【作者】 董文会；

【导师】 常发亮；

【作者基本信息】 山东大学 ， 模式识别与智能系统， 2015， 博士

【摘要】 随着信息技术的发展及社会安防需求的不断提高,传统的人工监控技术已不能满足人们的需求,智能化和主动化的目标监控技术成为目前主要的发展方向。作为智能监控的重要组成部分,多摄像机监控网络中的目标连续跟踪成为研究的热点,并具有较强的理论意义和广泛的应用价值。多摄像机目标连续跟踪包括单摄像机的目标跟踪和目标在摄像机之间的交接跟踪两个组成部分。与传统的单摄像目标跟踪相比,由于摄像机数目的增多,该课题引入了许多新的问题,如摄像机网络监控场景复杂、各摄像机参数不同及目标在各摄像机中的运动轨迹不同等。如何克服这些难题,实现目标在多摄像机网络中的准确跟踪是智能监控领域亟待解决的问题之一。本论文针对非重叠视野下多摄像机目标连续跟踪进行了研究,重点研究了鲁棒的单摄像机目标跟踪算法、多摄像机视野的拓扑结构学习、多摄像机下目标表观模型的学习和目标匹配及多摄像机间目标的准确交接等几个方面的内容。论文的具体工作及研究成果如下：1)研究了单摄像机目标跟踪方法,并提出了稀疏表达多特征融合的单摄像机跟踪算法。本算法以粒子滤波为总体跟踪框架,包括观测概率模型和状态模型两部分。观测概率模型通过将多个特征进行稀疏表达融合的方式构建,状态模型为六参量的高斯分布。为了使该算法能适应目标在运动过程中的各种变化,本文在跟踪算法中加入了具有自适应的模板更新算法。实验表明,该跟踪算法可以处理跟踪过程中可能发生的光照改变、遮挡、姿态改变和背景复杂等情况,实现准确而有效的目标跟踪。2)研究了非重叠视野下多摄像机视野的拓扑结构学习方法。摄像机视野的拓扑结构包括进／出口区域(节点)、节点间的转移时间分布及节点间的关联性三个要素。本文通过对监控视频中目标跟踪信息的统计,提出了非重叠视野下多摄像机视野的拓扑结构学习方法。首先采用高斯聚类方法对各摄像机视野中的进／出口区域进行学习,然后通过累加交叉相关及高斯拟合的方式对节点间的转移时间分布进行学习,最后利用互信息理论对节点间的关联性进行精简,去除虚假关联的节点,从而获得准确的拓扑结构。通过对摄像机视野拓扑结构的学习可以获得多摄像机目标连续跟踪过程中需要的时空推理信息(如当目标运行到某摄像机出口节点后选择哪一个摄像机作为交接摄像机、目标从出口节点到交接摄像机的入口节点需要的时间范围等),为合理调度摄像机提供依据。3)研究了目标在多个摄像机视野中运动时表观变化的特点,提出了基于特征树的目标表观模型学习和目标匹配方法。本文对不同摄像机下的目标表观进行了学习,提出了基于特征树的目标表观模型学习和目标匹配方法。通过对不同特征分别进行非督导聚类获得不同特征的树状结构,从多个角度对目标进行全面描述。当在不同摄像机下目标表观发生改变时,可以根据获取的目标新样本对特征树进行更新,并对目标之前表观具有记忆能力。目标匹配采用特征树投票的方式,候选目标需要经过特征树每一层相似性的计算。这种层层筛选的方式,可以大大减少计算量,提高匹配效率。4)研究了目标在多个摄像机之间的交接问题,提出了基于继承和学习的多摄像机目标交接算法。目标在多个摄像机之间的准确交接是多摄像机监控网络中目标连续跟踪能否成功的关键。本文将时空约束信息和目标表观模型进行融合,提出了基于继承和学习的多摄像机目标交接算法。该算法将单摄像机目标跟踪、拓扑结构学习及目标表观模型信息融入整体交接跟踪框架。首先通过拓扑结构学习获取目标在整个网络中运动时的时空信息,消除模糊因素的干扰。当确定跟踪目标后,通过单摄像机目标跟踪算法对其进行跟踪并构建目标在该摄像机下的表观模型。在目标交接启动时,将该模型传递给交接摄像机。交接摄像机利用上一摄像机传递的表观模型在一定的时间范围内对相应的入口区域进行目标匹配。若检测到目标,则交接完成并在该摄像机下继续跟踪,然后利用跟踪结果对目标表观模型进行再学习。随着目标在摄像机之间的跟踪,目标的表观模型将越来越精确,使得目标在摄像机之间的交接和连续跟踪更顺畅。综上所述,本文围绕多摄像机监控网络中的目标连续跟踪问题进行了多个方面的有益尝试并获得较好的实验结果,为该课题的研究提供了新的解决思路。更多还原

【Abstract】 With the development of information technology and the increasing demands of social security, the traditional surveillance technologies can not meet people’s needs. Intelligent and automated surveillance technologies are becoming the main development direction. As an important part of intelligent surveillance, target continuous tracking in multi-camera surveillance network becomes the hot research spot, which has an important theoretical significance and extensive application value. Target continuous tracking in multi-camera surveillance netwok not only includes visual tracking with single camera, but also includes the target handoff among cameras. Compared with the traditional visual tracking with single camera, many new difficulties are brought due to the increasing number of cameras. Such as the complexity of the surveilled scenes, the different parameters of cameras, the different traces of the targets, and so on. How to overcome these difficulties and achieve accurate target continuous tracking in the multi-camera network is one of the urgent tasks in intelligent surveillance field.This dissertation focuses on target continuous tracking method in multi-camera surveillance network with non-overlapping view fields. The main content includes the robust visual tracking algorithm in single camera, the topology learning method of camera views, the appearance model construction of the tracked target and target matching in multi-camera network, and the accurate target handoff among cameras.The main research work and the results of this dissertation are as follows:1) Visual tracking method in single camera is studied and multi-feature sparse representation based visual tracking method is proposed. The algorithm uses particle filter as the framework, including observation model and state model. The observation model is constructed by multi-feature sparse representation. The state model uses Gaussian distribution with six parameters. In order to adapt the variations of the target during tracking, the adaptive template update algorithm is added to the tracking algorithm. Experimental results show that the visual tracking algorithm can handle many complexity conditions, such as illumination variation, occlusion, pose variation and cluttered background, and achieve accurate and effective tracking results.2) Topology learning of the multi-camera views with non-overlapping view is studied. The topology of the multi-camera views includes entrance/exit zones (nodes), transition time distribution between the nodes and the links between the nodes. The topology learning method of the multi-camera views is proposed through the statistical information of target traces in the surveillance videos. Firstly, the entrance/exit zones are learned by Gaussian clustering. Then, the transition time distribution between two nodes is gotten by accumulated cross correlation function and Gaussian fitting. Finally, the false links between the nodes are removed by mutual information theory and the accurate topology is recovered. According to the topology of the multi-camera views, the time-space information can be gotten,which can be used as a reference for camera scheduling in the multi-camera target continuous tracking (For example, when the target arrives at the exit node of one camera, the handoff camera and the travel time range can be gotten from the topology).3) The appearance variation characteristics of the target in the multi-camera network are also studied. The feature tree based appearance model construction method and target matching method are proposed in this dissertation. Based on the research of appearance model construction methods and key problems existing, combining the characteristics of the target continuous tracking in multi-camera surveillance network, the dissertation presents the appearance model construction method and target matching method based on the feature trees. The trees are constructed by unsupervised clustering with different features, which can describe the target comprehensively. When the appearance varies during the tracking, the feature trees are updated by the new samples of the target. The feature trees not only can adapt to the appearance variation but also can recall the former appearance of the target. Feature tree voting is used for target matching. The similarity between each candidate and the target should be calculated in every layer of the feature tree. The screening method can greatly reduce the computational complexity and improve the effiency of the matching.4) After studying the key problems of the target handoff among cameras, the target handoff algorithm in multi-camera based on inheriting and learning is proposed in this dissertation. The accurate target handoff among cameras is the key factor for the successful continuous tracking in the multi-camera network. By fusing the time-space information and the appearance model, the target handoff algorithm based on inheriting and learning is proposed. This algorithm integrates the visual tracking with single camera, the topology of the camera view fields and the target appearance model into the whole target handoff framework. Firstly, the space-time information is obtained from the topology to eliminate the disturbances of the uncertain factors. When the target is determined, it is tracked using the single camera visual tracking method and the appearance model in this camera is constructed. Once the target handoff is triggered, the appearance of the target will be transferred to the handoff camera. Then, the handoff camera searches the target at the corresponding entrance zone during the transition time interval using the appearance model. If the target is detected, the handoff camera will continue to track the target in its field of view and the target appearance model will be updated using the tracking results. With the continuous target tracking in the multi-camera network, the appearance model of the target will be more and more accurate, which will make the target handoff and tracking smoother and smoother.In conclusion, the research works of this dissertation are useful attemptes in many aspects of target continuous tracking in multi-camera surveillance network and good results are obtained. These works will provide a new idea for the target continuous tracking in multi-camera surveillance network.更多还原