Hello people,
I am using a machine learning approach to be able to detect gestures using a 9-axis MEMS IMU sensor.
I am collecting the raw data from accelerometer, gyro and magnetometers for the appropriate gestures made in the air using the IMU sensor, compute the features, train the model (Eg. SVM model from Scikit learn), save it and use it later for predicting the gestures.
I am able to record the data, train and able to detect the gestures successfully, but only in one orientation. So, the model could not predict if I draw a '0' in the air holding the IMU sensor rotated 90 degrees clockwise for example, or any other orientation for that matter because It has not been trained in that orientation. So please shed some light so as to be able to detect the gestures irrespective of the orientation I am holding the IMU sensor. Please ask questions if this info is insufficient. Thanks in advance :)
Thats an interesting project Nikhilesh. And congrats for coming thus far.
The gestures in 3D space vary with respect to the placement and orientation of recipient sensors. In your case, they depend greatly on the alignment with respect to gravity.
So, even though your system is able to detect all gestures in one orientation, it is highly unlikely that it will be able to do the same when placed in perpendicular. I am afraid, you will have to train your system for all orientations.
Dear Nikhilesh Karanam ,
I agree, partially, with Nidhi Gupta . Probably, you will need to make some data augmentation or capturing more data samples to train your system for all orientations. However, did you try to apply small temporal sliding windows, before learning stage? I believe that small temporal windows (i.e., 1/5 of the original sample) could help the system more robust to orientation.
Please, consider to see this paper:
Conference Paper An analysis of eating activities for automatic food type recognition
In this paper, the authors propose to split each sample into action units. This process is similar to my proposed idea.
I hope this helps you.
- Badges
- Science topic