K-SVD based dictionary learning algorithms for accelerometer signal processing

Abstract

K-SVD is a very powerful dictionary learning algorithm which is used in signal processing and machine learning. The learned dictionary can then be used for classification problem. Human Activity Recognition is one of the very important classification-problem. HAR has various application such as Healthcare and Assistive Technologies, Ambient Assisted Living, Security and Surveillance, Fitness and Sports Tracking etc. Human Activity Recognition is a process wherein the activity performed by human is analyzed and classified into its respective classes. The dictionary learning process for the HAR problem is done by algorithms such as K-SVD, MOD (Method of Optimal Direction, Maximum Likelihood Method etc. In this thesis, we have extensively discussed about the K-SVD based dictionary learning algorithms. K-SVD based dictionary learning algorithms are variants of the original K-SVD algorithm such as Discriminative K-SVD, Approximate K-SVD, Block K-SVD, Parallel K-SVD, Structured K-SVD etc. Although K-SVD is a very powerful dictionary learning algorithm on its own, with ever increasing dictionary size for complex human activity recognition, the computational burden increases drastically and will consume more time to converge to a solution. Therefore, the algorithm must not only be powerful but also should have less computational burden. The variants of the original K-SVD algorithm aims to converge to a solution faster and improve efficiency as well. One of the variants is Discriminative K-SVD. As the name suggest, this algorithm helps in learning a dictionary with not only representational power but adds a discriminative power to the dictionary as well. The dictionary learned with a discriminative power is more suitable for the application in classification problem. This variant tends to increase the efficiency of the classification problem. Approximate K-SVD is another variant of the K-SVD based dictionary learning. This variant tends to decrease the computational burden by approximating the solution without calculating the actual solution. This helps the algorithm to be executed much faster. However, the accuracy of the dictionary learning process may get slightly lowered. The sparse coding is also an integral part of the dictionary learning process. There are various sparse coding algorithms such as OMP (Orthogonal Matching Pursuit), Basic Pursuit, Focal Underdetermined System Solver (FOCUSS) etc. In this thesis, we have discussed five different sparse coding algorithms. They are OMP, Batch OMP, POMP (Projection Based Orthogonal Matching Pursuit), OLS (Orthogonal Least Square) and LAOLS (Look Ahead Orthogonal Least Square). The K-SVD algorithm and its variant (D-KSVD and A-KSVD) is implemented with five different sparse coding algorithms to solve a bi-class classification problem of human activity recognition on 3 different datasets. All these has been covered in this thesis. The future scope on this topic has also been discussed in the conclusion.