Manipulative Action Recognition for Human-Robot Interaction

摘要

Recently, human-robot interaction is receiving more and more interest in the robotics as well as in the computer vision research community. From the robotics perspective, robots that cooperate with humans are an interesting application field that is expected to have a high future market potential. A couple of global and also mid-sized companies have come up with quite sophisticated robotic platforms that are designed for human-robot interaction. The ultimate goal is to place some robotic assistant or companion in the regular home environment of people, who would be able to communicate with the robot in a human-like fashion. As a consequence, the "hearing" as well as the "seeing" --as the most prominent and equally important modalities --are becoming major research issues. From the computer vision perspective, robot perception is more than an interesting application field. During the last decades, we can note a shift from solving isolated vision problems to modeling visual processing as an integral connected component in a cognitive system. This change in perspective pays tribute to important aspects of understanding dynamic visual scenes, such as attention, domain and task knowledge, spatio-temporal context as well as a functional view of object categorization. The visual recognition of human actions is in the center of all these aspects and provides a bridge for a non-verbal as well as verbal communication between a human and the robot, which both are highly ambiguous. It enables the robot's anticipation of human actions leading to a pro-active robot behavior especially in passive, more observational situations. Furthermore, it draws attention to manipulated objects or places, embeds objects in functional as well as task contexts, and focuses on the spatio-temporal dynamics in the scene.