Objective

Learn to use dynamic time warping (DTW) to recognize a time-based signal--in this case sensor data describing a gesture--among a group of pre-defined and learned gestures.

Steps and observations

Begin by setting up a mobile phone to send accelerometer data to a computer using UDP over a wireless network.

1. Click on "clear" to clear the SVM model
2. Choose the gesture number you wish to learn by clicking on the corresponding squre GUI
3. Hold your mobile phone in the orientation the is the begining point for the gesture; turn ON "learn" and perform the gesture; then turn OFF "learn"; repeat steps 2-3 for two more gestures.
4. Click on "train" to create the DTW model
5. Hold your mobile phone in the orientation that is the begining point of a gesture; turn ON "classify" and perform the gesture; look at the output of "ml.dtw" for the classification result; repeat with other gestures.

Comments

• How else can this technique be used?
• How many gestures can you recognize with this technique?
• During classificaiton, forcing the user to indicate the start and stop of a gesture is a major hinderence (requires an additional controller/sensor, forces the user to be partly responsible for the gesture recognition). Can you devise a system where the start/stop action in the classification stage can be automated?