Deep learning based architecture for humanoid pose (skeletal) estimation from partial depth data


Project no.: S-PD-24-29

Project description:

The aim of this project is to develop a novel occlusion-sensitive method for humanoid pose estimation and skeletal tracking from limited or impaired data. As a result, the trainee’s scientific qualification will be improved through practical scientific activities and the exchange of scientific ideas. The developed method will be based on deep learning architecture, but will also employ steps of signal preprocessing and feature extraction. A major challenge is to develop the occlusions-aware model. Since the model is aimed at siting persons, prior knowledge of kinematic constraints may be applied. The results will be valuable for several areas, such as telemedicine, education, and others. For example, the developed model can be implemented for systems requiring pose estimation of sitting people, such as posture correction, or telerehabilitation systems. The system would be tracking the movement of the upper limbs of a sitting person behind a table, in a wheelchair, or in other situations, when the lower part of a body is obstructed by the camera. The results of the project, such as the code of algorithms, and raw data will be open source, available to the scientific community, and it will meet the trends of this growing field.

Project funding:

Research Council of Lithuania (RCL), Projects of Postdoctoral fellowships funded by the state budget of the Republic of Lithuania

Period of project implementation: 2024-02-01 - 2026-01-31

Project coordinator: Kaunas University of Technology

Rytis Maskeliūnas

2024 - 2026

Department of Multimedia Engineering, Faculty of Informatics