Human Activity Recognition from Real-World Depth Images

Master Practical Training Project/ Master’s thesis

Status: available  taken (but if you would like to work on a similar topic, contact Irene Ballester)

Supervisors: Irene Ballester, Martin Kampel

Problem Statement

Human Activity Recognition (HAR) in computer vision, a pivotal area for healthcare, security, and robotics, often relies on privacy-invading RGB cameras. To enhance HAR accuracy while safeguarding privacy, this project employs deep neural networks (GNN, CNN, transformers) with point clouds or skeleton data extracted from real-world depth videos, targeting robust HAR in settings like long-term care facilities.


Advance HAR using depth videos for robust activity recognition in complex environments like care facilities while respecting privacy.


  1. Implement State-of-the-Art Skeleton Extraction Models from Depth Data: Research and implement leading-edge algorithms for precise skeleton extraction from depth data.
  2. Preprocess and Annotate Real-World Data: Prepare depth images from care facilities, and annotate activities to enhance dataset quality.
  3. Adapt Algorithms for Real-World Data: Modify models to handle noisy, diverse real-world data, including data augmentation and robust feature extraction.
  4. Develop HAR Models from Skeleton and Point Cloud Data: Design GNNs and CNNs specialized for HAR using skeleton data and point clouds, capturing temporal and spatial relationships.
  5. Comparison between models and report

Irene Ballester