Research On Smoke Detection Based On Wavelet Analysis In Open Environment

Online video monitoring of forest smoke in an open environment is of greatsignificance for protection of forest resources and sustainability of social production.The paper implemented an intelligent embedded online monitoring system, which isbased on platform TI DM642, for smoke detection in an open environment by thesupport of the projects of major science and technology in Zhejiang province--solarenergy-based forest fire monitoring technology and system development. The mainwork and contributions in this paper can be described as following:(1) Build implementation of software and hardware platforms of a videosurveillance system.To implement video capture, video playback in the local and teletransmission ofon-site images, a hardware platform framework of the smoke algorithm platform wasbuilt on the hardware framework. Here, regarding DM642 as the main module of thehardware platform. On the other hand, for the purpose of implementing smokealgorithm and communication with other modules on the software framework,building multi-task scheduling, inter-thread communication and interrupt processingby making use of DSP / BIOS to implement the software equipped with smokealgorithm and complete the task of information exchange between video camera andthe remote host.(2) Designment of an on-line smoke detection algorithm fitted openenvironment.In this paper, presenting an intelligent on-line algorithm for smoke detection,which is suitable for an open environment. The algorithm is modeling based on thestatistical modeling to extract the suspicious region by using combination of color andKALMAN models and adaptive GMM model, then the contour and regionalcharacteristics of smoke are extracted in the time domain based on thewavelet-expression. After that, multi-frame characteristics of smoke in the timedomain were regarded as the inputs of BP neural network judged as smoke andpseudo-smoke. Finally, the paper used the classifier’s outputs to determine whetherthere was occurring fires. In this paper, smoke model, built according to the analysisof a large number of forest datum in open environment, can implement the adaptationof parameters consistent with all-weather light variation to extract good smoke suspicious areas. Take consideration of dynamic characteristics, the wavelet-basedexpression of contour and regional characteristics of smoke was presented for itsadvantage of reflecting smoke’s variation characteristics so as to distinguishnon-smoke and smoke. In order to reflect the variation of smoke’s dynamiccharacteristics in the time domain better, a neural network determine mechanismbased on pattern was presented here. According this classification, the input smokefeatures are vectors composed of multi-frame and only a judgment is made to reducethe false detection of smoke effectively.(3) Completion of the overall debugging of the system in an open environment,This issue is processed by the support of the projects of major science andtechnology in Zhejiang province. The overall platform of smoke algorithm systemcomposed of hardware and software had been completed since 2008. Moreover,selection of dynamic characteristics in time domain and transplant of smoke detectionalgorithm had been implemented. At last, the overall test of system was processedboth in the school forest and Shaoxing forested area, which is stable and able todistinguish smoke and pseudo-objects correctly. The project was inspectedsuccessfully on May 30, 2011.