基于深度学习和动态时间规整的人体运动检索

Human Motion Retrieval Based on Deep Learning and Dynamic Time Warping

随着计算机动画在各种应用中的日益普及,市场上出现了很多人体运动捕获设备,人们使用这些设备制作了大量的人体运动数据库。为了节约成本和高效地利用已有数据资源,提出了一种基于深度学习和动态时间规整相结合的人体运动检索方法。该方法包括两个主要阶段,在学习阶段,针对运动数据库中的运动序列,首先利用模糊聚类获取运动代表性帧,进而建立关键帧图像集合,然后应用深度神经网络学习关键帧图像集合,得到自动编码器,再应用自动编码器提取各个关键帧运动姿态的特征,建立运动特征数据库。在运动检索阶段,针对待查询运动序列,根据阶段1获取的自动编码器对每一关键帧图片提取特征,进而使用基于曼哈顿距离的动态规划方法计算待查询运动与数据库中运动的相似度,并根据相似度量值对检索结果进行排序。最后通过实验验证了该方法的有效性。

With the popularity of computer animation in various applications,human motion capture equipment is produced on the market,which can be used to produce lots of human motion databases.To reduce cost and utilize the existing data resources better,we propose a method of human motion retrieval based on depth learning and dynamic time warping.It consists of two main phases.In the learning,first of all,we obtain motion representative frame by the fuzzy clustering in viewof the sequence in the motion database and set up a collection of key frame images.Then we use the deep neural network to learn the collection of key frame images for the automatic encoder and following the automatic encoder to extract the feature of each key motion frame for establishment of the motion feature database.In the motion retrieval,the automatic encoder attained by former phase extracts the feature of each key frame motion image. We use the dynamic programming method based on Manhattan distance to calculate the similarity between the motion sequences queried and the motions in the database,and sort the search results according to similarity measures.Finally the experiment proves the effectiveness of the proposed method.