无重叠视域多摄像机目标匹配算法研究

发布时间：2018-11-28 13:47

【摘要】：随着社会的进步,在重要场合对多摄像机的需求越来越大。单摄像机的目标跟踪方法已经日臻完善,但是由于其监控范围小,较难的应用于大范围的目标跟踪中。用摄像机覆盖所有的监控区域容易造成摄像机的浪费,所以只需要覆盖重要场合,这就造成了摄像机之间存在一定盲区,盲区的存在将摄像机之间运动目标在时间和空间上分离,目标在摄像机之间的成像特征不同导致匹配非常困难。本文主要工作以及创新点如下：(1)针对传统的VIBE算法提取的目标前景无法解决目标粘连和阴影问题,本文提出了使用VIBE与HOG+SVM特征分类相结合的多运动目标检测算法。利用VIBE算法检测出运动目标区域,将检测出的前景送入HOG特征的SVM分类器,分割得到正确目标,实验证明该算法具有较好的鲁棒性。(2)在多摄像机目标匹配中,针对不同环境下的每个特征在摄像机目标匹配中的作用不一样,选择不同特征决定匹配效果的好坏,本文选用基于颜色补偿的颜色直方图匹配；基于LMNN的H分量匹配,基于扩散距离改进的SURF特征匹配。颜色直方图可以消除摄像机之间的光照差异；基于LMNN的H分量特征能够加大不同目标的距离,使相同目标距离变小,便于匹配度的区分；基于扩散距离改进的SURF算法利用扩散距离的特性,消除了目标的误匹配,提高了匹配效率。(3)将摄像机之间的不可见区域训练一个拓扑关系,将拓扑关系作为一个匹配特征,作为摄像机对目标的选择,超过阈值的目标不进行匹配,在阈值之内的特征才用于摄像机之间的匹配。(4)将所有的特征进行D-S多特征融合,消除误匹配,获得不同情况下的最好匹配精度。
[Abstract]:With the development of society, the demand for multi-cameras is increasing on important occasions. The target tracking method of single camera has been improved day by day, but it is difficult to be used in a wide range of target tracking because of its small monitoring range. The use of cameras to cover all the monitoring areas is likely to lead to waste of cameras, so only important occasions need to be covered, which results in a certain blind area between cameras. The existence of blind area separates the moving targets between cameras in time and space. The difference of imaging characteristics between cameras leads to the difficulty of matching. The main work and innovations of this paper are as follows: (1) aiming at the problem that the target foreground extracted by the traditional VIBE algorithm can not solve the problem of target adhesion and shadow, a multi-moving target detection algorithm combining VIBE and HOG SVM feature classification is proposed in this paper. The VIBE algorithm is used to detect the moving target region, and the detected foreground is sent into the SVM classifier of the HOG feature to segment the correct target. The experiments show that the algorithm is robust. (2) in the multi-camera target matching, the proposed algorithm is robust. In view of the different function of each feature in the camera target matching under different environment, the color histogram matching based on color compensation is chosen to select different features to determine the matching effect. H component matching based on LMNN and SURF feature matching based on diffusion distance. The color histogram can eliminate the illumination difference between cameras, the H component feature based on LMNN can increase the distance of different targets, make the distance of the same object smaller, and facilitate the discrimination of matching degree. The improved SURF algorithm based on the diffusion distance eliminates the mismatch of the target and improves the matching efficiency. (3) the invisible region between the cameras is trained with a topological relation. The topological relation is regarded as a matching feature and the target is selected by the camera. The target above the threshold is not matched. The features within the threshold are used to match the cameras. (4) all the features are fused with D-S multi-features to eliminate the mismatch and obtain the best matching accuracy in different cases.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.41

【相似文献】