Human Activity Recognition based on Compressed Sensing

Fast and accurately recognizing human activities is a major challenge in wireless body sensor networks. In this paper, with the help of wearable sensors, the recognition of human activities is well solved from a compressed sensing perspective. Our proposed method first compresses the original sensing data collected using wearable sensors by applying random projections on each sensor node before the data are transmitted to the central node. On the central node, each activity test sample can be approximately represented as a linear combination of all the activity samples in the training set. Then the L1 minimization method is applied to derive the sparse representation of the test sample. Residuals between the activity test sample and its corresponding sparse representation form with only nonzero entries in each activity class are calculated. The activity test sample is finally classified into the activity class with the smallest residual. Our sparse representation based activity recognition method is evaluated on a real-world dataset and its experimental results are also compared with those of Support Vector Machine and Hidden Markov Model. Outcomes of the numerical experiments show that our proposed method achieves the highest recognition rate among the three methods, which verifies the effectiveness of our method and demonstrates that sparse representation classification method based on random projections is promising in recognizing human activities using wearable sensors.