EVALUASI KINERJA MODEL ESTIMASI POSE BERBASIS KECERDASAN BUATAN PADA SISTEM TANGKAP GERAK MONOKULER
Advancements in artificial intelligence technology in the field of computer vision have brought forth various pose estimation models widely accessible on the internet. These models are capable of automatically and accurately recognizing and tracking the positions, orientations, and movements of o...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/75669 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | Advancements in artificial intelligence technology in the field of computer vision have brought
forth various pose estimation models widely accessible on the internet. These models are
capable of automatically and accurately recognizing and tracking the positions, orientations,
and movements of objects in images or videos. As a result, various sectors, including sensing,
training, face recognition, and virtual reality, have experienced remarkable transformations
thanks to the utilization of pose estimation technology.
Evaluating the performance of pose estimation models is a crucial step to identify their
strengths and weaknesses in various usage scenarios. By testing the models in representative
and realistic scenarios, we can assess how reliable they are and identify areas that require
improvement. The pose estimation models used are BlazePose, MoveNet, and PoseNet, and
they will be evaluated using Percentage of Correct Parts (PCP) and Percentage of Detected
Joints (PDJ). In this pose estimation model, three experiments will be conducted to assess how
well the models can handle different cases. The experiments include finding the maximum
detectable distance of the models, rotation tests from three different camera perspectives, and
occlusion experiments. From these three experiments, MediaPipe BlazePose could detect poses
at a maximum distance of 10 meters. BlazePose also performed well in detecting poses in
rotation cases, achieving PCP and PDJ values of 91.67% for the center and left cameras and
100% for the right camera when rotated up to 70 degrees. Moreover, BlazePose successfully
handled the occlusion cases in this experiment, with a PCP value of 58.33% for cases 1 and 2,
and 0% for case 3. The PDJ values were 75% for case 1, 66.67% for case 2, and 0% for case
3. |
|---|