Servicenavigation

Project Details

Back

Projekt Print View

Transfer Learning for Human Activity Recognition in Logistics

Subject Area Human Factors, Ergonomics, Human-Machine Systems
Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
Production Systems, Operations Management, Quality Management and Factory Planning
Traffic and Transport Systems, Intelligent and Automated Traffic
Term from 2016 to 2023
Project identifier Deutsche Forschungsgemeinschaft (DFG) - Project number 316862460
 
Final Report Year 2025

Final Report Abstract

Warehousing plays a crucial role in logistics, with manual labor still dominant despite automation, in which order picking is costly and time-consuming, especially in high-wage regions. In particular, demand unpredictability complicates warehouse optimization, making efficient planning essential. Traditional time studies are expensive. Consequently, automated human activity prediction with sensors can improve efficiency in both warehouse optimization and human ergonomics. HAR classifies movements using signals from videos, motion capture, or on-body sensors like accelerometers and gyroscopes. On-body sensors are affordable, reliable, and non-invasive. HAR is widely applied in healthcare and smart homes. However, HAR faces challenges due to human motion variability, lack of standardized definitions, limited datasets, and class imbalances. Multi-channel time-series HAR is, in general, a challenging classification task. Human activities and movements show a large variation. Humans carry out in similar manner activities that are semantically very distinctive; conversely, they carry out similar activities in many different ways. Transfer learning reuses knowledge from a source domain to improve learning in a related target domain, addressing data limitations. Sharing feature representations is a type of transfer learning method, such as attribute-based representations. Attribute-based representations, such as human-activity taxonomies, support Transfer-based HAR. However, defining transferable manual activities in warehousing is difficult due to the complexity of human motion. Traditional HAR assumes fixed activity classes, but warehouse tasks are highly dynamic. As layouts and technologies evolve, HAR models must adapt continuously. Motivated by the success of semantic attributes for representing classes in the context image or scene classification and document analysis, human activities can be likewise represented by such a collection of semantic attributes. This project proposes a method for Transfer Learning for M-HAR using attribute representations and parameter transfer. Thus, this project recorded logistics (order-picking and packaging scenarios) HAR data from inertial sensors and Optical Motion Capture systems. Performed manual annotation of the recordings with labels consisting of both class and human motion attribute labels. The project’s results on the recorded dataset and on a realistic dataset provided by MOtionMiners GmbH prove the feasibility of transfer learning by testing on real industrial applications. In addition, various facets of annotation, such as manual annotation, semi-automated annotation and retrieval-based annotation, were experimented upon, with positive results pertaining to retrieval-based annotation being demonstrated. This project was tested on realistic data provided by MotionMiners GmbH showing the potential of the proposed method. An annotation tool with these possibilities was released as part of this project. Finally, the work motivated research in interesting directions such as human reidentification, dataset bias based on person idiosyncrasies and a tutorial or format for dataset creation.

Publications

 
 

Additional Information

© 2026 DFG Disclaimer / Copyright / Privacy Policy

Textvergrößerung und Kontrastanpassung