Layered Architecture for the Control of Micro Saccadic Tracking of a Stereo Camera Head

The paper describes a 3-layered architecture for the control of the stereoscopic eye-saccade system of a stereo-camera head 1 mounted on an autonomous vehicle. The 0-level is a proportional feedback controller providing a microsaccadic 2 control for eye movements enabling the head to foveate and track targets but requiring iteration through the vision system with the attendant computational overhead. The 1-level provides the feedforward inverse kinematics for saccadic eye movements allowing a ballistic movement to replace the 0-level control loop. The training data is provided by the feedback error signal from the 0-level controller. The 2-level is an adaptive lattice filter which is used to track moving targets. The filter is 'trained' using vision error-feedback from previous saccades. The filter learns to predict the future target position in the next image. This is used by the inverse kinematics module to generate the eye movement commands for the appropriate predictive saccade. 1 The stereo camera rig used for this work comprises a 3-link kinematic chain, whose degrees of freedom are rotations around the following axes: i) Pan: a vertical axis corresponding to the 'neck'; ii) Tilt: an axis at right angles to the neck; and iii) Verge: each camera ('eye') can rotate independently around an axis at right angles to the tilt axis. The rig has been constructed so that the centres of rotation of the tilt and pan links coincide, and the centres of rotation of left and right verge and the tilt links coincide. The length of the tilt link is approximately 12.5 cm for each eye (i.e. the head is about 25 cm wide); the length of the verge link (i.e. approximately how far the centre of rotation is from the focal centre of the camera) is 5cm so that tilting the eye also produces a small translation. It is also of note that the right camera has been mounted with a 5 degree heterophoria and about 2.5 degrees of cyclotorsion. Stepper motors control the head and give a maximum saccade velocity of 50 degrees/second. Microsaccades are generally used to refer to the very small saccades which, if they have any function at all, may be used to correct errors arising from drift during fixation of a stationary target (Carpenter, 1988). We use the term microsaccadic tracking to describe a form of tracking which uses small vergence saccades (ranging in size from a few minutes of arc to two degrees), characterised by a fast movement stage, followed by a 80ms image capture stage during which the eyes remain stationary. In humans, this form of tracking may not normally occur in isolation but seems to be an important component of pursuit movements (Carpenter, 1988, page 55). BMVC 1992 doi:10.5244/C.6.39