【作者】 李强；

【导师】 王赓；

【作者基本信息】 上海交通大学 ， 软件工程， 2016， 硕士

【摘要】 目标追踪一直是计算机视觉领域的一个重要课题,在视频监控、航空探测、人机交互等领域有着广泛的应用前景。尽管不断有新的目标追踪算法被提出,但是由于目标运动过程中会受到光照、姿态和尺度变化以及遮挡等一系列不确定因素的干扰,因此研究设计实现高效鲁棒的追踪算法仍然是一个具有重要实用价值的技术难题。本文针对复杂场景下的目标追踪进行研究,重点分析了TLD^[1]（Tracking Learning Detection）和压缩感知^[2]两种追踪算法,结合两者的优势,设计实现了一个基于TLD框架和压缩感知理论的高效目标追踪方案。主要研究成果如下:（1）改进中值流追踪算法,以目标区域中的FAST特征点替代均匀分布的特征点,用邻域一致性约束过滤追踪失败的特征点,提高追踪器的稳定性。（2）利用Kalman滤波对目标进行运动估计,结合目标窗口的置信概率,自适应确定目标的搜索范围,减小检测器的搜索区域,在保证追踪准确性的基础上提高检测器的效率。（3）将压缩感知检测应用于TLD框架,基于PN（Positive and Negative）学习^[3]策略和追踪器尺度信息对压缩感知追踪进行改进。用改进后的压缩感知算法结合最邻近分类器对目标进行检测,提高检测器的准确性。改进的追踪算法结合了TLD和压缩感知方法的优势,在鲁棒性和准确性上均有所提升。实验结果表明,相比原有的目标追踪方法,改进后算法在兼顾效率的基础上能够更加准确稳定地追踪目标,满足实时追踪的要求。更多还原

【Abstract】 Object tracking has always been an important subject in computer vision,which can be applied to many domains,such as video surveillance,aeronautical detection,human-computer interaction and so on.Although there have been new algorithms proposed,it is still a challenging and practical task to develop effective and efficient tracking algorithm due to the undetermined factors such as illumination variation,pose change,scale variation and occlusion during tracking.Aiming the object tracking under complicated environment,this paper places emphasis on TLD^[1]（Tracking Learning Detection）and compressive tracking^[2].By combining the advantages of both methods,an effective tracking algorithm is proposed based on TLD framework and compressive sensing theory.The main research achievements include following aspects:（1）Modify the median-flow tracking algorithm by replacing uniformly distributed interest points with FAST feature points and using neighborhood consistency constraint to filter tracking failures in order to improve the performance of tracker.（2）Estimate the object motion using Kalman filter and determine the search region adaptively with the confidence probability of the object window to reduce the search area of the detector for improving detection efficiency on the premise of tracking accuracy.（3）Employ compressive sensing in TLD framework,and improve compressive tracking based on PN（Positive and Negative）learning^[3]and scale information from tracker.Moreover,use improved compressive sensing combined with nearest neighbor classifier to improve the detection stability.By combining advantages of the TLD and compressive tracking,the modified tracking algorithm improves the tracking stability and accuracy to some extent.The experiments show that,compared to original tracking methods,the improved algorithm can locate the object more accurately and stably,taking efficiency into consideration.更多还原