【作者】 刘锐；

【导师】 孙涵；

【作者基本信息】 南京航空航天大学 ， 计算机应用技术， 2009， 硕士

【摘要】 移动目标检测与跟踪是计算机视觉学科研究的一个重点课题。在视频监控、安防布控、武器装备等方面都有广泛的应用。经过多年来国内外研究者的共同努力,移动目标检测与跟踪技术已经得到长足的发展,取得了很多突破性的研究成果。但在实际复杂的现场环境中,光照变化、目标遮挡、阴影干扰、形变影响等的存在,给该技术提出了新的挑战。要很好的解决这些问题,并达到实际应用的要求,需要设计实时性和鲁棒性兼顾的新算法。本文深入学习若干核心算法,并在此基础上做了新的尝试,本文主要工作如下:1.深入学习了国内外移动目标检测与跟踪的经典方法,并通过实验比较,了解各自优缺及适合的应用条件。在此基础上,提出了一种基于多特征目标建模的Mean-Shift及Particle Filter相结合的目标跟踪算法。目标建模融合了颜色、边缘、纹理等一般目标特征,另外还借鉴图像质量评价研究领域中的结构相似度(SSIM)概念,提取目标的结构信息作为第四个融合子特征。改进了多特征融合方法,得到鲁棒性较高的目标模型。在改进的MSPF(Mean-Shift & Particle Filter)跟踪框架下最终实现对移动目标的跟踪。通过实验验证该算法跟踪准确。2.提出了一种基于行扫描的目标提取算法,该算法对图像进行逐行扫描,直接提取目标,不需要对全图进行链码跟踪。该算法在嵌入式DSP开发中结合DMA双缓冲技术,可以大大节省算法时间开销。3.将多特征融合的MSPF目标跟踪算法移植到ADSP BF561开发平台上,解决移植中的图像格式转化等若干问题。结合基于行扫描的目标提取算法,优化目标跟踪算法在该平台上的实现。在移植基础上做C语言级和汇编级代码优化,进一步提升算法执行效率,最终满足目标跟踪应用的实时性要求。更多还原

【Abstract】 Moving target detection and tracking is one of the most important subjects in computer vision. It has important and practical value together with wide developmental prosperity in video surveillance, security defense, and military hardware and so on. After years of joint efforts made by domestic and foreign researchers, this technology has been developed rapidly and achieved a lot of significant results. However, there are new challenges existing in practice complex environment, illumination changes, target occlusion, shadow interference, the impact of the presence of deformation and so on. To solve these problems and compensate the practical need of the requirements, it is necessary to design new algorithm of real-time performance and high robustness. This paper deeply studied many core algorithms and did new trial based on those. The main works in this paper are as follows:1. By deeply investigating many classical methods of moving target detection and tracking, this paper points out their advantages, disadvantages and the appropriate application conditions of them. This paper proposes an algorithm based on multi-featured modeling, Mean Shift and Particle Filter combining. Object modeling combines the color, edge, texture and other general features. Using the concept of the structural similarity (SSIM) in the field of image quality evaluation for reference, this paper extracts the structural information of the target as the fourth sub-feature. Also,the paper improves the multi-feature fusion method to get a higher robustness target model. Finally this paper track objects in the improved MSPF (Mean Shift & Particle Filter) tracking framework and test the algorithm by numerous experiments.2. The paper presents an object extraction algorithm based on line scanning. This algorithm scans the image line by line to extract object, other than searching object in whole area of images. In embedded DSP development, this algorithm is combined with DMA and ping-pong buffer, which could greatly improve the speed of the operation.3. This paper transplants the multi-feature fusion MSPF tracking algorithm to ADSP BF561 development platform. Many problems in migration are Solve including image format conversion and so on. The line-scanning object extraction algorithm is used to optimize this tracking algorithm in the implementation on this platform. C-level and assembly-level code optimization are also included in to further improve the efficiency to meet the real-time requirement of the target tracking application.更多还原