Date

2022-4-29

Author

Ersoy, Mustafa

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

500
views

326
downloads

Cite This

In this study, we propose an extendable, synthetic human pose estimation dataset named “Metupose”. Pose estimation aims to determine the pose of a person by detecting joints in an image or video. Dataset was created in Blender 3D software and with varying human objects and environment. It is also used to enhance the accuracy of pose estimation models in the literature. Metupose dataset contains 178000 images. Images have 1 to 4 people in it, where there is a total of 402000 people exist in these images. When we train different pose estimation models from the literature with our dataset, we observed an accuracy increase in all model/dataset cases. Our second contribution is the source code to create new images for the dataset. Although, Metupose contains large number of images for most of the applications, users may need to create their own custom dataset or want to increase the number of images. We provide original Blender 3D files and a simple configuration file so that users can create new dataset easily. Normally, creating a real-world dataset is time consuming and open to labelling errors. The advantage of our study is that datasets of desired sizes can be created from software, without any error, in an automized way. We finally, trained a pose estimation model with our Metupose dataset and integrated the trained model into Nvidia Jetson Nano that is equipped with a Raspberry Pi Camera and evaluated human tracking performance. Results indicate that single board computers offer a low-cost alternative to be used in human-robot interaction studies.

Subject Keywords

Pose estimation, Extendable dataset, Synthetic dataset, Human tracking

URI

https://hdl.handle.net/11511/97333

Collections

Graduate School of Natural and Applied Sciences, Thesis