Erkennen von Haltungen

Type: Semester thesis
Student: Marianne Michel, Claudia Frischknecht
Advisor: Corinne Mattmann
Project: Backmanager

This study is situated in the context of the Backmanager Project and is based upon a series of measurements, recorded using a wearable system prototype. A total of 27 strain sensors were mounted on this garment, measuring at a frequency of 35Hz. The recorded series included nine test subjects performing 29 postures and 11 posture sequences.

In a first step the described measurement data was thoroughly analysed which included considering the specific sensor values of each posture. Different classification algorithms were then applied to this data, in order to find those classifiers which best satisfied the given requirements in terms of running time and percentage of correctly classified instances. The executed tests resulted in the selection of the Support Vector Machine (SMO) and Naive Bayes (NB) classifiers, which were therefore used for the ensuing analysis.

Subsequently, the robustness of the system was examined, so as to determine its behaviour in a noisy environment and the postures sensitivity with respect to sensor data variations. To this end, different aspects and levels of noise were added to the measurement data. The results were then visualized in form of confusion matrices. Pairs of postures, which are often mutually confused, were successfully detected by examining the symmetries in the respective confusion matrices. These postures were found to be characterized by being highly similar such as the two often confused postures “rotated to the right while sitting” and “body rotated to the right while standing”.

Another objective of this analysis considering robustness was to find certain postures, which are least sensitive to small data variations and were therefore often correctly classified, even if high levels of noise were added. The postures, which were identified as robust, were uniquely characterized by high elongation values for particular sensors that showed smaller values in other postures, such as the posture “Torso strongly bent forward, hands clasping ankles”.

A next task was to determine if a user's data can be correctly classified with a model built on other test subjects' data. The corresponding tests showed that a classifier trained on and applied to the posture data of a single subject performs significantly better (approx. 99% correctly classified instances) than one trained on user-independent measurement data (approx. 67% correctly classified instances).

A subsequent analysis regarding the classification of posture sequences, yielded similar insights. Movement recognition was achieved by applying a Hidden Markov Model algorithm. The observations were continuous and a HMM with Gaussian outputs was therefore demanded, in which the unknown parameters were estimated using the Baum-Welch method. Again, tests considering user-dependent as well as user-independent classification were conducted, where the latter showed a percentage of approx. 20% correctly classified instances -- a very unsatisfactory outcome compared to the approx. 85% of correctly classified instances reached in the user-dependent tests.

The significantly worse results of the unseen-user classification are assumed to be caused by the differences in the test subjects' performance of postures and posture sequences, their differing stature as well as their individual physical abilities.

The presented evaluation led to the insight, that the considered wearable garment prototype is primarily adequate for applications in which the treated postures are highly distinct. For the classification of similar postures one should, if possible, try to find a better positioning of the strain sensors so that each posture possesses pronounced characteristic elongation values. For the case, that a better classification result cannot be reached by altering the sensor positioning, the usage of different sensor types might also deserve consideration.

The gained insights show that, if these results cannot be improved by further analysis, the model should be trained on the end-user, in order to achieve sufficiently good results.

The consolidated findings attained in this thesis establish a foundation for the implementation of the desired garment and provide an outlook for further studies.