Accelerated Sparse Representation for Human Activity Recognition

Human activity recognition using wearable sensors plays a significant role in many applications. How to accurately and quickly recognize various activities based on wearable sensors draws more and more attentions. This paper proposes an accelerated sparse representation classification method based on random projection and k-nearest neighbor for human activity recognition. Random projection is first applied to reduce the dimensionality of the activity signal sampled from each wearable sensor. To optimally reconstruct an activity test sample, some nearest neighbor training samples from a few near neighbor classes of the test sample are selected to form a reduced training sample set, based on which the test sample can be sparsely represented. And then the activity class of the test sample is determined by solving an L1 minimization problem. The effectiveness of our method is experimentally validated on an open Wearable Action Recognition Database by recognizing nine common human activities performed by 20 subjects. Our method achieves the highest average recognition rate (92.56%), which beats the traditional sparse representation classification method and the conventional near neighbor algorithm. Meanwhile, the runtime of our method is significantly reduced compared to the traditional sparse representation classification method.