A Novel Approach towards Pattern and Speed Invariant Holistic Analysis of Dynamic Gesture Recognition System

This study focused to develop a user-friendly gesture recognition system with multiple speed and gesturing pattern. The differences in speed and gesturing pattern by different users immensely affect the accuracy of the gesture system if the features are spatiotemporal in nature. Existing trajectory features in literature are spatiotemporal in nature. The trajectory data represents the gesture as an ordered sequence of direction in 2D retaining its temporal information. Such systems are not user-convenient and restricted to limited applications. Therefore, we proposed a new approach of feature extraction from the dynamic gestures which are not trajectory-based and yet pattern and speed invariant with minimum computational complexity. In the proposed system, after the gesture is smoothened, it is represented as a binary image unlike a 2D sequence of trajectory points in the Euclidean space. The image is sub-divided into local instances to extract magnitude of Zernike moment and histogram features from each instance. The study is carried out on 16 keyboard characters from literature which are highly likely to have different patterns from different users. The proposed approach improved the accuracy by 8.65% (histogram + ANN model) and 2.05% (Zernike moment + ELM model) than existing spatial trajectory based system for pattern variation.