Realizing a Hand Gesture Recognition System for Indian Classical Dances Collection Visualization and Comparative Study of Features of a Bharatanatyam Mudra Dataset

Abstract

newline Indian classical dance forms use hand gestures(mudras), facial expressions, and body newlinemovements to communicate the intended meaning of the performance to the audience. newlineThe classical dances involves several intricate body movements which are to be studied newlinefor years under the expert supervision of gurus(teachers). It is this dedicated training newlinethat gives the performers perfection in their rendering. Modern times have witnessed a newlineshortage of such experts. There also exists a difficulty in comprehending the meaning of newlinemudras, facial expressions, and body movements of the classical dancer. This difficulty newlinein interpreting the classical dance gestures occurs especially for common people who newlinehave not been trained in classical dances. These two reasons have hindered the reachability newlineof the dances to new learners and dance enthusiasts all over the world leading newlineto a decline in the popularity of classical dance forms. Developing an e-learning tool newlinecan help in the revival of these dance forms by encouraging self-study and also aiding newlinea layman in understanding the performance of an expert classical dancer newlineAn exhaustive Bharatanatyam Mudra dataset was collected and made open. The newlinedataset consisted of 15,396 single hand gesture images of 29 classes and 13,035 double newlinehand gesture images of 21 classes. Dataset visualization was then performed as a part newlineof exploratory data analysis. newlineOn observing that the dataset was classifiable, the next step was obtaining the newlineoptimum feature vectors by observing their classification performance. Initially, raw newlinepixels and Histogram of Oriented Gradients(HOG) of the images were used as feature newlinevectors. These were classified using Support Vector Machines(SVM), Logistic newlineRegression(LR), Decision Tree(DT) and Random forest(RF) algorithms and their performances newlinecompared. Local feature descriptors such as SIFT(Scale Invariant Feature newlineTransform), SURF(Speeded Up Robust Features), ORB(Oriented Fast and Rotated newlineBrief), KAZE and KAZE Extended, A-KAZE (Accelerated KAZE), BRISK (Binary newlineRobust