Jörg Moldenhauer

Analysis of Human Motion for Humanoid Robots

A great challange in robotics is to make robots more like humans. One important aspect is to make robots move like humans and recognize their motions. Both tasks are based on human motion trajectories and require a proper modelling. To accomplish these tasks, we acquire data from complex motions like setting the table, pouring water into a cup, or stirring the content of the cup. The objectives of our studies are to identify the subject doing the motion and to detect slight changes in motion constraints with automatic classification methods. For that purpose, we present reliable methods based on Elman networks and hidden Markov models. We develop the model parameters and compare the performance of the methods, especially, the classification results and the suitability for the classification tasks.

Classification of phases in human motions by neural networks and hidden Markov models

A proper modeling of human motions plays a crucial rule for many motion processing tasks. In particular, models for the automatic classification of elementary motion phases are highly important for the interaction between man and machine. In this work, we present different approaches for this modeling task based on neural networks and hidden Markov models. Both approaches yield reliable classification results. We show that even simple instances of the models work well if proper motion features are determined. A comparison of the different approaches shows the reasons for this behavior and leads to essential consequences for further modeling approaches

Similarities and varieties in human motion trajectories of predefined grasping and disposing movements

Human movement is characterized by smoothed trajectories and individual solutions relating to joint angle configuration. For the transformation of human movement characteristics into humanoid robots knowledge about basic strategies and the nature of human motion is necessary. The objective of this study is the inter-individual and intra-individual comparison of movement trajectories and time series of joint angles for several predefined object manipulations. For that purpose the kinematics of arm, hand and finger motion was measured via 3D video motion capturing and analyzed. Graphs of object trajectories, joint angles and finger tip distances (grasp state) were displayed for an example motion: moving to a cuboid, grasping the object and putting it to a heightened target position. The varieties in motion of different subjects and within repeated motion cycles of one and the same subject are presented. Besides this we show, how test subjects with different body height solve the motion task.

Context Orientated Motion Generation: A New Scheme for Humanoid Robot Control

There are two quite different kinds of approaches to the generation of motions for a humanoid robot. The first one optimizes for the robots pose, i.e. the movement should look similar to human motions (Zordan and Hodgins, 1999), (Dasgupta and Nakamura, 1999), (Riley et al., 2003), (Kuffner et al., 2003), (Erol et al., 2003), and (Ijspeert et al., 2002). The second one attempts to make the robot perform a given task and is thus focused on the accurate movements of the robots end effectors (Asfour et al., 2000) and (Yigit et al., 2003). In this work, we present a completely new scheme for the motion generation of a humanoid robot called context oriented motion generation. This scheme incorporates the pose oriented approach and the task oriented approach. It is based on a classical trajectory generator and a new context specific motion classifier developed by our group. The trajectory generator creates a set of trajectories and thereby ensures that the given motion task is accomplished by all trajectories. Afterwards, the motion classifier evaluates this set of motions with respect to the given context and selects the optimal trajectory

Composition of complex motion models from elementary human motions

An appraisal of human motions and particular motion phases is essential for a good interaction between a human and a humanoid robot. We present a new method for the analysis of human motions and the classification of motion phases. The method allows an automatic composition of a motion model for a complex motion from several elementary models. The elementary models can be retrieved from a motion catalogue according to the requirements of a current motion processing task. The method is based on the analysis of the hidden states in a complex HMM and considers the context of all elementary phases in an entire motion sequence. The analysis of motion phases with the new model is computationally more efficient and yields better recognition rates than conventional motion analysis with HMMs and winner-takes-all strategy

Zielgerichtete Aufbereitung und Visualisierung dreidimensionaler medizinischer Ultraschallbilddaten

Es werden Aufbereitungsmethoden fur 3D-Ultra-schall(US)-Aufnahmen vorgestellt, die auf eine gute Darstellung diagnostisch wichtiger Strukturen bei Volumenvisualisierungen ausgerichtet sind. Die Aufbereitung basiert auf einer zielgerichteten Verwendung von Calderon-Zygmund-Operatoren, wobei die Eigenschaften des US-Bildgebungsprozesses maszliggeblich berucksichtigt werden. Anhand von Volumenvisualisierungen fur US-Aufnahmen aus klinischen Untersuchungen wird die Wirksamkeit der Methodik verdeutlicht.

Merkmalsinduzierte Aufbereitung medizinischer Ultraschallbilddaten

Fur die Aufbereitung von 3D-Ultraschalldaten werden Verfahren entwickelt, die eine robuste Trennung einzelner Objektbereiche bewirken. Dies erfolgt durch eine approximative Rekonstruktion lokaler Gewebewerte und eine darauf aufbauende merkmalsgesteuerte adaptive Filterung. Die Wirkung der Algorithmen wird anhand von Volumenvisualisierungen klinischer Testdatensatzen demonstriert.

Aufbereitung medizinischer Bilddaten

Zusammenfassung. Medizinische Bilddaten durchlaufen von der Datenakquisition bis zur Diagnose beziehungsweise Operationsplanung die verschiedensten Verarbeitungsschritte. Die frühzeitige Kompensation auftretender Störungen in dieser Verarbeitungskette bietet den anschließend angewandten Algorithmen deutlich verbessertes Datenmaterial. In diesem Zusammenhang werden Verfahren zur Kompensation von Bewegungs- und Metallartefakten in tomographischen Aufnahmen vorgestellt, sowie neue Methoden der modellbasierten Registrierung von Datensätzen unterschiedlicher Modalitäten präsentiert.Abstract. Medical images pass several data processing steps from data acquisition to diagnosis and operation planing. Artifacts which are introduced in this data processing chain should be compensated as early as possible to yield better data sets for the algorithms subsequently performed. In this regard we present new methods for compensation of metal and motion artifacts in computerized tomography and describe a new approach to model based image registration of different image modalities.

Waveletbasierte Visualisierung von Ultraschall-Volumendaten

In dieser Arbeit1 wird die Notwendigkeit fur die Entwicklung spezieller Verfahren zur Visualisierung von Ultraschall-Volumendaten erlautert. Daraus folgernd wird ein neues waveletbasiertes Visualisierungsverfahren vorgestellt. Das Hauptaugenmerk gilt dabei der Konstruktion einer sogenannten Transferfunktion zur Berechnung von Voxeltransparenzwerten. Weiterhin wird in die Berechnung der Transferfunktion eine Rauschreduktion integriert. Abschliesend werden anhand von Beispielbetrachtungen die Visualisierungsergebnisse untersucht.