Hierarchical detection of wildfire flame video from pixel level to semantic level

The importance of flame detection cannot be ignored in a wildfire video surveillance system due to disturbance of heavy fog and challenging of smoke detection. In this paper a novel method for hierarchical detection of wildfire flame video is presented. Specifically, wildfire flame images are gradually recognized from low level visual features of pixel based to high level semantics of video clip based. For all the pixels of one image, the pixels which meet color rules and motion characteristics are labeled as flame colored pixels. The candidate flame region roughly generated by flame-like pixels is divided into non-overlapped image blocks. The sparse representation of the blocks are defined and recognized by learned dictionaries to more accurately segment candidate flame region and exclude some non-flame regions. To reduce the cost of computation, the proposed method detects one Frate (Frate denotes one frame rate) frames instead of one frame at a time by using a sliding time window. Flicker features and spatiotemporal features extracted from video clips of the size Frate are used to build semantic model of wildfire flame video recognition based on mathematical model of meaning. Experimental results show that the proposed approach can effectively segment flame region and significantly improve the performance of wildfire flame detection.