面向Kinect骨骼运动数据优化的堆叠双向循环自编码器

Stacked bidirectional recurrent autoencoder for Kinect skeleton motion data refinement

深度相机Kinect获取的人体骨骼运动数据含有大量噪声并且骨骼节点较少,数据细节层次较低。针对该问题,文章提出一种用于优化Kinect骨骼运动数据的网络,该网络由6个双向循环自编码器堆叠构成,通过堆叠结构提高数据的平滑自然性,并在训练阶段利用隐变量约束确保骨骼运动数据细节层次提高后仍具有合理的骨骼结构。在运行阶段,采用滑窗处理方式获得长序列的优化结果。实验结果表明,该网络得到的优化后数据具有更好的平滑性并能保持更为合理的骨骼结构,能够达到用低精度Kinect设备获取高精度动捕数据的目标。

The human skeleton motion data obtained by Kinect contains a lot of noise,few bone nodes and a low level of data details.To solve this problem,a network consisting of a stack of six bidirectional recurrent autoencoders is proposed to refine Kinect skeleton motion data.The smoothness and naturalness of the data are improved benefiting from the stack structure.Owing to the hidden variable constraints during the training phase,the level of motion data details is improved and the reasonable skeleton structure is guaranteed.In the running phase,the sliding window processing method is used to refine long-sequence data.Experiments show that the refined data obtained by the network has better smoothness and can maintain a more reasonable skeleton structure,which can achieve the goal of acquiring high-precision motion capture data from low-precision Kinect equipment.