A Machine Learning Based Approach for Dynamic Hand Gesture Recognition in Human-Robot Interaction

Today's modern world is equipped with many smart robotic gadgets and devices to make our day-to-day jobs easier. These robotic devices need to capture human orders using different kinds of audio-visual inputs, hand gestures being one of them. Therefore, hand gesture recognition is an important step in human-robot interaction. The response of the robot directly depends on the time taken in classification of the performed gesture by the human. However, processing all the redundant frames in the visual input increases the response time of the device. In this paper, we propose a machine learning based approach to dynamic hand gesture recognition, by extracting only unique key-frames from the input visuals. To obtain them, firstly the hand skeleton landmarks are focused in order to obtain the important spatial features. The similarity in these landmark features is analyzed and used to determine the unique key-frames from each input video. The unique key-frames so decided by similarity detection are further input to SVM machine learning model to perform hand gesture classification, achieving a classification accuracy of 79.30% at 20 key-frames within a training time of 1241 sec, which is comparable to the one using all frames, taking a training time of 6455 sec. The obtained results have been further compared with regular feature extraction method using PCA, showing improved results of proposed method over PCA.