【作者】 杨治国；

【导师】 王太宏；

【作者基本信息】 湖南大学 ， 信息与通信工程， 2015， 硕士

【摘要】 机器视觉伴随着硬件性能的快速提升得到了快速的发展,而行人检测与跟踪技术是机器视觉领域研究的热点方向,同时也是最基本的任务。它在安防、交通安全和人机交互等领域有着非常重要的研究价值和广阔的市场前景。虽然近十年来,大量研究人员致力于行人检测与跟踪算法的研究,但是由于行人的固有特性(肤色、衣着各异、非刚性),使得目前行人检测与跟踪的效果远远没有达到实际应用的标准,因而该方向仍然是机器视觉领域内的研究热点。本文主要针对目前行人检测实时性差的问题,利用视频监控中行人的运动信息这一先验知识并对行人的位置进行预测,提出了一种快速的行人检测方法。同时考虑到行人是一种非刚性目标,跟踪过程中,行人的外观模型容易发生较大的变化。为了适应行人的外观变化,为CMT算法中采用的模型提出了一种模型更新方法,使其更适合于长时间跟踪。本文的主要工作与创新点阐述如下:实时性差是限制行人检测在应用中推广的最主要的因素,本文针对该问题,提出了一种在监控视频下如何快速提升行人检测速度的算法。该算法利用场景中行人是运动目标这一特性,提出了率先分割前景目标的方式,减少了待筛选区域。同时通过设计良好的预测方式,对确定为行人的区域进行预测,得到行人在新一帧中的位置,这种预测方式可以极大地减少计算量,是实现实时行人检测关键的一步。为了保证预测结果的可靠性,本文利用BRISK关键点特征为场景中每一个行人创建了临时模型,最后对预测区域利用临时模型进行校验。并通过实验从精度上和时间上验证了本文提出的算法的性能。基于部件的模型更具应对目标变形、遮挡的能力。由于行人非刚性的特点以及实际场景中容易发生遮挡,因此,利用基于部件的模型可使行人跟踪更加鲁棒。但是基于关键点模型的CMT算法,仅仅利用第一帧的信息作为目标模型,使得其在长时间跟踪中性能表现不佳。为了解决CMT算法存在的这些问题,在本文中,首次提出了一种更新关键点模型的方法,同时为了配合这种动态模型,利用累积求和求积的方式对目标的旋转角度与缩放尺度进行估计。为了确保最后结果的可靠性,引入凝聚型分层聚类算法对最后与模型匹配的关键点进行聚类,从而排除因目标形变,背景所带来的异常关键点的影响。最后,在50个视频序列上的测试结果表明,改进后的算法(AMT算法)比原CMT算法具有更加优异的性能。更多还原

【Abstract】 With the rapid improvement of hardware performance,computer vision is developing fast.Pedestrian detection and tracking technology is the hotspot of computer vision and also the most fundamental application of computer vision.It plays an important role in surveillance,traffic security and human-computer interaction,so it is of great research value and has a bright market prospects.Although lot of researchers have been dedicated to improve the algorithm of pedestrian detection and tracking for decades,due to the inherent characteristics of pedestrians(different color and dress,non-rigid),the performance of pedestrian detection and tracking is far from practical application,so it’s still a hotspot in computer vision.This paper is mainly focus on improving the real-time performance by taking advantage of the information of pedestrian movements to predict their positions.In consideration of that pedestrian are non-rigid objects,the models of pedestrian change a lot.We propose a method to update models in CMT algorithm to adapt to the change of pedestrian appearance,making it more suitable for long-term tracking.Our main work and innovation are elaborated in this paper as follows:The state-of-art pedestrian detection methods are mainly restricted by weak real-time performance.To alleviate this problem,we propose a fast pedestrian detection method by taking advantages of the motion information of pedestrian in this paper.It regards pedestrian as moving objects,focusing on moving region to remove useless background information,to speed up the following detection.By well-designed predicting method,we can predict the new location of the pedestrian in a new frame.This procedure reduces the amount of computation greatly,so it’s the key for real-time pedestrian detection.To enhance the reliability of our method,we build a temporary model for every pedestrian in the scenario,and verify the predicted location with the temporary model.Finally,we experiment on various scenarios to verify the improved performance of our method.The part-based model is more suitable to deal with the deformation and occlusion of targets.Pedestrian is non-rigid and easy to be covered by other object,so part-based model enhances the robustness of pedestrian tracking.But the keypoints based model method CMT algorithm is initialized by the information of the first frame,making it not suitable for long-term tracking.In order to alleviate problems ofCMT algorithm,we are the first to propose a method to update the keypoints based model,and modify CMT algorithm to adapt to the dynamic model.To enhance the reliability of the result,we introduce the hierarchical clustering algorithm,checking the consistency of the matched keypoints,to remove outliers.Finally,the experiment results on 50 video sequences show that our modified method(AMT algorithm)are superior to the original CMT algorithm.更多还原