DEVELOPMENT OF A MARKERLESS MOTION CAPTURE SYSTEM BASED ON OPENPOSE
In previous research, several models to predict the 3D Ground Reaction Forces of running had been created using LSTM (Long Short-Term Memory). The models take kinematic data obtained through motion capture as input. However, there were disadvantages to the currently developed method in ITB Biomec...
Saved in:
| Main Author: | |
|---|---|
| Format: | Theses |
| Language: | Indonesia |
| Subjects: | |
| Online Access: | https://digilib.itb.ac.id/gdl/view/70198 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | In previous research, several models to predict the 3D Ground Reaction Forces of
running had been created using LSTM (Long Short-Term Memory). The models
take kinematic data obtained through motion capture as input. However, there were
disadvantages to the currently developed method in ITB Biomechanics Lab such as
the need for a dark room, or the tendency of markers to shake while the subject is
moving. To overcome this, human pose estimation based on deep learning and
computer vision can be applied as a markerless motion capture system. There was
a huge leap in this field back in 2020, by a bottom-up approach model OpenPose.
This model can predict human pose in real-time even for multi person problem. But
recently, it was reported that OpenPose is not as accurate as marker-based or sensorbased
motion capture. Hence, to alleviate this inaccuracy, it is proposed to fine-tune
the original OpenPose with motion capture data so that it will learn and perform
better for motion capture task especially for running or gait. But to get the finetuning
done, the subject and marker should be clearly visible. Later, the frames will
be inpainted to hide the marker shown in the image leaving only the subject for the
input data. Lastly, after training it was discovered that the model has good accuracy
with the average of 17-pixel error. |
|---|