Person recognition and person independent letter recognition in airwriting via motion sensors

Today, diversity in human-computer interaction is progressively increasing, enabling input data to be provided in smart environments through the recognition of both individuals' activities and self-identity. Recently, there has been a growing trend in utilizing gestures in the air, particularly as a tactile interface. This method of data input eliminates numerous practical limitations and emerges as a favorable option, especially for individuals with disabilities. This thesis focuses on the problems of letter recognition and person recognition for a system that allows text input by writing in the air as an alternative to touch-screen text entry. To address this objective, MPU6050 sensors (gyroscope and accelerometer) were integrated into an Arduino board and designed as a wearable device by placing them on a glove. The purpose of this glove is to convert hand movements into signals, thereby creating a data source for person and letter recognition. In this study, raw signal data corresponding to 29 Turkish letters were processed, resulting in the creation of two datasets (Dataset-1, Dataset-2) each containing feature data for each letter. Dataset-1 contains features generated by applying the Fourier transform to the raw data, while Dataset-2 contains features produced using the Wavelet transform on the raw data. Each dataset comprises 10401 letter samples collected from 15 different individuals. Both person and letter recognition utilized Decision Trees, Random Forest, K-Nearest Neighbors (KNN) algorithm, Support Vector Machine, Artificial Neural Networks, and Subspace KNN methods. All the classifiers were tested on both datasets using cross-validation. The test results, within the framework of the examined parameter values, indicate that the Fourier transform outperforms the Wavelet transform for both letter recognition and person identification. Moreover, the Subspace KNN classifier demonstrated greater effectiveness compared to other methods. This thesis offers specific contributions to the problem of person recognition in air-writing and serves as a guide for researchers working in the field of air-writing recognition.