modeling result is shown in (a), and the mask of candidate regions is shown in (c). The second row shows a video with moving background. The candidate fire region mask shown in (e) is found by a multi-level threshold.

Evaluation and Results

Extensive testing with a dataset comprising 64 video clips demonstrated the system's high precision (93%) and recall (82%) rates, outperforming existing methods by improving recall significantly. The evaluation indicates that while the system is highly effective, early-stage fire detection, particularly for small flames, remains a challenge due to subtle visual characteristics.

Conclusion

This research presents a significant advancement in fire detection systems, emphasizing empirical evaluation and optimal feature utilization to develop a reliable, real-time detection framework. The study contributes valuable datasets and software, encouraging further improvements and adaptations in real-world applications. Future work should explore enhancing initial combustion detection sensitivity and refining the accuracy of distinguishing between fire and fire-like objects in challenging scenarios.