| Fire accident is one of the most serious threat to people’s lives and property and fire prevention is of great significance. At present, smoke and heat detectors are widely used in all kinds of places, while these contact detectors have certain limitions in detecting height. So video fire detection, with merits such as non-contact, fast response, large detection range, active detection mode and visibility, has become a new technology in security and fire safety industry. This technology is already widely applied in spacious buildings, factories and forest.However, most of the detector are flame detectors or infrared detectors, the video smoke detector, which has a much faster response is still in the research phases.There are still many problems need to be solved. So the relate methods and theories on video smoke detection are studied systematically. This thesis mainly proceeds in following aspects:(1) A smoldering smoke video library is built for the study. Two experiment platforms have been built in the standard laboratory of fire and low-pressure cabin located in the state key laboratory of fire science. Wood and cotton are use to generate smoldering smoke, and the combustion processes in different low pressures have been recorded with high definition video cameras, infrared video cameras and high speed video cameras. After arranging, standardizing and encoding, a smoke video library has been built for analysis.(2) A saliency based early smoke region segmentation method has been proposed to extract the suspected smoke region at the beginning of the detection. According to the human visual attention model, smoldering smoke region can be defined as a moving region with turbulence and grey color, so the smoke region can be segmented with the saliency based detection method based on the top-down visual attention model. Firstly, a non-linear enhancement method has been used to promote the smoke region. Then moving foreground calculated with Gaussian mixture model was used to construct a motion energy function. Combined with the function and the saliency map, the suspected smoke region can be segmented. The comparison result of several smoke detection methods show that the proposed method has better performance and the computing speed is fast enough for real-time fire detection.(3) A track-before-detect smoke detection algorithm framework is proposed. In order to promote the accuracy and robustness of the algorithm, after suspected smoke region segmented, each suspected smoke region is tracked within a time window. And the spatio-temporal features are used to identify the fire. By analyzing the physical process of smoke formation and the component of smoke, two features, the entropy and contrast of smoke texture, which can be used to describe the turbulence behavior are proposed, and the efficiency of the turbulence features is verified. The turbulence, color and motion features are used to train a smoke recognition support vector machine classifier, and then multiple classification results in a time window are used to estimate the possibility of a fire. Performance of the algorithm is tested with a group of smoke and non-smoke videos. Experimental results show that the algorithm framework is effective and reasonable. By changing the components in the framework, flame detection and other functions can also be quickly completed. Finally the effect of smoke concentration on video smoke detection is analyzed and it is pointed out that the smoke concentration is an important factor which must be considered in research and industrial application.(4) A video smoke detection system is developed. With the proposed smoldering smoke segmentation method and the track-before-detect method, a video smoke system is designed. Then the system is also tested with smoke videos under different pressures to see if the system is feasible. After region segmentation and feature extraction, features of smoldering smoke, the brightness, mean and variance of the speed, texture entropy and contrast, turbulence kinetic energy, turbulence intensity are analyzed statistically. Multiple classifiers are used to test the features under different pressures. The result shows that the features and methods we get in previous study can be used for video smoke detection under low-pressure. |
PDF Full Text Request