Recognition of Human Activities Using Fast and Adaptive Sparse Representation Based on Wearable Sensors

Methods based on sparse representation have achieved success in human activity recognition. However, these methods, either emphasize too much on the sparsity but neglect the correlation in the activity training set, or underline the correlation but ignore the discriminative ability of the sparsity. Besides, the recognition speed of these methods is not fast enough. This paper proposes a fast and adaptive sparse representation classification method for human activity recognition, taking both correlation and sparsity into consideration. Random projection is first used to reduce the dimensionality of the activity signals. And then a simplified near neighbor algorithm is proposed to quickly and optimally reduce the training set. Next, the class of the test sample is determined by solving an adaptively joint L1-norm and L2-norm minimization problem using our pro-posed Alternating Direction Optimization Method. The effectiveness of our method is finally validated on an open Wearable Action Recognition Database. Results demonstrate that our method beats the traditional sparse representation classification method and the conventional near neighbor algorithm in accuracy, and the runtime of our method is also much less than the traditional sparse representation classification method.