PCA & HMM Based Arm Gesture Recognition Using Inertial Measurement Unit

This paper presents a novel arm gesture recognition approach that is capable of recognizing seven commonly used sequential arm gestures based upon the outputs from Inertial Measurement Unit (IMU) integrated with 3-D accelerometer and 3-D gyroscope. Unlike the traditional gesture recognition methods where the states in the gesture sequence are irrelevant, our proposed recognition system is intentionally designed to recognize the meaningful gesture sequence where each gesture state relates to the contiguous states which is applicable in the specific occasions such as the police directing the traffic and the arm-injured patients performing a set of arm gestures for effective rehabilitation. In the proposed arm gesture recognition system, the waveforms of the inertial outputs, i.e., 3-D accelerations and 3-D angular rates are automatically segmented for each arm gesture trace at first. Then we employ the Principal Component Analysis (PCA) - a computationally efficient feature selection method characteristic of compressing the inertial data and minimizing the influences of gesture variations. These selected features from PCA are compared with those standard features stored in pattern templates to acquire the gesture observation sequence that satisfy the Markov property. Finally, the Hidden Markov Model is applied in deducing the most likely arm gesture sequence. The experimental results show that our arm gesture classifier achieves up to 93% accuracy. By comparing with the other published recognition methods, our approach verifies the robustness and feasibility in arm gesture recognition using wearable MEMS sensors.