Feature based segmentation and clustering on forest fire video

There are some reported segmentation and clustering techniques used to analyze, browse and retrieve videos. They work on general video clips but features of special applications need to be considered to improve the accuracy. This paper focuses on segmentation and clustering of forest fire videos. Not only general frame features, but also fire features are used to segment videos and cluster scenes. Four frame features, color histogram, dominant colors, percentage and average of dominant colors, are defined to measure the similarity between frames, and key frames are extracted to index the video division. Flame color region is defined in HSV space, and the pixels with fire are tracked in each key frame to help decide the frame content as with or without fire. The segments with the same content are then clustered into one scene. Fire features are also utilized to locate the exact starting frame and ending frame with fire, which may be required in fire detection. Experiment results show that the proposed approach works efficiently on forest fire videos and is helpful for automatic fire decision.