Labanotation Generation Based On BLSTM-CTC Network With Continuous Motion Capture Data

Labanotation is an effective way to represent and record human body movements. The existing methods of generating Labanotation can only complete recognition after segmentation, which are less suitable for real scenes. We proposed a novel method to process continuous motion capture data, which can simultaneously implement segmentation and recognition in this paper. Firstly, we extract the Line feature which contains the location and direction information. The Line feature fully reflects the relative spatial relationship between adjacent joint points, thus it shows good spatial characteristic. Secondly, especially for long time series, the Bidirectional Long Short Term Memory has good time characteristics and long-term memory function, which can learn the further history and future information on the timeline. Therefore, BLSTM can catch more important information with continuous motion capture data on the timeline. Thirdly, Connectionist Temporal Classification is applicable for the case when the length of the input and output sequence are not aligned. Our dataset happens to be the case that the input and output sequence are not aligned. Thus CTC is suitable for processing continuous motion sequences of variable length. According to the experiments, this method based on BLSTM-CTC shows good performance in simultaneously completing segmentation and recognition of the continuous motion sequences of different length. Compared with other methods, our experiment results achieved a better result, which demonstrates the effectiveness of our pro-posed method.