【作者】 祝博荟；

【导师】 丁永生；

【作者基本信息】 东华大学 ， 控制理论与控制工程， 2013， 博士

【摘要】 视频监控系统中,监控视频中出现的人都是其重点关注的对象,因此智能监控系统需要拥有对行人进行检测,再识别,跟踪的能力,以便进一步对行人目标的行为进行分析。这就要求监控系统拥有可靠的行人检测以及行人再识别技术。然而由于行人姿态复杂多变、尺度变化明显并且应用场景易受背景、光照、阴影、摄像头参数等应用环境的干扰,使得行人检测以及行人再识别技术目前尚无在可靠性和速度方面都令人满意的解决方案。本论文针对这一情况,利用深度图像不受光照变化影响以及同一物体空间信息的一致性,通过研究背景消除、人体分割、深度与视觉信息融合,视角识别,关键帧选择等问题,建立了基于深度与视觉信息融合的行人检测与再识别模型。首先本文针对行人检测易受遮挡以及光照变化干扰的问题,提出了融合深度与视觉信息的行人检测方法。利用深度图像不受光照影响的特性将其引入行人检测,避免检测过程中来自照明变化的干扰。将行人检测问题转换为对行人头部的检测,减少了遮挡与姿势变化对检测结果的影响。在深度图像与彩色图像上分别建立头部检测器,并通过决策级信息融合得到漏检率更低的头部检测器。利用同一物体表面深度信息连贯性,提出了基于图论的人体提取方法,使得待检测行人只要头部能够检测到就能用其提取全身像素,将行人与背景分离。实验证明了该方法提高了应对遮挡以及人体姿态变化干扰的能力。然后利用人体曲面上两点间的测地距离不变特性以及人体骨骼所含的语义及空间信息,提出了基于人体骨骼的空间距离特征,并设计了基于此特征的人体部位识别算法。最后通过实验对该算法的可行性进行了验证。接下来针对现有的人体外貌模型易受人体姿势以及摄像机视角变化的干扰导致行人再识别错误的问题,从人体各个部位分别提取外貌特征,并将其与基于骨骼的空间特征相结合,建立了基于深度和视觉信息融合的人体外貌模型,提高外貌模型的鲁棒性与可区分度,从而实现行人再识别性能的提高；提出了基于再识别概率最大化准则的行人相似度函数训练方案,运用免疫进化算法得到最优的相似度函数,并通过实验验证运用该准则训练得到行人再识别方案要优于基于其他训练方法的的行人再识别方法。最后,对多镜头行人再识别技术进一步分析,提出了基于支持向量机的行人视角识别,从而解决视角变化对行人再识别的干扰；针对多镜头下图像冗余的问题,进一步提出了基于人体骨架的关键帧提取技术,实现对不同姿态行人的选择；建立了包含全局特征与周期性局部特征的新型人体外貌模型。实验结果验证了该模型能提高行人再识别的识别率与鲁棒性。文章末尾总结了总结了论文的研究内容,指出了研究中存在的不足,展望了下一步的研究方向。更多还原

【Abstract】 In video surveillance systems, human appearing in surveillance video are the object of its focus, so intelligent monitoring systems need to have a ability of pedestrian detection, re-identification, and tracking, so as to further analyze the behavior of targeted pedestrians. This requires the monitoring system has a reliable technology for pedestrian detection and re-identification. However, due to pedestrian posture complexity and variability, scale changes, as well as the fact that application scenario is susceptible to interference from application environment, such as background, light, shadows, camera parameters, pedestrian detection and re-identification technology is still unsatisfactory in terms of reliability and speed at present.Upon this situation, applying the principle that depth image is robust against illumination changes and same object keeps consistency of space information, the thesis established pedestrian detection and re-identification model based on fusion of depth and vision information, through research on background subtraction, human body segmentation, fusion of depth and vision information, viewpoint identification, keyframe selection and other issues.Firstly, we proposed pedestrian detection based on fusion of depth and vision information, as pedestrian detection is vulnerable to interference of occlusion and illumination changes. Depth image is introduced to pedestrian detection to avoid interference of illumination changes as depth images is characterized by robustness against illumination variation. And pedestrian detection problem is transformed to detection of human head in order to eliminate the influence of occlusion and posture changes on detection result. Then, the thesis built head detector respectively for depth image and color image, and employed decision-level information fusion to obtain head detector with lower miss rate. By the light of depth information continuity of the same object surface, graph theory-based human feature extraction methods were proposed, which makes extraction of the whole body pixel possible as long as pedestrians’heads can be detected, so that pedestrian and background can be seperated. Experiments show that the method improves the ability to counter interference of occlusion and posture changes.According to invariability of geodesic distance between two points on the human body surfaces, as well as applying context and space information contained in human skeleton, we proposed spatial distance features based on human skeleton and designed a human part detection algorithm based on these feature. Finally, experiments verify the feasibility of this algorithm.After that, we built human appearance model based on fusion of depth and vision information through extracting appearance model from all of human parts and then combining it with skeleton-based spatial information, for avoiding re-identification errors existing in the current human appearance model as those models are susceptible to posture and camera view changes. The method improves robustness and discrimination of the appearance model, thus achieves enhancement on pedestrian re-identification performance. Then, training scheme of pedestrian similarity function based on maximization of re-identification probability was proposed. We used immune evolutionary algorithm to get the optimal similarity function and verify by experiment that our pedestrian re-identification scheme trained with this rule is superior to pedestrian re-identification method that are trained with other rules.At last, further analysis was made on multi-shots pedestrian re-identification. We proposed pedestrian viewpoint identification method so as to make re-identification immune from interference of viewpoint changes; Due to image redundancy problem under multi-shots, key frame extraction technology based on the human skeleton was proposed, for achieving selection of pedestrians with different postures; and we established a new human appearance model, which contains global features and cyclic local features. Experimental results demonstrate that the model can improve re-identification accuracy and robustness.In the end, we summarize the content, advantage and deficiency of the paper, and narrate further research direction.更多还原