基于大场景多视频的运动车辆检测与连续跟踪研究

发布时间：2018-06-02 14:39

本文选题：车辆跟踪 + 车辆检测　；参考：《中国矿业大学》2015年硕士论文

【摘要】：实现智慧城市,首先必须“感知城市”,其中视频传感器因其全天候、形象直观、高覆盖、实时性等优点得到广泛运用。目前城市视频监控以道路交通监控与安全监控为主,视场窄、功能单一,且不能实现对目标的连续、实时跟踪,在“感知城市”中没有发挥应有的作用。本文针对目前视频传感器遍布城市这一现状,研究多视频联合的大场景下车辆连续跟踪所涉及的关键问题。主要工作和所取得成果如下:(1)提出多视频联合的大场景下运动车辆连续跟踪的思路。首先在单视场实现运动车辆的检测,确定跟踪目标;然后对跟踪目标进行特征描述,实现单视场的车辆跟踪;最后当车辆处于重叠视域时,利用目标交接算法将多相机视域中的目标进行关联,从而实现跨相机车辆的连续跟踪。(2)基于优化混合高斯模型的运动车辆检测。经典混合高斯模型背景建模时主要存在三方面不足:匹配高斯分布选择时未综合考虑分布权重及自身匹配程度、模型参数进行更新时模型参数学习率r固定不变以及背景显示时只显示权值方差比最大的分布。本文在GMM基础上,对匹配高斯分布选择、模型参数更新、背景显示三方面进行优化。实验证明优化算法提取车辆轮廓清晰,对场景变化有较强适应性,甚至对强光与树叶晃动等扰动都能良好处理,具有一定实用性。(3)基于Kalman滤波与颜色纹理联合特征的自适应Mean Shift单视场运动车辆跟踪。在车辆跟踪过程中,本文首先构建纹理与颜色联合特征直方图来描述目标;然后利用Kalman滤波进行目标运动估计,预测目标在当前时刻的位置;最后在预测位置的可信区域内进行联合特征的Mean Shift最佳匹配区域检测,且搜索窗口的大小根据前一时刻最优目标区域的矩信息调整。实验证明本文自适应Mean Shift跟踪算法,能够克服相似目标和目标形变的干扰,最大程度减少迭代搜索时间,即便目标被部分遮挡仍能保持有效跟踪。(4)基于视野分界线的跨相机运动车辆连续跟踪。跨相机车辆连续跟踪问题的本质是目标交接,本文从基于视野分界线的目标交接算法入手,着重研究存在重叠视域的跨相机车辆交接。首先利用基于投影不变量的算法生成视野分界线,再通过目标与视野分界线的关系判定其所属区域,最后通过欧氏距离与SIFT特征匹配双重条件进行相同目标的关联,提高匹配准确度,更好地实现目标交接。
[Abstract]:In order to realize a smart city, we must first "perceive the city", in which the video sensor is widely used because of its all-weather, visual image, high coverage, real-time and so on. At present, urban video surveillance mainly focuses on road traffic monitoring and safety monitoring, with narrow field of view, single function, and can not achieve continuous, real-time tracking of the target, which does not play its due role in "perceived city". In view of the fact that the video sensors are all over the city, this paper studies the key problems of continuous vehicle tracking in large scale scene with multi-video joint. The main work and results are as follows: 1) the idea of moving vehicle continuous tracking in multi-video joint large scene is proposed. First, the moving vehicle detection is realized in the single field of view, and the tracking target is determined. Then, the tracking target is described to realize the vehicle tracking in the single field of view. Finally, when the vehicle is in the overlapping field of view, The target handover algorithm is used to correlate the targets in the multi-camera horizon to realize continuous tracking of cross-camera vehicles. The moving vehicle detection is based on the optimized hybrid Gao Si model. There are three main shortcomings in the background modeling of classical mixed Gao Si model: the distribution weight and its matching degree are not considered when selecting the matching Gao Si distribution. When the model parameters are updated, the learning rate of the model parameters is invariant, and the distribution of the maximum ratio of weights and variances is only shown when the background is displayed. On the basis of GMM, this paper optimizes the matching Gao Si distribution selection, model parameter updating and background display. The experimental results show that the optimized algorithm can extract the vehicle contour clearly and has a strong adaptability to the scene change, and can deal with the disturbances such as strong light and leaf sloshing, etc. An adaptive Mean Shift single field moving vehicle tracking algorithm based on Kalman filter and color texture features is presented. In the process of vehicle tracking, this paper first constructs the texture and color histogram to describe the target, then uses Kalman filter to estimate the target motion, and then predicts the position of the target at the current time. Finally, the Mean Shift optimal matching region of joint features is detected in the trusted region of the predicted location, and the size of the search window is adjusted according to the moment information of the prior optimal target region. Experimental results show that the proposed adaptive Mean Shift tracking algorithm can overcome the interference of similar targets and target deformation and minimize the iterative search time. Even if the target is partially occluded, it can keep effective tracking. 4) continuous tracking across camera moving vehicles based on the visual field demarcation line. The essence of the continuous tracking problem of cross-camera vehicles is the target handover. This paper focuses on the cross-camera vehicle handover algorithm based on the visual field dividing line of vision. Firstly, the projection invariant algorithm is used to generate the visual field boundary, and then the region is determined by the relationship between the target and the visual field boundary. Finally, the Euclidean distance and SIFT feature matching double conditions are used to correlate the same target. Improve the accuracy of matching, better achieve the transfer of objectives.
【学位授予单位】：中国矿业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U495;TP391.41

【参考文献】