Fire Detection in Color Images Using Markov Random Fields - Advanced Concepts for Intelligent Vision Systems - page 92

Fig. 2. Source frame, detected blocks with high energy threshold, and corresponding
component with lower energy threshold

This is exploited in a two-stage classifier. The energy function is calculated on
4x4 pixel blocks and first thresholded on an energy level T 1 allowing the entire
fire areas to pass the criterion, as well as some spurious detections in the non-
fire areas. In the second stage, the calculated energy is thresholded on a level
T 2 <T 1 , allowing only series of blocks near the edges of the flames. Only the
first-stage connected components which contain blocks from the second stage
are retained, resulting in much fewer false detections. An example can be seen
in figure 2.

4P orman e

The color occurrence probability distributions P f and P b weretrainedonaset
of 6 ground truth images of fire, and 6 additional background images featuring
a variety of settings. The fire images are video frames depicting four different
fires, captured by different types of cameras and from different viewing angles.
The images also exhibit a wide variety of camera settings, from underexposure
to oversaturation and varying degrees of focal sharpness. The additional back-
ground images were included for training balance, as the video frames were
all captured in an industrial environment and therefore featured similar back-
grounds. This training is intended to be universal, so no retraining is required
for use in different circumstances. However, results may improve further for very
specific scenarios when the method is trained on the according scenario-specific
imagery.
The performance of the fire detection system was evaluated on over 49,000
video frames and compared to the fire detection method proposed by Celik
et al. [14], which defines a set of rules in Cr-Cb color space based on three

Table 1. Performance statistics of Celik et al. compared to the proposed method

Method Detection rate False alarm rate

Celik 2008 99.95% 50.93%

Proposed (first stage only) 99.98% 42.82%

Proposed (first and second stage) 99.57% 21.80%