Active View Point Control for Detection of Dynamic Targets in Human-Robot Interaction
Object detection is a fundamental task for robots, and is particularly challenging in the dynamic setting of humanrobot interaction because of the user’s movements. In this work, we put forward a practical viewpoint control mechanism for object detection in human-robot interaction. We not only consider constraints from vision but also incorporate the low level robot kinematics to guarantee reachability of the desired viewpoint. Using a linear time cost score function, our system selects very fast the viewpoints, which in turn leads to a smooth user interaction experience. We also provide a method to learn from human demonstration the score function weights to fit the task’s preference.