弹性动作关联序列的模式检索

视频检索要求在数据库中能够找到与待检索视频序列相关的序列,该相关序列也可以是数据库中某个序列的部分序列,为此提出了一个动态编程方法.该方法综合了计算信息领域的编辑距离算法和信号处理领域的动态时间规整(DTW)算法,并克服了它们都不能直接用于动作检索的局限性,由于可以通过调节参数使所提方法转换为上述2种方法,所以可把它们看作是所提方法的特例.通过在不同数据库上的实验,以及和编辑距离算法和动态时间规整算法的比较,发现在相同的误检率条件下,所提算法的检测率要好于其他2种算法.

Action pattern retrieval based on elastic action sequence correlation

If there is one video for indexing and a database of video to be indexed,the objective is to identify similar sequences which can be a sub-sequence of a long sequence in the database.A dynamic programming formula is proposed based on both the well-known edit distance in computer and information sciences and the dynamic time warping(DTW) method which cannot be used in action retrieval. The proposed method can be changed to edit distance or DTW through modifying parameters,so both methods can be regarded as special cases of the proposed method.The usefulness of the proposed method is demonstrated in action pattern retrieval,where performance comparable to the state-of-the-arts is obtained during empirical experiments on a number of challenging datasets including several standard databases.In particular,the true positive rate of the proposed method is better than the other two methods under the condition of the same false positive rate.