【作者】 刘献如；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 控制科学与工程， 2011， 博士

【摘要】 视频图像序列目标跟踪是目前计算机视觉领域研究的热点问题之一,它广泛应用于智能监控、智能交通、汽车辅助驾驶、精确制导等民用和军事领域,因此,对图像序列中的感兴趣目标进行检测与分析具有非常重要的现实意义。论文围绕图像序列目标跟踪中的运动目标检测、目标跟踪、多目标数据关联算法和遮挡处理等重点与难点问题展开研究,并将这些方法应用于视频监控以及相关领域,为其目标检测与跟踪提供理论支持与应用指导。论文的主要工作包括以下四个方面。(1)针对复杂场景下运动目标难以提取和分割问题,在深入分析现有算法的基础上,提出一种新颖的运动目标检测方法,该方法首先对相邻图像帧进行基于SAD的块匹配获取视差图,然后通过对视差图及其直方图分析,提取场景中的运动目标及其运动信息,实现场景中动态目标检测；对于车载摄像机的目标检测,采用毫米波(MMW)雷达与图像序列相结合的方法实现智能车对前方障碍物的检测,针对多传感器之间标定需传感器内外参数等复杂性操作问题,提出一种基于运动检测的毫米波雷达与摄像机之间的标定方法,该方法有效地实现毫米波雷达数据与图像数据之间的映射；针对智能车辆根据车道线进行安全导航的要求,将改进的Otsu图像分割算法与KF滤波估计算法相结合,实现对结构化道路车道线的鲁棒检测与实时跟踪。(2)对常用跟踪滤波器卡尔曼滤波(KF)、扩展卡尔曼滤波(EKF)、无迹卡尔曼滤波(UKF)和粒子滤波(PF)算法进行了深入研究与实验分析,实验结果表明,UKF算法较之其它滤波算法具有较好的综合滤波性能；对EKF建议和UKF建议的PF算法以及PF算法进行深入研究与实验分析,实验结果表明UKF建议的PF算法具有较高的估计精度,同时进一步验证了UKF算法的优越性；针对Mean shift算法由于其理论缺陷容易导致跟踪快速运动目标时容易跟丢的问题,提出一种基于无迹卡尔曼波的Mean shift实时目标跟踪算法,该算法先用UKF算法对目标状态进行估计,再由Mean shift算法在此估计邻域定位目标的准确位置,实验结果表明改进的Mean shift目标跟踪算法较传统Mean shift算法具有较好的跟踪性能。(3)针对多目标跟踪时的数据关联问题,对最近邻(NN)、概率数据关联(PDA)、联合概率数据关联(JPDA)和多假设(MHT)等多目标数据关联算法进行深入研究与实验分析,实验结果验证了MHT算法能有效处理密集杂波环境下的数据关联问题；针对图像序列多目标数据关联的特殊性,提出一种基于UKF-MHT的多目标数据关联算法,该方法先采用核直方图描述目标,引入巴氏系数度量观测与先验目标的相似性,并由此建立假设关联矩阵。将所提出的多目标数据关联方法应用于场景中多目标数据关联,实验结果表明所提出的多目标数据关联算法的有效性和可行性。(4)针对视频图像序列目标跟踪中的遮挡问题,提出一种带遮挡关系的UKF-PF遮挡跟踪算法。该算法首先对目标状态进行预测估计,并对可能存在的遮挡目标进行遮挡检测,再对遮挡目标及遮挡区域进行自适应栅格分块；然后对遮挡区域栅格块通过相似性度量进行分类,由此建立目标之间的遮挡关系矩阵；最后,采用基于UKF的自适应粒子滤波算法,结合目标之间的遮挡关系,选取有效粒子的观测值更新对应目标粒子的状态。对多段图像序列进行实验测试与比较分析,结果表明所提出的带遮挡关系的UKF-PF遮挡跟踪方法的有效性和可行性。更多还原

【Abstract】 Object tracking in image sequence is one of the hottest problems in the field of computer vision. The object tracking technique is widely used both in military and civilian entities, such as intelligent supervision, intelligent transportation system, vehicle assistance driving system and precision guidance. So it is quite significant to study on the interest object detection and tracking in image sequence. The dissertation mainly focused on some vital and difficult problem of the object tracking technique including moving object detection, object tracking, multiple object data association and occlusion tracking and so on. Some research methods will be used in the visual surveillance and relative application fields, and supply theory supports and technique director for the object detection and tracking based on image sequence. The dissertation mainly contains the following four aspects.(1) Aiming at the object detection and segmentation problems in the complex scenes, a novel object detection method was proposed based on studying the present detection method. The proposed method firstly got the disparity image between two adjoin image frames based on SAD block matching method. And then moving objects and their moving information were extracted from the disparity image. The proposed method could effectively segment most moving objects for the static camera system. While for the object detect system with the camera on vehicle we used camera combined millimeter-wave radar to detect the obstacles in front of the intelligent vehicle. Being dead against to the complicate of the calibration needing internal and external parameters between multiple sensors,a simple and quick calibration method between millimeter-wave radar and camera was proposed based on motion detection method. According to the need of safe navigation of the intelligent vehicle based on lane a robust detection and real time lane tracking algorithm based on improved otsu’s method was proposed for the structured road.(2) Some tracking filters such as Kalman Filter(KF), Entended Kalman filter(EKF), Unscented Kalman filter(UKF) and Particle filter(PF) were studied in detail and experimentally analylized. The results show that the UKF has better filter performance comparing against other filters. Some found studing on the UKF proposal PF,PF based on EKF and PF were carried on in detail, and simulating results show that the UKF-PF has better estimation precision.At the same time it also presents that UKF is superior for state estimation. Aiming at the theory shortage of the Mean shift, which easily resulted in miss the tracked object with high moving velocity, a improved mean shift tracking algorithm based on UKF was proposed. And the propsed method firstly used UKF algorithm to estimate the object state.And then the Mean shift algorithm was applied to locate the object precisely. The experiments show that the UKF-Mean shift tracking has better tracking performance both in tracking accuracy and time costing than the traditional Mean shift.(3) Data association problem is vital for multiple objects tracking. Nearest Neighborhood(NN), Probability Data Association(PDA), Joint PDA(JPDA) and Multiple hypothesis tracking(MHT) and some popular data association algorithms were studied foundly and analyzed. Simulating results show that the MHT data association has better filter performance than JPDA. Aiming at the speciality of the image object, a kind data association algorithm based on UKF-MHT was proposed.The proposed algorithm utilized the kernel histogram presenting targets.And the Bhattacharyya factor was introduced into evaluate the similarity between observation and prior targets. According similarity the hypothesis association matrix was built for MHT algorithm. Some experiments show exciting results for the proposed data association algorithm.(4)Considering the occlusion problem during the object tacking process a kind of UKF-PF occlusion tracking algorithm with occlusion relation was proposed. The algorithm firstly declared the occlusion and occlusion region by the prediction. And then occlusion objects and occlusion regions were segmented grid blob by a adaptive size method. The occlusion relation matrix was constructed by evaluating the similarity of the occlusion region and prior objects. Finally the UKF-PF with occlusion relation algorithm was applied to track occlusion objects. The state of the tracked object is updated by the observation of the valid particles. The propsed occlusion tracking algorithm is applied into several image sequnces with different occlusion situations.And the results show that the proposed algorithm is efficient and feasible.更多还原