Comparison of Recognition Accuracy on Dynamic Hand Gesture Using Feature Selection

Abstract

Dynamic Hand Gesture Recognition is carried out in various studies to read patterns. Various sensors can be used to capture dynamic hand movement patterns. The results of the initial reading are usually in the form of raw data that must pass initial processing. Advanced processing is carried out to obtain features that will be trained using various classifiers. The recognition process without feature selection activities will reduce the accuracy of pattern recognition during the classification process. Seeing the many shortcomings in the implementation of the initial data processing, this study will present some initial processing examples to produce features that are relatively good for the data training process. The method used is the Gaussian mixture model and the selection of predictors for the classification process. The sensor for recording dynamic hand movements used in this study is Leap-motion. There are three dynamic hand gestures were used in this study. The data used were 4609 coordinates spread in 30 features. The classifiers used are k-NN with Euclidean distance metric without feature selection process compare to the process with feature selection. The results obtained from this study are the availability of examples of feature selection models in the form of Gaussian mixtures and some accuracy of the results of processing comparison. The lowest prediction, without and with feature prediction, has a slightly range from 99.7% to 100%.