Analyzing Behavioral Biometrics of Handwriting Using Myo Gesture Control Armband

Abstract

Through the last few decades, computer technology has gradually merged into our everyday lives. Computers and sensors are embedded in an increasing amount of household items, enabling us to monitor and remotely control our connected devices from apps on our smartphones. The technology interfaces are also evolving along with new technologies. Among the up and coming digital interfaces are wearable technology. The Myo gesture control armband (GCA) is an example of tools which aims to make the communication from computer to human more seamless and intuitive. The Myo GCA is a multi sensor armband containing 8 surface electromyography sensors which measure electrical activity originating from skeletal muscles in the upper forearm. It is also equipped with a 9-axis inertial measurement unit which can provide information on spatial arm movements of the users. Together these sensors enable its user to pass 6 configurable commands to a smart phone or Blue-tooth connected computer. In this thesis we explore the Myo armbands potential as a multi sensor for handwriting recognition. Data are sampled and manually extracted through a cumbersome time consuming process, using recorded video as a reference to the sampled Myo data. The subjects are given the task of writing 10 repetitions each, of the four capital letters: E, L, O, and R. A strong positive correlation between same class letters within subjects has been proven in all of the four sensor types, where the orientation data yields the highest correlation coefficient values, while the sEMG data yields the lowest. Statistical similarity between same class letters has been found through singular value decomposition, where again orientation data yields the highest values, while sEMG scores the lowest of all sensor types. In an attempt to cross subject classification though k-NN, with k = 1, k = 3, and k = 5, the 1-NN classifier yields a minimum success rate of 58\% across the four letters. This is considerably better that what we would expect from a random assignment of letter classes. In the last part of the results, a similarity search by DTW is attempted. This yield poor results, with a classification success rate of around 10$\%$ on average across letters.