Human Action Recognition

Theoretical (Analytical):

Practical (Implementation):

Literature Work:


In order to enable large-scale analysis of human behavior in team sports such as soccer, it is necessary to develop automated methods for the extraction of individual actions from existing data sources. The most common available source consists of video data (e.g., soccer).

Problem Statement

  • Meaningful action recognition of soccer video data (event annotation) is still performed manually & very time consuming
  • Existing methods often focus only player detection but not their performed actions
  • However, information about player action is fundamental for subsequent analysis


  • Research of existing methods for video-based human action recognition
  • Implementation of Deep Learning models for action recognition in soccer on unique & massive soccer data (600+ matches)
  • Evaluation of applicability of developed methods to different use-cases in real scenarios to provided gold standard datasets


  • Good to exceptional Python, OpenCV, and tensorflow/pytorch skills,
  • Basic knowledge in computer vision fundamentals
  • Advanced knowledge about deep learning models



  • Scope: Master
  • 3 Month Project, 6 Month Thesis
  • Start: immediately