Human activity recognition based on ELM using depth Images

Abstract

Human Activity Recognition (HAR) has gained considerable research interest in recent decades due to its vast applications especially in the fields of medicine, surveillance, human-machine interaction, gaming and entertainment. Feature extraction is a key step in HAR algorithms. However, at present most research is focused on common features such as spatial domain and frequency domain features. Such features lack context and are not comprehensive in nature. Unfortunately, building a comprehensive feature space of human activities is difficult due to the vastness and uncountable nature of human actions. This leads to the challenging problem of designing a HAR system that uses context-based feature extraction of human actions. In this work a comprehensive contextual feature space for human activity recognition is presented using depth image,the total number of fratures is 11. in classification aspect, extrem learning machine uses only a single iteration in the training stage to determine the output weights. extrem learning machine is extremely effective as it tends to achieve the global optimum compared to the traditional FNN learning methods which might get trapped in a local optimum. The drawback of ELM algorithm holds an infinite number of degrees of freedom for approximating a given data set. These infinite degrees of freedom are a consequence of the random nature of the weights assigned between the input and the hidden layer. A possible potential improvement in performance in this research can be achieved by assigning the weights based on an objective functionan optimization of the (ELM) using the meta-heuristic. Harmony Search Algorithm which is a part of meta-heustric and Tansig activation function which remove un needed hidden neuron are also presented in this work. The presented approach hence solves the problem of the infinite degree of freedom of the input weights as well as restricting the number of neurons in hidden layer, thus increasing the performance of the ELM algorithm. The optimized ELM algorithm is then used to perform the classification of the developed context based on feature space. The accuracy achieved was 100% during training and 94.95% during testing with throw action and 100% during training and 100% during testing without throw action. Gready optimization of the ELM with HSO has acehived an accuracy of 94.95%. Moreover, 60% of the features have achieved an accuracy of over 100%. Thus, the approach can be utilized to perform the human activity recognition for various purposes.