Emotion is a natural, instinctive state of mind emanating from one's circumstances, mood or relationships with others. Emotion can be characterized primarily by the psycho-physiological expressions, biological reactions, body interaction and mental states. In social interaction, emotional component serves as an important element in communication, response and conveying information. Every day, the human body has evolved to perform sophisticated tasks to carry information about emotions. Recent years have seen a significant expansion in research on computational models of human emotional processes primarily in body interaction. However, most researchers fail to address the problem in preprocessing technologies and mainly rely on traditional methods to interpret emotion. Thus, this thesis aims to develop improved emotion recognition methods from human actions which includes the extraction of dynamic descriptors (distance, speed, magnitude of acceleration and magnitude of jerk) and statistical features (mean, maximum, minimum, standard deviation, median, log-energy, RMS and entropy) from joint position data, feature selection/reduction (Relief-F, fast correlation-based filter (FCBF), correlation feature selection (CFS), linear discriminant analysis (LDA) and principle component analysis (PCA), feature weighting methods (feature weighting based on fuzzy Cmean (FWFCM) and feature weighting based on binary encoded output (FWBEO)) and recognition of emotions using different classifiers (K-nearest neighbor (KNN), fuzzy K-nearest neighbor (FKNN) and probabilistic neural network (PNN)). After feature extraction, irrelevant/redundant features were removed using Relief-F, FCBF, CFS, LDA and PCA. Further, to reduce the higher degree of overlap among the relevant/non-redundant features, this thesis proposes FWFCM and FWBEO based feature weighting methods to enhance the discrimination ability of the features and also to minimize the degree of overlap among the features. Different emotion recognition experiments such as subject dependent, subject independent, gender dependent and gender independent were carried out. The performance measures such as overall accuracy and g-mean were considered for the evaluation of the classifiers. The experimental results demonstrate that the proposed feature weighting methods (FWFCM and FWBEO) are effective to classify emotion in human action with a maximum accuracy of 100%.
