Simple and Robust Automatic Detection and Recognition of Human Movement Patterns in Tasks of Different Complexity

Abstract

In many different research areas it is important to understand human behavior, e.g., in robotic learning or human-computer interaction. To learn new robotic behavior from human demonstrations, human movements need to be recognized to select which sequences should be transferred to a robotic system and which are already available to the system and therefore do not need to be learned. In interaction tasks, the current state of a human can be used by the system to react to the human in an appropriate way. Thus, the behavior of the human needs to be analyzed. To apply the identification and recognition of human behavior in different applications, it is of high interest that the used methods work autonomously with minimum user interference. This paper focuses on the analysis of human manipulation behavior in tasks of different complexity while keeping manual efforts low. By identifying characteristic movement patterns in the movement, human behaviors are decomposed into elementary building blocks using a fully automatic segmentation algorithm. With a simple k-Nearest Neighbor classification these identified movement sequences are assigned to known movement classes. To evaluate the presented approach, pick-and-place, ball-throwing, and lever-pulling movements were recorded with a motion tracking system. It is shown that the proposed method outperforms the widely used Hidden Markov Model-based classification. Especially in case of a small number of labeled training examples, which considerably minimizes manual efforts, our approach still has a high accuracy. For simple lever-pulling movements already one training example per class sufficed to achieve a classification accuracy of above $95\%$.