Motion Analysis

The Chair of Digital Signal Processing and Circuit Technology conducts research on technical assistance systems for marker-less motion analysis. One field of application is the monitoring of trainings in the therapy environment. The aim is to detect incorrectly executed movement sequences with the aid of a skeleton previously detected on the image data. Machine learning methods are used to detect incorrect movements. This involves learning a classifier with typical training errors images.

Fig 1: System Overview

In system design and development, the necessary invariance of the system against persons of different sizes and stature poses a particular challenge. The aim is therefore to minimize the configuration effort for new patients, whose skeletal data can nevertheless differ considerably from one another with the same movements. For this purpose, methods for normalizing skeletal data of different individuals are investigated.

Fig 2: Local and hierarchical coordinates

Furthermore, the Incremental Dynamic Time Warping (IDTW) approach has been extended to generate feature vectors for the classifier. It compares the currently executed exercise in real time against a reference measurement and determines the difference between the reference skeleton and the skeleton of the current frame for relevant joints.

Fig 3: Visualization of the extended principle of the IDTW

Further work includes the efficient extension of the existing exercise catalogue by new exercises.