An algorithmic approach for break static and dynamic gesture recognition

In this dissertation, a novel approach for recognizing static and dynamic hand gestures by analyzing the raw data streams generated by the sensors attached to the human hands is proposed. We utilize the concept of 'range of motion' in the movement and exploit this characteristic to analyze the acquired data for recognizing the static and dynamic signs. It is shown that since the relative 'range of motion' of each section of the hand involved in any gesture is a unique characteristic of that gesture, it provides a unique signature for that gesture across different users. We also define a signature for each dynamic sign based on the deviation from the predicted movement and show that this signature is independent from raw data and uniquely defines the gesture. Based on these observations, our approach for hand gesture recognition which addresses two major challenges is proposed: user-dependency and device-dependency. Furthermore, it is shown that our approach neither requires calibration nor involves training. Our approach then applied for recognizing multiple sets of data including ASL (American Sign Language) and AUSLAN (Australian Sign Language) signs and it is shown that these signs can be recognized with no training. Our preliminary experiments demonstrate more than 80% accuracy in sign recognition for the both data sets. A subset selection approach based on bio-mechanical characteristics is also proposed which provides a simple yet effective technique for Multivariate Time Series (MTS). This approach applied to recognize ASL static signs using Neural Network and Multi-Layer Framework Neural Network and it is shown that the same accuracy can be maintained by selecting just 50% of the generated MTS data.