Research on Digital Archives for Human Motion Data and Video Data

在视频数据库的研究中，对视频数据进行索引是实现快速、准确检索的必要条件。在索引视频数据时，最重要的过程是将视频分割成称为场景的连贯单元，而不仅仅是我们称为剪切的部分视频。人工分割是准确的，但它需要大量的时间和关于视频内容的背景知识。通过综合提取视频中的颜色、文本和时间信息，我们开发了一种自动分割视频的方法，并重点研究了人体运动数据的标注。运动数据具有以下特征：身体某些部位的运动会影响身体其他部位的运动。我们将这种依赖称为运动关联规则。因此，标记运动数据的任务等同于通过使用运动关联规则来表达运动的任务。由于人体运动数据具有上述相关性，我们将运动数据视为多数据流。为了发现运动关联规则，降低任务成本，我们将高维的运动数据转换为多个低维的符号序列。每个符号唯一地表示原始运动。符号序列可以通过基于内容的相似片段的分割和聚类来获得。

In researches on video database, indexing video data is necessary for quick and correct retrieval. The most important process in indexing video data is to segment video into coherent units called scenes, more than just partial video that we call cuts. Manual segmentation is accurate, but it takes a lot of time and background knowledge about the content of video. We have developed the method to segment video automatically, by integrating color, textual, and temporal information extracted from video.Furthermore, we have focused on tagging human motion data. The motion data has the following features : movements on some body parts influence those on other body parts. We call this dependency as motion association rule. Thus the task of tagging motion data is equal to the task of expressing motion by using motion association rules. Because human motion data has above dependency, we treat motion data as multistream. In order to find motion association rules and cut off the cost of the task, we convert motion data of high dimension into multiple sequences of symbols of lower dimension. Each symbol uniquely represents primitive motion. The sequences of symbols can be obtained by content-based segmentation and clustering of similar segments.