The Research On Early Forest-Fire Smoke Detection Method Based On Nonparametric Feature Extraction

The traditional forest-fire monitoring mainly relys on the observatory or the unmanned aerial vehicle(UAV).When monitoring a large area of the forest, those traditional methods not only cost too much but also are difficult to assure the monitor quality. As a result, it often delays the time and causes huge financial losses and ecological damage. The smoke detection technology basing on the video breaks through the traditional method of forest-fire monitoring, it will be more accurate and earlier to detect the smoke with less manpower and resources. In this paper, the problem of recognition accuracy and false alarm rate in current forest-fire smoke detection method are studied.Firstly,this thesis introduces the research background and status of the forest-fire monitoring and simply describes the mainly facing problem in forest-fire smoke detecting. Basing on the research of smoke’s dynamic characteristics and static characteristics, several typical classic moving target detection algorithms and feature extraction algorithms are summarized and analyzed in this paper.Secondly, in order to detect the slow-moving smoke from many kinds of jamming targets in the wild, the double-background model is improved. This model uses the existing motion area detection algorithm’s defect which often classifies the smoke target as the background image. This improvement is inspired by inter-frame difference method and background difference method, which improves the detection capabilities about slow moving targets.Finally, there are still some problems in current smoke detection algorithms. For example, in most cases, the eigenvector’s selection and configuration are still relying on the human experience, which will not guarantee the detection results in optimum. In addition, traditional non-parametric feature extraction methods are mostly basing on the Euclidean Distance. When the Euclidean Distance of samples is equal, there will be some bound about weight assignments. In order to solve these problems, we propose an early forest-fire smoke detection method basing on the non-parametric feature extraction. This algorithm uses the cosine distance instead of Euclidean Distance to measure the similarity of two samples and introduces the kernel function. Simulation results demonstrate that the proposed algorithm can identify the smoke target from a variety of interference objects accurately and have a better recognition performance.