Research And Development Of Real-time Fire Detection System Based On Embedded Vision

Image based fire detectors are new horizons of the fire detectors family. Theyintegrated technologies such as intelligent control, image processing, pattern recognitionand embedded system. They have benefits than traditional ones such as shorter alarm time,ignoring the indoor space size, supporting the outdoor environment accurate positioningand etc.Currently Image based fire detectors have not been widely used. The main reason isthat, the working environments are complex, and the immature fire detection technology isstill in experimental stage. In addition, the traditional detectors have mature markets, andbecome strong competitions. So it is fewer cases in which new detectors are provided toengineering applications. The complete solution of image-base fire detectors is provided inthis paper. The main works are listed below:(1) Considering the effect and price, we chose the method of camera with NIR filter tocapture infrared image, and provide a selection solution of camera and filter.(2) We analyze the disadvantages of the global luminosity algorithm. After doing studyon local geometry features, we research the detection algorithm based on features fusion.(3) We write programs for the programmable intelligent industrial cameras, transplantprograms from computers to embedded devices. Final the detectors can work on the alarmnetwork independently without computers.(4) We optimize the algorithm in the way of parallel computing. For example, we useOpenMP/CUDA to improve the speed of moment invariants algorithm by1.3-19times onparallel CPU/GPU.(5) According to a mount of experiment results, we prove the correctness and stabilityof the algorithm and measure the actual effect of optimized parallel computingThere are several innovations this paper.(1) We provide a fire detection algorithm based on local geometric features, It includesheat area extraction,5local geometric features of the heat area, heat area classificationbased on minimum risk Bayes decision and multiple features fusion, multiple sampling toreduce error rates.(2) We try to combine the algorithm based on local geometric features and algorithmbased on global luminosity features in paper [1], and enhance the speed and effectiveness.As the experimental result, our algorithm is better than algorithm in paper [1]. (3) We use OpenMP and CUDA to optimize the algorithm in the way of parallelcomputing on computer, provide complete implementation details and optimization results...