Regresar

Real-Time Hand Gesture Recognition Using KNN-DTW and Leap Motion Controller

Abstract:

The advances in artificial intelligence have been boosting the development of applications in different fields. Some of these fields work in real-time and based on development and implementation in machine learning models. The problems presented by these fields of study involve pattern recognition, which consists of selecting a label and specify the instant of time that defines the gesture. In this context, the paper presents a specific model of real-time hand gesture recognition using the leap motion controller and machine learning algorithms. This model recognizes five static gestures. The gestures are open hand, fist, wave in, wave out, and pinch. The modules used to develop the model are data acquisition, pre-processing, feature extraction, classification, and postprocessing. The input signals to the model are the spatial positions and the direction. These signals are pre-processed by a translation and rotation matrix. Also, these are normalized and smoothed by a Butterworth filter. For the feature extraction module, we use the windows division technique. The model was tested with a window of 7 and 10 features. This signal is delivered to the KNN classifier. It tested the model with k = 1, and k = 3, and with DTW as a distance metric. We are varying the signal warping parameter in w = 1, 5, and w = 7. Finally, the model report 92.22% of classification accuracy and 77.64% of recognition accuracy in 287 ms.