Inertial sensor based recognition of 3-D character gestures with an ensemble classifiers

We present a 3-D input medium based on inertial sensors for on-line character recognition and an ensemble classification scheme for the recognition task. The system allows user to write a character in the air as a gesture, with a sensor-embedded device held in hand. The kinds of sensors used are 3-axis accelerometer and 3-axis gyroscope generating acceleration and angular velocity signals respectively. For character recognition, we used the technique of FDA (Fisher discriminant analysis). We tried different combinations of sensor signals to test the recognition performance. It is also possible to estimate a 2-D handwriting trajectory from the sensor signals. The best recognition rate of 93.23%, in case we use only raw sensor signals, was attained when all 6 sensor signals were combined. The recognition rate of 92.22% was reached if the estimated trajectory was used as input. Finally we tested the ensemble method and the generalization rate of 95.04% was attained on the ensemble recognizer consisting of 3 FDA recognizers based on acceleration-only, angular-velocity-only and handwriting trajectory respectively.