Figure 3:
a), c) and e) Idealised behaviour of fixation error signal. b), d) and
f) Experimental behaviour of viewpoint error signal.
The experimental setup consists of a JVC colour CCD camera with an Ernitec controllable lens, attached to the end-effector of a PUMA762 robot arm (see figure 4 a). The camera can be controlled as if it were independent of the arm by rotating it about and translating it along any of its three axes.
Figure 4:
a) Robot/camera system. b) Target objects. c) Experimental scene.
The objects used as targets were the Halloween masks as shown in figure 4 b. These masks were chosen for a number of reasons. Their relative complexity, arbitrary shape and 3-dimensional nature makes them more realistic objects for an unconstrained environment than, say, a blocksworld scenario. Of course, their distinctive colours makes them more suitable for the current system which is only, at present, extracting colour information. For each of the movement categories described in section 5 the robot was moved accordingly, relative to a target Halloween mask. At each fixation point the correlation measure from section 4.3 (error signal) was recorded. The fixation grid experiment was equivalent to fixating an even grid of points over the scene in figure 4 c. The results are shown in figure 3 . Due to practical difficulties of determining a long line of sight along which the camera could move without violating safety protection, the camera was kept stationary and the camera zoom varied for the 'distance to object' category (see figure 3 d). The experimental results compare favourably with the predicted behaviour shown on the left hand side of figure 3 , though the fixation grid does show unevenness at the extremities of the scene. The critical issue being that in each plot there is a gradual progression to a minimum point which a controller could use to position the camera appropriately.
Rupert J Young