Flame Detection Algorithm In Video Image And Its Optimized Implementation Based On DSP

The fire is one kind of common natural disasters, it serious harmed people's life and property. Traditional fire detectors are mostly based on heat detect, light detect or smoke detect, they may easily be affected by height, square, air movements, dust or temperature of the surveillance area. Fire detector based on images can surmount these difficulties; it will work better in large spaces, and has broad application prospects. Flame detection is an important aspect of fire detection. The main research object of this paper is flame detection algorithm in large spaces with complicated background and its optimized implementation based on embedded DSP system. The key of research includes three respects:segmentation of flame in video image; flame feature extractions and recognition; algorithm migration and optimization based on DSP system.Traditional flame segmentation algorithms may easily be affected by complex environment and some of them are too complex. After researching on motion property of flame and considering of practical application environment, a motion detect algorithm is proposed to segment the flame region. The algorithm is based on background subtraction and the background model of every pixel contains multiple uniform distributions. The current value of a pixel will be matched with its background model to decide the pixel is moved or not. The algorithm can overcome the interference of complex background and it's easier and more suitable for embedded system when compared with motion detection algorithm based on Gaussian mixture model.The study on flame features includes two aspects:static features and dynamic features. Four flame features such as color, area changes, flickering and region movement are selected to discriminate the real fire regions from the foreground regions of flame segmentation. The color judgment of flame can be more suitable for different flame when we judge the rate of flame color pixel count and the area of connected regions. The slope of fitting line is used to represent the increasing of flame area; the position changes of top and bottom points of fire region are used to represent the flickering feature of flame, with these methods descriptions of flame features are more intuitive; feature extractions of flame are easier, so the system is more effective to detect a flame.Finally, the flame detect algorithm will be migrated to DM6437 DSP system. Based on basic optimization method and arithmetical features of DM6437 DSP, the algorithm will be optimized from two aspects:memory optimization and code optimization. Memory optimization includes memory allocation and optimization with EDMA and Cache, while code optimization includes C code optimization and linear assembly language. The test result shows that the new algorithm and its optimization method can meet the requirements of DSP system and achieve real-time flame detection in video image.