Study of techniques for motion image retrieval based on motion information

在这项研究中，我们的目标是开发主要基于运动信息检索MPEG编码的运动图像的技术。研究结果如下：1.为目标检索系统选择了一种基于草图、基于特征的模型。对于特征值，我们使用每个宏块的平均颜色和运动向量的直方图。获取视频场景中颜色和运动矢量的直方图，丢弃它们的时空分布信息。为了在维护信息的同时提高检索速度，提出了一种分层匹配技术。我们得到的结果是，运动矢量在一种极坐标系中的表示，而不是在直角坐标系中的表示，给出了一种有效的运动矢量的相似性度量，很好地符合人类的视觉特征。然后将该相似性度量实际应用于运动图像检索中，并验证了其相对于传统的直角坐标表示法的有效性。基于运动矢量和颜色直方图的匹配算法要求用户输入关于目标场景的前景和背景对象以及摄像机运动的所有信息。然而，由于这并不总是可能的，我们已经修改了算法，使得可以在不知道目标场景的所有信息的情况下执行直方图匹配。我们结合了本研究中开发的所有基本算法，构建了一个运动图像检索原型系统。在原型系统上进行的多次检索实验证明，该系统能够搜索到与用户目标场景接近的场景。

In this research, we aimed at developing techniques for retrieving MPEG-encoded motion images based mainly on motion information. The following are the results of the research :1. We selected a sketch-based, feature-based model for the target retrieval system. For the feature values, we used histograms of the average color and motion vector for each macroblock.2. Taking the histogram of the color and motion vector in a video scene drops their spatio-temporal distribution information. In order to maintain the information as well as to speed up the retrieval, we proposed a hierarchical matching technique.3. We obtained the result that a representation of motion vectors in a kind of polar coordinate system rather than in the orthogonal coordinate gives an efficient similarity measure for motion vectors which well match human visual characteristics. We then actually applied the similarity measure to the motion image retieval and confirmed its efficiency over the conventional orthogonal coordinate representation.4. The matching algorithm based on motion vector and color histograms requires the users to input all the information with respect to the foreground and background objects and camera motions for the target scene. Since this is, however, not always possible, we have revised the algorithm in such a way that a histogram matching can be performed without knowing all the information for the target scene.5. We combined all the fundamental algorithms developed in this research to make a prototype retrieval system for motion images. Numbers of retrieval experiments on the prototype system have confirmed that the system can search for scences that are close to user's target scenes.