Active Visual Observer Integration of Visual Processes for Control of FixationArchitecture for controlling an agile cameraThe basic ideaThe basic system layoutThe attention mechanismsStandard control systemDesigning the ArchitectureLayout of Aalborg headCyclopean representationControlArchitecture for an equal eye dominance control scheme.Notes to the figureCompleting the ArchitectureThe modified cyclopean control architectureAn Experimental SystemFixation distance for combined disparity and accomodation controlAttention selectionThe receptive fieldsAnother Active Visual ObserverMajor parts of the current systemSystem descriptionThe system implementation schema (the diamond indicates one frame delay in the feedback)Motion detection schemaThe AlgorithmTarget segmentationThe target segmentationDisparity segmenttaionDisparity selectionIntegrationThe experimental platformReal time pursuitSlide 32Slide 33Figure ground segmentation extracts the target from previous sequenceFigure ground segmentation,cont.Motion detection returns areas that possibly belong to moving targetTraget segmentation returns areas that are believed pursued targetPursuit, can handle occlusionTraget pixels extracted from the previous sequenceTarget pixels extracted without disparity cue. Attention shifts to the second moving perosnPursuit during target expansionNotesPursuit of a rotating umbrellaMotion detection with real time motor control feedbaclTracking the white boxSlide 46Active Visual ObserverIntegration of Visual Processes for Control of FixationKTH (Royal Institute of Technology, Stockholm)and Aalborg UniversityC.S. Andersen and H.I.ChristensenArchitecture for controlling an agile camera•Basic system facilitates three low level processes:•Fixation•Tracking •Attention selection and shiftingThe basic idea•Is that a tight coupling between the lowest visual processes, referred to as the basic system, and the sensing apparatus, with known latencies, is imperative for successful operation in dynamic environments. Following the biological inspiration, the basic functionality of a camera head are: fixation, gaze shift and smooth pursuit.•A system capable of addressing these aspects of active vision will be capable of fixating on an object, and maintaning fixation while it is moving, or during ego motion of the head.The basic system layoutThe attention mechanisms•The attention mechanism will allow for selection of interesting (salient?) points from the input data. The system can perform selection of fixation points, fixation and tracking.•Below is a standard control system for a DC motor with tachometer feedback , with “normal” appearance at the top and the control schematic at the bottom.Standard control systemDesigning the Architecture•Biologists have argued convincingly that eye movements typically are performed in two separate stages, Version and vergence, with both eyes participating in both motion patterns, while fixating at some point in the space.The version angle is the direction of gaze for an imaginary eye positioned between the two rotation centers in next figure.Layout of Aalborg headCyclopean representationThe version angle relies on the two vergence motor settings.The pan motor contributes however, along with the vergence motor to the direction of gaze.istiltandisvergencetheisversionthe ;Control•We may use one camera as leading and the other following.•The visual process of tracking in the leading eye approach is roughly equivalent to performing control of version and tilt in the cyclopean representation, while fixation corresponds to the process of vergence control. Hence renaming the modules and utilizing a different representation the basic control architecture may facilitate equal eye control as shown in figure belowArchitecture for an equal eye dominance control scheme.Notes to the figure•It should be noted that the figure only displays the “forward control lines” . Actually there are feedback signals form the hardware to the visual processes, as well as communication between the individual processing modules. The signals in the system is as described earlier the actions issued by the processing modules, which in this case is vergence,version and tilt angle adjustments. Thus the close connection with the actual control of hardware is still maintainedCompleting the Architecture•So far we presented only the mechanical control associated with the eye movements. The system has addiotnal rotational degree of freedom, the pan. There is alos motorized lenses with 3 degrees of freedom: focal length (zoom), focus (accommodation) distance and aperture.The modified cyclopean control architectureAn Experimental System•Final system relies on correlation based stabilization fro the left and right camera. The computed image slip from the two cameras is combined to form the error signal for control of the version and tilt angles. While a disparity estimate could be computed from the target location in the image pair, it has been chosen to perform an explicit disparity extraction by correlating the images. This provides redundant information but it also allows for a more robust control since a loss of disparity information does not necessarily mean that version and tilt control cannot be performed and vice versa.Fixation distance for combined disparity and accomodation controlAttention selection•The figure below shows how the system selected areas of high contrast. Using the centroid of the receptive field as fixation point, the fixation has been shifted resulting in vergence-version-tilt angle changes as shown to the right of the figure below.The receptive fieldsAnother Active Visual Observer•Binocular Active Vision System that can attend to and fixate a moving target, in particular is capable of figure-ground segmentation. This work focuses on occlusions of other both stationary and moving targets and integrate three cues to obtain an overall robust behavior, ego-motion, target motion and target disparity.Major parts of the current system•Selecting a target•Control of the system for saccade and pursuit•Measuring the speed of the target for pursuit•Measuring and selecting a disparity for pursuitSystem description•Fundamental skills are: fixation, target pursuit and target discrimination•The full system includes the integration of three cues for target selection and target discrimination. These are used by the moving observer to smoothly pursue
View Full Document
Unlocking...