The Design of Human-Following Control

Fig.4. shows a human's walking velocity actually measured by ISpace. The estimated human position data by DIND is not the proper control input required for a mobile robot to follow a human since estimated position data contains errors in the form of calibration error and image processing error. When the heads of humans are measured by vision sensors in a short sampling period, their velocity and direction also change drastically.

図 4: Human walking as measured by ISpace

Both the estimated position of humans and human walking action are unstable. The direction and velocity of human walking sometimes changes suddenly and unpredictably. In this case, the mobile robot may fail to follow the human and lose the stable movement by a sudden change in the velocity input. When conventional control law is used, which is derived only from the distance between the human and the mobile robot, it is considered that a mobile robot cannot follow the human. In order to solve such problems, a special control method for human following is required.

The following need to be considered in the derivation of a new control law for human following: There are fundamental differences between a mobile robot and a human in the level of motion. A human is able to move freely by foot. However, since a differential wheel velocity type mobile robot of non-holonomic constraints is adopted as the agent of the ISpace in the current system, the robot cannot move as freely as does a human. To trace a human who walks freely, a control strategy that overcomes the limitations imposed by the non-holonomic constraints of the robot is needed. Solutions for establishing a control strategy for human following are considered as follows: One solution is that the control law absorbs the kinematic difference between the human and the mobile robot. Another solution is that the presumption based on the detailed analysis of the human's motion characteristic is included into the control law. In the latter solution, a detailed human model is needed for the human's motion analysis. In addition, since an intention of the human is closely related to the human's motion, it is difficult to presume the human's walking course. The first solution is adopted in this research.

In order to overcome the above problems, we propose a virtual spring model. The proposed control law is derived from the assumption that a human and a mobile robot are connected by a virtual spring. The input velocity to a mobile robot is generated on the basis of an elastic force of a virtual spring in this model. In the proposed control system, the virtual spring works as a low pass filter and absorbs adverse fluctuations. The proposed virtual spring model is able to absorb the gap between the motion of the human and that of the mobile robot. Since the point of application of the elastic force differs from the rotation center of the mobile robot, as shown in Fig.5, the non-holonomic restriction of the mobile robot is overcome.