【作者】 闫建强；

【导师】 马莉；

【作者基本信息】 杭州电子科技大学 ， 模式识别与智能系统， 2011， 硕士

【摘要】 研究开发基于视频图像的嵌入式分析设备对林区火灾的早期预报、减少森林资源损失具有重要意义。以DSP为核心的嵌入式系统具有体积小、功能完备的优点,同时结合机器视觉领域的图像处理技术,使其成为基于视频图像的森林火灾监控系统的可行方案。本文在浙江省重大科技专项——基于太阳能的森林火灾监测技术与系统研制项目的支持下,在TI公司的DSP芯片TMS320DM642的基础上设计和开发了林火烟雾检测系统,以DM642为核心搭建系统硬件平台,基于实时可裁剪的DSP/BIOS内核设计了RF5系统软件框架,完成了系统多线程、多任务的实时调度。烟雾是林区火灾发生早期的显著特征,对烟雾的及时检测是提高火灾预警能力的一个有效途径。根据开放环境下火灾烟雾检测的应用需求,提出了一种基于多特征融合以及神经网络/支持向量机判定的在线烟雾检测算法。该算法首先提出了适应开放环境且抗光照影响较好的烟雾颜色模型,并与Kalman运动检测图像进行融合得到疑似烟雾区域;接着利用小波变换等方法对疑似区域提取烟雾的轮廓不规则性、面积扩散性、烟雾区域模糊性以及烟雾边缘的低频振荡性四个时空域特征,作为离线训练完成的BP神经网络/支持向量机SVM分类器的四个输入;最后根据分类器的输出来判定所监测的林区是否发生火情。本系统通过现场调试和实验表明,系统稳定可靠,有较高的实效性,系统设计和资源分配正确合理;本文烟雾检测算法能够实时进行视频烟雾检测,对开放环境较为鲁棒,能较好的满足开放环境下火灾烟雾的监测预警,基本达到了预期效果。更多还原

【Abstract】 It is of great significance to take a research and develop a system of video image analysis for early warning of forest fires and reduce the loss of forest resources based on embedded technology. Because of its small volume and fully functional, the embedded system taking DSP as the core are adopted .Combining with the technology of image processing in machine vision field, it’s also becoming a good choice for the forest fire monitoring system. The paper developed the fire smoke detection system based on TMS320DM642 made by TI, took DM642 as the core kernel, and planned the RF5 software framework by DSP/BIOS to complete the multitasking real-time scheduling by the support of the projects of major science and technology in Zhejiang -- solar energy-based forest fire monitoring technology and system development.Smoke is a remarkable characteristic for early forest fire, the timely detection of smoke is an effective approach to improve the ability of fire warning. According to actual application requirements of smoke detection in the open environment, an on- line algorithm based on multi-feature integration and neural network or Support Vector Machine (SVM) was proposed. Firstly, a smoke-color model adapted to open environment and resistance to light was presented, then, got the suspected smoke area with Kalman motion detection by image fusion. Secondly, four spatial characteristics of smoke were extracted as follows: contour irregularities, area of avian, ambiguity of smoke region, low frequency oscillation of smoke use wavelet and so on, as the inputs of BP neural network and support vector machine trained completed offline. Finally, the paper used the classifier’s outputs to determine whether there was occurring fires.The debugging and experiments’results show that the system is reliable and higher effective, the design of system and resource allocations are correctly reasonable. The algorithm of paper can realize real-time video smoke detection under certain conditions, and adapts to the open environment, meets the requirements of fire smoke warning in open environment and desired effects.更多还原