Recognition of Walking Motion Using Support Vector Machine

This paper presents a motion recognition method which combines Support Vector Machine (SVM) and state machine. We applied our method to the recognition of walking motions. We divide walking motion into five states (right leg up, right leg down, left leg up, left leg down, not walking). Based on subject’s posture that is acquired using a motion capture device, our method recognizes the subject’s current walking state as well as the subject’s walking speed. We use the velocity of primary body parts (hands, feet, and pelvis) as a feature vector. Based on a trained model, a SVM detects the subject’s current state. However, it is difficult to consider the subject’s previous states with SVM. Therefore, we introduced a state machine in which the subject’s current state is determined based on the previous state and the recognized state by the SVM. The walking speed is also computed from the state transition speed on the state machine. This paper also describes the results from some experiments that were made to evaluate the accuracy of our system.