In this paper, hashing learning is applied for proposing a method of high efficient motion capture data encoding and content-based retrieval. In the preprocessing step, for each clip sequence of motion capture data in the database, we first extract feature from a number of consecutive frames, called sliding window, to avoid local sensitive of the motion actions, then combing all the sliding window features to form the feature of the motion clip. For a given motion clip, we use a hierarchical strategy to do retrieval. First we hashing encode the represen-tative frames of the clip, then search the entire database using the hashing code to retrieve the first several similar window-clips. Among the selected clips, we finally output those ones that are with minimum accumulated errors computed frame by frame. It is demonstrated by the experiments that the proposed method has the advantage in timing performance when doing retrieval from a very large scale motion capture database.%针对动作捕捉数据高维度、检索耗时问题，提出一种基于哈希学习的高效编码和快速检索算法。该算法对动作捕捉库中的每个运动序列，顺序将固定帧数的帧序列作为滑动窗口，以滑动窗口为单位抽取特征，将一个帧序列其所有滑动窗口特征作为其特征表达，通过哈希学习方法对每个序列进行哈希编码；检索时采用分层检索策略，对待检索序列提取其关键帧，用哈希编码方法检索库中与关键帧相似的若干窗口帧，并通过均匀帧采样计算帧间误差进一步筛选包含相似窗口帧的序列段，获得数据库中与待检索序列相似的若干相似序列段。实验结果表明，文中算法可实现从大规模动作捕捉数据库中快速检索相似序列段。
展开▼
