无视野重叠区域的跨摄像机行人跟踪关键技术研究

发布时间：2018-05-18 15:48

本文选题：智能交通系统 + 跨摄像机行人跟踪　；参考：《华南理工大学》2016年博士论文

【摘要】：无视野重叠区域的跨摄像机行人跟踪是指运用计算机视觉、模式识别和机器学习等技术,在无视野重叠区域的不同摄像机间对相同行人进行检测、跟踪和再识别,从而实现对行人进行长时间、远距离的持续跟踪。目前,它已成为智能交通系统中的研究热点与难点问题。然而,由于在实际的交通监控环境中存在复杂的背景、光照变化、行人间遮挡、运动形变、摄像机拍摄视角和摄像机间自身属性差异等因素干扰,使得无视野重叠区域的跨摄像机行人跟踪面临着巨大的困难与挑战。本文针对上述干扰因素,对无视野重叠区域的跨摄像机行人跟踪过程中的检测、跟踪和再识别展开了深入研究,其主要的研究工作包括以下内容。一、本文提出了一种显著性二值化Haar-like特征(SLBH),该特征在融合了LBP特征和Haar-like特征优点的基础上,增加了对行人显著性信息的描述。另外,基于整体特征的行人检测算法容易受部分遮挡和运动形变的影响,本文提出了一种多部件验证的双层行人检测算法。该算法首先采用低维的SLBH特征检测整个行人,对于落入模糊区域的候选行人利用行人部件进行进一步检测,最后利用贝叶斯估计判断模糊区域的候选行人。理论分析和实验结果表明所提算法能有效的降低行人检测的虚警率。二、本文在粒子滤波框架下提出了一种权值自动调整的多特征融合的行人跟踪算法来自动适应外界环境的变化。同时,为了减少因多特征融合而造成粒子特征提取时间开销增加的缺陷,本文提出了一种类似积分图技术的方法来加快粒子区域特征提取速度。理论分析和实验结果表明所提算法能有效的适应周围环境变化,提高了行人跟踪准确率。三、行人再识别则主要完成两个摄像机间相同行人的确认工作。由于运动形变、部分遮挡和光照变化、摄像机的拍摄角度、曝光时间等自身属性差异等因素都直接影响行人再识别的准确度。因此,为了增加行人再识别算法对上述影响因素的鲁棒性,本文提出了基于分组相似度对比模型的行人再识别算法。该算法首先通过构建图库里行人样本的原型并以样本原型相似度特征代替底层特征来增加特征的语义信息。其次使用图像序列代替单幅图像作为查询图像,并采用系统抽样的方法将图像序列分组,同时设置了组内样本图像间进行相似度计算而组间不计算的规则来构建个体差异性特征。最后结合AdaBoost分类器实现行人再识别。此外,为了增强欧氏距离在两特征向量间相似度测量上的区分能力,本文提出了显著性差异距离。理论分析和实验结果表明所提算法可有效的抵抗行人在交通场景中的各种仿射变换、噪声等。四、为了增强时空信息在无视野重叠的摄像机网络中的可靠性,本文提出了基于加权时间的摄像机网络拓扑结构估计算法。该算法假设行人在不同摄像机间的转移时间满足高斯分布,通过对在平均转移时间附近到达的行人给予较大的权值,而远离平均转移时间到达的行人给予较小的权值,进而突出不同时间到达的行人对网络摄像机拓扑结构估计所做的贡献。在此基础上,结合行人检测、跟踪和再识别算法提出了时空线索融合的无视野重叠区域的跨摄像机行人跟踪算法。理论分析和实验结果表明所提算法能有效的提高摄像机网络拓扑结构估计的准确性,可更好的辅助于无视野重叠区域的跨摄像机行人跟踪。
[Abstract]:Cross camera pedestrian tracking without field of vision refers to the use of computer vision, pattern recognition and machine learning techniques to detect, track and re identify the same pedestrians in different cameras in the no field of overlap, so as to achieve a long and long distance tracking for pedestrians. At present, it has become an intelligent transportation. However, because of the complex background in the actual traffic monitoring environment, illumination changes, pedestrian occlusion, movement deformation, camera angle of view and the difference of the self property between cameras, the cross camera pedestrian tracking in the non field of vision overlaps faces great difficulties. In this paper, this paper studies the detection, tracking and re recognition of cross camera pedestrian tracking in the unoverlapping region. The main research work includes the following contents. First, this paper proposes a significant two valued Haar-like feature (SLBH), which combines the LBP features and Haar-li. In addition, the pedestrian detection algorithm based on the whole feature is easily affected by partial occlusion and movement deformation based on the advantages of Ke. In this paper, a double pedestrian detection algorithm with multi component verification is proposed in this paper. First, the algorithm uses low dimensional SLBH features to detect the whole pedestrian. The candidate pedestrians in the fuzzy region use the pedestrian component to carry out further detection. Finally, the Bayesian estimation is used to judge the candidate pedestrians in the fuzzy region. The theoretical analysis and experimental results show that the proposed algorithm can effectively reduce the false alarm rate of pedestrian detection. Two. In this paper, a multi feature fusion of weight automatic adjustment is proposed under the particle filter framework. The pedestrian tracking algorithm adapts to the change of the environment automatically. At the same time, in order to reduce the defect that the time overhead of the particle feature extraction is increased because of the multi feature fusion, a method similar to the integral graph technique is proposed to speed up the velocity of the particle region feature extraction. The theoretical analysis and experimental results show that the proposed algorithm is effective. To adapt to the changes of the surrounding environment, the accuracy of pedestrian tracking is improved. Three, the re recognition of pedestrians mainly completes the confirmation work of the same pedestrians between two cameras. Due to movement deformation, partial occlusion and illumination change, camera shooting angle, exposure time and other factors of the self attributes are directly affected by the accuracy of pedestrian recognition. In order to increase the robustness of the pedestrian re recognition algorithm to the above factors, a pedestrian rerecognition algorithm based on the group similarity contrast model is proposed in this paper. First, the semantic information of the feature is increased by constructing the prototype of the pedestrian sample in the library and replacing the underlying feature with the sample prototype similarity. Secondly, the graph is used. The image sequence is replaced by the image sequence as the query image, and the image sequence is grouped by the system sampling method. At the same time, the individual difference characteristics are set up by the similarity calculation between the sample images in the group and the non calculating rules between the groups. Finally, the AdaBoost classifier is used to realize the rerecognition of the rows. In addition, in order to enhance the Euclidean distance, Two the distinguishing ability of similarity measurement between eigenvectors, this paper presents a significant difference distance. Theoretical analysis and experimental results show that the proposed algorithm can effectively resist all kinds of affine transformation and noise in pedestrians in traffic scene. Four, in order to enhance the reliability of spatio-temporal information in the camera network without field of vision, this paper puts forward this paper. A network topology estimation algorithm based on weighted time is proposed. The algorithm assumes that the time of pedestrians' transfer between different cameras satisfies Gauss distribution and gives greater weights to pedestrians arriving near the average transfer time, while pedestrians who are far away from the average transfer time give smaller weights, and then they are different. On the basis of the pedestrian detection, tracking and rerecognition algorithm, a pedestrian tracking algorithm with a spatio-temporal clue fusion is proposed. The theoretical analysis and experimental results show that the proposed algorithm can effectively improve the network extension of the camera network. The accuracy of estimation of flutter structure can be better aided by cross camera pedestrian tracking in non visual overlap area.
【学位授予单位】：华南理工大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】