Study On Image Restoration And Human Detection In The Building During Fire Initial Stage

Based on the CCTV system, the building manager can get the real-time information in the building in the form of images. This provides a great convenience for the daily maintenance of the building. Thus the CCTV system has become an integral part of current buildings. However, when a fire breaks out in the building, due to the contamination of the smoke, the images that are taken photo by the CCTV system become fuzzy. The fire fighters cannot obtain the information in the fire scenario from the CCTV system. The personnel evacuation and the personnel rescue become much harder. Therefore, it is worth developing an image sharpening technology to restore the ability of the CCTV system.It is the key moment to guide the personnel evacuation and rescue the trapped occupants during the fire initial stage. And when fire is at the initial stage, the smoke in the fire scenario is in low concentration, this will be convenient for us to research the image restoration method. In this paper, we study the image restoration method at the building fire initial stage from two ways:light design and software development. After the human image is clear, combining the human detection algorithm to find the trapped people in the fire scenario in a very short time. The specific works are as follows:A model of light transmission in the smoke layer is established, and the effects of the light transmitted flux, the incident light wavelength, the smoke particle concentration, and the smoke particle size are analyzed. Considering the light scattering effect of the smoke particles, the radiative transfer equation (RTE) that includes the multiple scattering is applied to calculate the light transmitted flux. The n-heptane smoke particle is used as the research object in this work, and the single scattering albedo, the asymmetry parameter as well as the extinction cross section of the single smoke agglomerate were calculated which are the input parameters of the RTE. The discrete ordinates method (DOM) was used to solve the RTE. The results show that the light transmitted flux decreases with the growing smoke size and smoke layer particle concentration, and increases with the growing light wavelength. The experiments with respect to the incident light with different wavelength transmitting in the smoke layer are carried out, and the results are consistent with the light transmission model predictions. A fast smoke removal algorithm is developed to refine the people images contaminated by the smoke. Based on the Retinex theory, the brightness of the image is determined by the illumination of the environment and the surface reflection of the object. The smoke in the fire scenario only changes the characteristic of the illumination, while the surface reflection of the object is not varied. Thus, the process of the smoke removal is the process to solve the surface reflection of the object. Considering the characteristic of the smoke, this paper improved and compared the different surround functions in the Retinex algorithm, and last the Gaussian Pyramid filter algorithm and the pixel logarithm table algorithm are combined to develop the fast and effective smoke removal algorithm which named GL-Retinex. Based on GL-Retinex algorithm, the operation time to deal with an image of704x576reduces from5150ms to198ms, which is nearly26times. The GL-Retinex algorithm is applied to deal with the smoky images in different light conditions (incandescent lamp, emergency light and infrared light), and the results show that the GL-Retinex algorithm in this paper can effectively remove the contamination of the smoke. When in the emergency light environment, the largest concentration of smoke which we can find the human in the smoky images after disposed by GL-Retinex algorithm is up to4.81dB/m.A dataset regarding the physical shapes of the occupants in particular fire scenario is built, and a personnel classifier is trained to detect the occupant in the monitoring images.6normal shapes were taken into account in the dataset, which are upright run, stoop run, single hand cover the nose upright run, single hand cover the nose stoop run, both hands cover the nose upright run and both hands cover the nose stoop run. This paper collected4767positive images through experiment, online and INRIA person dataset. Considering the assembly occupancies which include the corridors, stairs, chairs, etc.,2000negative images were collected. Extracting the histogram of gradient (HOG) features of all images in the person dataset and applying the Adaboost cascade classifier to train the HOG features, the personnel classifier which is suitable in fires can be obtained.300smoky images (200positive images and100negative images) were applied to test the performance of the personnel classifier. Before testing, the GL-Retinex algorithm is applied to refine the test images and the histogram equalization method is applied to increase the contrast of the test images. In this way, the detection accuracy of the personnel classifier will be increased. The test result shows that the detection accuracy of the positive images are more than54%in all illumination conditions, and the detection error rate of the negative samples are all lower than8%. Although these values are small because of the influence of smoke, they can at least meet the requirement of practical engineering.Based on the image human detection method (static method), a dynamic human detection algorithm is developed. The static human detection method needs to deal with the whole image to search for the people in the image; therefore the operation time is much longer. However, in fire scenario, the people are moving while the building is still. Based on this idea, this paper extract the moving area of the image through the operation of frame difference, morphological opening operation, cutting the small connected domain, combing the adjacent connected domain, etc. The human detection is only operated in the moving area. Thus, the operation time of the dynamic human detection method is reduced from550ms to250ms, which basically meets the real-time requirement of CCTV system. At the same time, as the detection area is reduced, the detection error rate of the negative images is also reduced to below6%.