Abstract: Abstract：It is very important for the development of human-computer interaction that how to accurately recognize various behaviors from the contact sensing data. Based on the fact that there is redundancy between different behavior data and the action of identifying behavior is often only at some key points, a Key-frame DynAmic sampling network KDAS for behavior recognition is proposed, which aims to mine key-frames representing behaviors, eliminate redundant frames of behaviors, and improve the discrimination of different behaviors from the data source. Then, the discrimination of their deep features is enhanced and the recognition accuracy is improved. First, a pre-processing module for behavior recognition based on BLSTM network is established, which is used to initialize the original perceived data state and obtain the initial behavior prediction result. Second, based on the initialization information of each data frame, a key frame selection network is established using BLSTM, through which the probability of each frame being selected is predicted and then the key frame is determined. Finally, the selected key frames are used for behavior prediction again. The two prediction results are used to form a utility function for the key frame selection network training. The experiment is conduct on three public data sets UCI HAR, Opportunity and UCI MHEALTH. The experimental results show that compared with several existing advanced behavior recognition models, KDAS can obtain higher behavior recognition accuracy.

Key words: Key words：contact sensing, behavior recognition, heterogeneous time series data, key fram, discriminative feature