Abstract-More and more researchers devote themselves to system development of life support for elders such as health condition monitoring, appliances control, emergency call, and so on. Activity recognition using wearable devices is one of the ways of system implementation. In this paper, a system of gesture recognition is designed and implemented based on a tri-axis accelerometer. Six pairs of one-stroke finger gestures are defined according to the sensing characteristic of the accelerometer. Acceleration data was collected from 20 subjects under the supervision by researchers. Gesture information was segmented automatically from successive time signal. Mean, standard deviation, and the relationship of position sequence of wave crest and trough are extracted as features for training and testing of gestures classifying. Several classifiers were selected for gestures recognition. Results show the KNN classifier got the best recognizing result with an average accuracy of 91.58%. The recognition rate is not influenced by which hand being used to perform the gestures.