Fire Detection for Early Fire Alarm Based onOptical Flow Video Processing
Suchet Rinsurongkawong1, Mongkol Ekpanyapong, and Matthew N. Dailey
Mechatronics, [email protected]
Microelectronics and Embedded systems, [email protected]
Computer Science and Information Management, [email protected]
Asian Institute of Technology, Pathumthani, Thailand
Outline
• Introduction• Methods• Experience result• Future work
Introduction
• Fire has always threatened properties and peoples’ lives.
• Most conventional fire detection technologies are based on particle sampling, temperature sampling, and smoke analysis,but fire detection systems using these technologies have limited effectiveness due to high false alarm rates.
• Because of the rapid developments in digital camera technology and computer vision system, there are many fire detection technologies which are introduced based on image processing.
Moving region detection
• Background subtraction:
• Be assumed to be a moving pixel if:
Chromatic features(1/3)
• The color of fire always appears in red-yellow range.
Chromatic features(2/3)
• To solve from a fire-like color.
Chromatic features(3/3)
• Besides, when the fire is in dark background environment without other background illumination, the fire will be the main light source. From this reason, the fire may display in a whole white color in an image. Thus, the intensity should be over threshold intensity IT .
Growth rate analysis
• The growth rate rule can be deduced as:
• Where Gi denotes quantities of the current frame to the n th frame.
• If the result is more than a reference Gr from the first detected frame, the moving object will be considered as a real flame.
Turbulent fire plumes
Turbulent fire plumes
• The frequency shows the cycle times of eddies effect per 1 second.
• Where f denotes a vortex shedding frequency in Hz for a fire of diameter D in meters.
Lucas-kanade optical flow pyramid
• The algorithm of LK is based on 3 assumptions.
1. “Brightness constancy”
2. “Temporal persistence”
3. “Spatial coherence”
Flow rate analysis(1/3)
• From the previous step, the LK optical flow can extract the motion velocity vector from each feature point.
• Where p and q denote the starting and the ending point of each feature point respectively. n refers to the number of feature points.
Flow rate analysis(2/3)
• The average flow rate of the first time of optical flow analysis is calculated as follow:
• Where Fa denotes the average flow rate of the first detected time for optical flow analysis. This first average flow rate will be used as a reference value for next n time calculation.
Flow rate analysis(3/3)
• variation of flow rate:
• Where Fv is the average flow rate from n time calculation,we will called it “variation of flow rate”. Due to the turbulent of flame, the variation flow rate of fire will give a significant value more than other moving objects.
Expermental result
• Find the flow rate threshold value
Method1 & method2
Result from method1
Conclusion and future
• In dynamic analysis, the combination of growth rate and Lucas-Kanade optical flow can extract the motion feature of fire, so this method can easily distinguish the disturbances which having the same color distribution as fire.
• In the future, the neural network will be applied to train the raising parameters composed of fire-pixels extracted at timeinterval fur increasing the reliability of fire-alarming. The use of neural networks, the statistical values must have highly enough in the training process.
Thanks for your attention!