Collection and Labelling of Training Data for AI Based Fall Detection in Experimentation Facilities

Livy Care, developed by HUM Systems from Berlin, offers advanced AI-driven solutions to enhance the quality of care for the elderly and alleviate the workload on caregivers. Their innovative technology includes smart sensors and telecare systems designed to improve safety and efficiency in both residential and home care settings. By integrating intelligent assistance, Livy Care aims to modernise traditional care practices, ensuring better safety and comfort for care recipients while providing peace of mind for their families.

HUM Systems GmbH developed a radar-based fall detection system, which can be installed in the home of patients who are living on their own and are at risk of falling. The system sends an alarm message when it detects a person lying horizontally, except when they are in bed, when the system should not send an alarm. In order for the solution to act in this manner, the bed had to manually be marked as a “safe zone” after installation of the system.

However, installing this system in hospitals and nursing homes introduces a new challenge, as beds in these settings are frequently moved due to transfers to different units or diagnostic facilities. Consequently, the fall detection system must incorporate an automatic "safe zone" detection to adapt to the dynamic environments of healthcare facilities.

To incorporate an automatic "safe zone" detection, HUM Systems aimed to train an AI algorithm for automatic bed detection and segmentation, which required a diverse set of training data. Since no existing database met these requirements, the CITAH team captured about 700 training images in the testing and experimentation facilities at the OFFIS Institute for Information Technology using HUM Systems’ fall detection system. 

Initially, the CITAH team assessed the requirements of the training dataset based on the experience and knowledge provided by HUM Systems. Collaboratively, it was evaluated which factors need to be varied when recording the training data set. The training images were captured within two different real-world laboratories: one resembling a fully functional apartment; the other resembling an intensive care unit at a hospital.

The resulting images of the training dataset systematically varied in lighting conditions, bed positions, camera position, and the presence of people, blankets, or pillows. As an additional service, the CITAH team meticulously labelled the data, enabling effective training of the AI algorithm on the labelled images.

The training data have now been used to train an AI algorithm, which automatically detects and segments the patient’s bed as “safe zone”. This new feature of HUM Systems’ fall detection system saves valuable time which would have been spent on manually marking "save zones" and it circumvents human error when forgetting to update the “safe zone” after a bed has been moved.

As the automatic “safe zone” detection was particularly requested by the end-users this new feature significantly improves the HUM Systems’ fall detection system. In fact, this new feature makes the fall detection more suitable and attractive for end-users in hospitals and inpatient care facilities, thus, increasing HUM Systems’ customer base.