Analysis of derived features for the motion classification of a passive lower limb exoskeleton
CovenanteeKarlsruher Institut für Technologie
Document typeMaster thesis
Rights accessOpen Access
Analysis of Derived Features for the Motion Classification of a PassiveLowerLimbExoskeleton The recognition of human motion intentions is a fundamental requirement to control efficiently an exoskeleton system. The exoskeleton control can be enhanced or subsequent motions can be predicted, if the current intended motion is known. At H2T research has been carried out with a classification system based on Hidden Markov Models (HMMs) to classify the multi-modal sensor data acquired from a unilateral passive lower-limb exoskeleton. The training data is formed of force vectors, linear accelerations and Euler angles provided by 7 3D-force sensors and 3 IMUs. The recordings consist of data of 10 subjects performing 14 different types of daily activities, each one carried out 10 times. This master thesis attempts to improve the motion classification by using physical meaningful derived features from the raw data aforementioned. The knee vector moment and the knee and ankle joint angles, which respectively give a kinematic and dynamic description of a motion, were the derived features considered. Firstly, these new features are analysed to study their patterns and the resemblance of the data among different subjects is quantified in order to check their consistency. Afterwards, the derived features are evaluated in the motion classification system to check their performance. Various configurations of the classifier were tested including different preprocessors of the data employed and the structure of the HMMs used to represent each motion. Some setups combining derived features and raw data led to good results (e.g. norm of the moment vector and IMUs got 89.39% of accuracy), but did not improve the best results of previous works (e.g. 2 IMUs and 1 Force Sensor got 90.73% of accuracy). Although the classification results are not improved, it is proved that these derived features are a good representation of their primary features and a suitable option if a dimensional reduction of the data is pursued. At the end, possible directions of improvement are suggested to improve the motion classification concerning the results obtained along the thesis.