Abstract
This paper introduces a new feature selection approach for human activity recognition systems to accurately recognize the human activities. We proposed normalized mutual information-based feature selection (NMIFS) method that will select good features extracted from numerous existing feature extraction techniques. The proposed method is an extension of the max-relevance and min-redundancy method. The ability of this method is to combine the strengths of different extraction techniques. However, the selection process might be influenced because of the inequality among the feature's classification power and the feature's redundancy. To escape this influenced selection, we normalize both terms by the proposed feature independent upper bound of the mutual information function. Moreover, we exploit the curvelet transform for feature extraction, and linear discriminant analysis for reduction of feature space. Moreover, we use the hidden Markov model for activity recognition based on our proposed method of feature selection. Finally, by using the benchmark datasets such as KTH and Weizmann datasets, we experimentally compare the proposed scheme with state-of-the-art. Simulation results show that the proposed scheme is not only more accurate for some datasets, but outperforms competing method by weighted average accuracy 98%.