非重叠视域多摄像机目标跟踪方法研究

发布时间：2018-07-12 20:27

本文选题：非重叠视域 + 多摄像机目标跟踪　；参考：《西安科技大学》2017年硕士论文

【摘要】：随着安防技术的迅速发展,生活场所中摄像机数目不断增加,监控覆盖区域也逐渐扩大。如何在大范围摄像机监控视域内实现运动目标连续跟踪的难题应运而生。不同于单摄像机的是多摄像机视域中光照、目标姿势和摄像机属性等因素更具复杂性,尤其在非重叠视域下目标的运动在时空上都是离散的,为多摄像机目标跟踪带来了巨大挑战。本文对非重叠视域多摄像机目标跟踪方法展开深入研究。1)要想实现大范围多摄像机视域内的目标连续跟踪,首先需要进行单摄像机下的目标检测与跟踪。为了更加完整的获取目标区域,设计了一种基于小波变换的混合高斯背景建模检测方法。利用小波变换去除噪声使目标信息更加清晰,然后建立混合高斯背景模型提取目标。实验表明该方法相比于其他常用的检测方法具有更高的准确性。在跟踪阶段,为了改善TLD算法实时性差、对光线突变敏感等问题,提出基于Kalman滤波的TLD目标跟踪方法。用Kalman滤波器代替TLD中的光流法成为新的跟踪器,在跟踪过程中利用TLD中检测器的结果更新跟踪器。最后综合跟踪器和检测器实现目标定位。实验表明本方法相对于TLD算法来说具有更高的实时性和准确性。2)为了减少摄像机视域之间不重叠而造成的时空差异,设计了一种基于高斯和互相关函数的拓扑估计方法来获取摄像机之间的时空关系。该方法利用单摄像机目标检测与跟踪方法来确定拓扑结点。根据多个时间窗口内的平均互相关函数估计结点之间的连接关系,降低噪声干扰。最后利用高斯分布描述每条连接上的转移时间概率分布,使拓扑关系更加稳定。实验表明本方法在不需要事先获取摄像机间连接关系的情况下,能准确估计出摄像机网络的拓扑关系。3)首先为了消除摄像间光线的差异,提出基于累积颜色属性转换模型的目标匹配方法。利用多张图像学习累积颜色属性转换模型,将不同摄像机视域内的图像进行颜色转换,缩小了不同视域内同一目标的表观差异。实验表明该方法显著提高了目标的识别率。其次,为了度量不同摄像机下目标的相似性,采用基于SIFT特征的目标匹配方法,先提取SIFT描述子建立表观模型,然后给出基于序列的目标匹配策略计算匹配度,使匹配结果更加可靠。实验证明该方法能够准确的在摄像机间进行目标匹配。4)为了实现最终的非重叠视域下的目标连续跟踪,提出一种基于拓扑关系和表观模型相融合的目标关联方法。将拓扑关系、表观模型和匹配策略融合到目标关联方法的框架中。首先根据拓扑关系确定候选目标集,利用累积颜色属性转换模型处理不同视域下光线突变的问题,然后建立表观模型,利用基于序列的匹配策略度量目标的相似性,将不同视域内的运动轨迹进行关联。该方法减少了目标关联的计算量,并提升了准确性。
[Abstract]:With the rapid development of security technology, the number of cameras in the living place is increasing, and the surveillance coverage area is also gradually expanding. How to realize the continuous tracking of moving targets in the field of wide range camera surveillance arises at the historic moment. Different from the single camera, the illumination in the multi-camera field is more complicated, especially the motion of the target is discrete in time and space, especially in the non-overlapping field of view. It poses a great challenge for multi-camera target tracking. In this paper, the method of multi-camera tracking in non-overlapping field of view is studied deeply. 1) in order to realize the continuous tracking of targets in the field of view of large range and multi-camera, it is necessary to detect and track the target under single camera first. In order to obtain the target region more completely, a hybrid Gao Si background modeling and detection method based on wavelet transform is designed. Wavelet transform is used to remove noise so that the target information is clearer, and then the mixed Gao Si background model is established to extract the target. Experiments show that this method is more accurate than other commonly used detection methods. In the tracking phase, in order to improve the real-time performance of the TLD algorithm and to be sensitive to the sudden change of light, a TLD target tracking method based on Kalman filter is proposed. The Kalman filter is used to replace the optical flow method in TLD as a new tracker. In the tracking process, the result of the detector in TLD is used to update the tracker. Finally, the target location is realized by synthesizing tracker and detector. Experiments show that this method has higher real-time and accuracy than TLD algorithm. A topology estimation method based on Gao Si and cross-correlation function is designed to obtain the temporal and spatial relationship between cameras. The method uses single camera target detection and tracking method to determine topological nodes. The connection relationship between nodes is estimated according to the average cross-correlation function in multiple time windows, and the noise interference is reduced. Finally, the transition time probability distribution on each connection is described by Gao Si distribution, which makes the topological relation more stable. The experimental results show that the proposed method can accurately estimate the topological relationship of camera network without obtaining the connection relationship between cameras in advance.) first of all, the difference of light between cameras can be eliminated. An object matching method based on cumulative color attribute transformation model is proposed. Based on the cumulative color attribute transformation model of multiple images, the color conversion of different cameras in the field of view is carried out, which reduces the apparent difference of the same object in the different field of view. Experiments show that this method can improve the target recognition rate significantly. Secondly, in order to measure the similarity of targets under different cameras, the target matching method based on sift feature is adopted. Firstly, the sift descriptor is extracted to establish the apparent model, and then the target matching strategy based on sequence is given to calculate the matching degree. Make the matching result more reliable. Experiments show that this method can accurately match the target between cameras. In order to realize the target continuous tracking in the final non-overlapping visual field, a target association method based on the combination of topological relation and apparent model is proposed. The topological relation, the apparent model and the matching strategy are fused into the framework of the target association method. First, the candidate target set is determined according to the topological relation, and the problem of light mutation in different horizons is dealt with by using the cumulative color attribute transformation model. Then, the apparent model is established, and the similarity of the target is measured by the matching strategy based on sequence. The motion trajectories in different horizons are correlated. This method reduces the computation of target association and improves the accuracy.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN948.41;TP391.41

【参考文献】