本文關(guān)鍵詞： 動作捕捉數(shù)據(jù) 相似檢索 哈希編碼　出處：《山東大學(xué)》2017年碩士論文　論文類型：學(xué)位論文

【摘要】：動作捕捉數(shù)據(jù)具有冗雜度高、數(shù)量級大、特征維度高等特點,使得動作捕捉數(shù)據(jù)在檢索時耗費時間較長。本文提出一種基于哈希學(xué)習(xí)的高效編碼和快速檢索算法以達(dá)到在大規(guī)模動作捕捉數(shù)據(jù)集中快速檢索相似序列段的目標(biāo)。哈希算法對于高維數(shù)據(jù)的編碼檢索方面具有較好的性能,本文將數(shù)據(jù)庫中的每一序列,以固定寬度的窗口,從起始幀開始依次向后滑動,每滑動若干幀統(tǒng)一編碼窗口內(nèi)的所有骨架幀,滑動窗口的策略很好地將序列的局部差異性考慮在內(nèi)。編碼方式采用哈希學(xué)習(xí)的策略,得到的哈希碼相比于原始特征長度更短,占據(jù)空間更少。序列檢索時采用分層策略,使用關(guān)鍵幀粗略搜索可能相似序列段,然后計算關(guān)鍵幀以及采樣幀的累積誤差,兩者加權(quán)以確定檢索排序。本文從兩個角度論證基于哈希編碼的序列檢索性能:一是比較使用哈希編碼與未使用哈希編碼的檢索時間;二是比較使用哈希編碼與未使用哈希編碼的檢索準(zhǔn)確率。實驗結(jié)果表明,采用哈希學(xué)習(xí)的方法可以有效地提升動作捕捉數(shù)據(jù)的檢索性能,這為大規(guī)模動作捕捉數(shù)據(jù)的管理與檢索提供了很好的選擇。
[Abstract]:The motion capture data has the characteristics of high jumbled degree, large order of magnitude, high characteristic dimension, etc. This paper presents an efficient coding and fast retrieval algorithm based on hashing learning to achieve the goal of fast retrieval of similar sequence segments in large motion capture data sets. Hashing algorithm has better performance in coding and retrieval of high-dimensional data. In this paper, each sequence in the database is slid backward from the start frame to the fixed width window, and every sliding frame is unified to encode all the skeleton frames in the window. The strategy of sliding window takes into account the local difference of sequence very well. The hashing learning strategy is adopted in the coding method, the hash code is shorter than the original feature length and occupies less space, and the hierarchical strategy is used in the sequence retrieval. Use key frames to roughly search possible similar sequence segments, and then calculate cumulative errors for key frames and sample frames. This paper discusses the retrieval performance based on hashing coding from two angles: first, comparing the retrieval time between using hash coding and not using hash coding; The experimental results show that the hash learning method can effectively improve the retrieval performance of motion capture data. This provides a good choice for large-scale motion capture data management and retrieval.
【學(xué)位授予單位】：山東大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41

【參考文獻(xiàn)】

相關(guān)期刊論文 前1條

1 李武軍;周志華;;大數(shù)據(jù)哈希學(xué)習(xí):現(xiàn)狀與趨勢[J];科學(xué)通報;2015年Z1期

，

本文編號：1507648