Accelerometer-based gesture recognition using dynamic time warping and sparse representation

In this paper, we propose a new accelerometer-based gesture recognition system. In this system, the start and end of the data collection process is automatically determined by acceleration waveform. In pretreatment phase, we propose a waveform compensation algorithm to solve the problems caused by the amplitude range of the accelerometer and use the coordinate transformation theory to alleviate the angle offset. In training phase, we use dynamic time warping (DTW) and affinity propagation (AP) to extract clusters and exemplars. We implement sparse representation for gesture recognition and propose a modified variable sparsity adaptive matching pursuit (MVSAMP) algorithm for signal reconstruction. This algorithm is more adapted to the characteristics of gesture recognition. In the classification stage, a method of weighted residuals is applied to improve the resolution of the best classification. To test the system’s performance, a dictionary of 10 gestures is defined and a database consists of 3800 samples is created from 14 participants. Test results have shown that the proposed system achieves a good performance in a variety of experiments on Android platform.