Study On Fire Smoke Detection Method In Video Surveillance System

Fire is a kind of un-controlled burning, smoke is one of the most important image remarkable characteristics in the early fire but no open flames. Thus, smoke detection and warning in the early fire is an important way of controlling fire development and prevention.Traditional fire smoke detection systems are influenced by the external environment distortion with high false positive rate and false negative rate. Due to increasingly stringent requirements of safety and rapid development of high technology, fire detection and warning way tends toward the the direction of the visualization and intelligence. Fire smoke detection methods in video surveillance system have advantages of good real-time, scene and wide detection range, so smoke in video image sequences can be directly recognized and lindage handled in time such as alarm, fighting etc. The thesis mainly studies fire smoke detection methods in video surveillance system, and the major research has:A filter algorithm based on grads similarity-fuzzy and an enhancement algorithm based on homogeneity measurement and fuzzy entropy are proposed to improve the visual effects of the low-quality video surveillance systems. Aiming at the suspected noises in filtering window, an adaptive spatial neighborhood information function based on grads similarity is defined to eliminate pulse noises,a fuzzy membership function of the center pixel in filtering window is defined to eliminate gaussian noises combined with the weighted average method. The experiment results show it can preserve details and filtering. The improved local fuzzy fractal dimension is introduced in the traditional homogeneity measurement method, a criterion based on fuzzy entropy measure is established, finally the enhancement algorithm is improved. The experiment results show it is an efficiencial image enhancement for details and contrast.A pixel intensity classification algorithm based on random strategy of background reconstruction is proposed to overcome the light interference. Background sample points are selectively collected and updated with spatial local correlation and temporal continuity of each pixel. The reconstruction has strong noise resistance, good image quality and better adaptability for light gradient, light sudden change and non-uniform illumination etc. The experiment results show the algorithm has better accuracy of moving target detection.Aiming at suspected smoke interference in traditional algorithms, an algorithm of smoke detection based on dark-channel prior is proposed. The local regional blocks are chosen by adaptive partition method, the fine and optimized transmittances are quickly obtained by guided filtering. Base on the video smoke imaging model, the optimized and smooth dark-channel image model is builded using atmospheric optical and fine transmittances. Above this, in the reconstructed background image and the current smoke image, suspected smoke regions can be captured. Then by frame difference and the following treatment, the accurate smoke targets can be detected. The experiment results show the algorithm largely improves the accuracy rate.An algorithm of smoke recognition based on wavelet fractal-fractal neighbour distance is proposed.Modifying fractal compression coding based on entroy and multi-scale matching, the local unit entroy feature is extracted, the coarse scale images are builded, the matchings procedures in range blocks on coarse scale and fine scale are carried on with the pre-rejection of the mismatched domain blocks, so the algorithm has low complexity and high matching efficiency, low distortion error. The database include smoke and non-smoke is established, the wavelet fractal feature is extracted, the similarity distance between images based on wavelet fractal-fractal neighbour distance is calculated. The experiment results show the algorithm largely improves matching rate.