Engineering topics
Publications and source records attributed to Yakimovsky, Y..
Digitizer and storage system allow rapid random access to video data by computer. RAPID (random-access picture digitizer) uses two commercially-available, charge-injection, solid-state TV cameras as sensors. It can continuously update its memory with each frame of video signal, or it can hold given frame in memory. In either mode, it generates composite video output signal representing digitized image in memory.
Obtaining accurate three-dimensional (3-D) measurement from a stereo pair of TV cameras is a task requiring camera modeling, calibration, and the matching of the two images of a real 3-D point on the two TV pictures. A system that models and calibrates the cameras and pairs the two images of a real-world point in the two pictures, either manually or automatically, was implemented at JPL. This system is operating and provides three-dimensional measurement resolution of plus or minus 5 mm at distances of about 2 m.
A few new algorithms for region growing in pictures were developed. These algorithms are a step toward finding a satisfactory solution to the image segmentation problem, and in-depth understanding of the problems of nonsemantic image segmentation. The algorithms utilize a sequential decision approach for region boundaries detection. The sequential decisions are supported by a stochastic algorithm that maintains local statistics of the region near the boundaries as the region grows. A few illustrations of the algorithm's performance are included.
RAPID is a system capable of providing convenient digital analysis of video data in real-time. It has two modes of operation. The first allows for continuous digitization of an EIA RS-170 video signal. Each frame in the video signal is digitized and written in 1/30 of a second into RAPID's internal memory. The second mode leaves the content of the internal memory independent of the current input video. In both modes of operation the image contained in the memory is used to generate an EIA RS-170 composite video output signal representing the digitized image in the memory so that it can be displayed on a monitor.
A description is given of the data base used in the perception subsystem of the Mars robot vehicle prototype being implemented at the Jet Propulsion Laboratory. This data base contains two types of information. The first is generic (uninstantiated, abstract) information that specifies the general rules of perception of objects in the expected environments. The second kind of information is a specific (instantiated) description of a structure, i.e., the properties and relations of objects in the specific case being analyzed. The generic knowledge can be used by the approximate reasoning subsystem to obtain information on the specific structures which is not directly measurable by the sensory instruments. Raw measurements are input either from the sensory instruments or a human operator using a CRT or a TTY.
Obtaining accurate three-dimensional (3-D) measurement from a stereo pair of TV cameras is a task requiring camera modeling, calibration, and the matching of the two images of a real 3-D point on the two TV pictures. A system which models and calibrates the cameras and pairs the two images of a real-world point in the two pictures, either manually or automatically, was implemented. This system is operating and provides three-dimensional measurements resolution of + or - mm at distances of about 2 m.
A description is given of the knowledge representation data base in the perception subsystem of the Mars robot vehicle prototype. Two types of information are stored. The first is generic information that represents general rules that are conformed to by structures in the expected environments. The second kind of information is a specific description of a structure, i.e., the properties and relations of objects in the specific case being analyzed. The generic knowledge is represented so that it can be applied to extract and infer the description of specific structures. The generic model of the rules is substantially a Bayesian representation of the statistics of the environment, which means it is geared to representation of nondeterministic rules relating properties of, and relations between, objects. The description of a specific structure is also nondeterministic in the sense that all properties and relations may take a range of values with an associated probability distribution.
The sensory subsystem of the JPL robotics project is described. The first sensory instrument is a pair of TV cameras, which provides a stereo view of the area of interest. The second is a laser range finder, which can be aimed in a requested direction and measure distance to a surface point that reflects the beam back. Both instruments are controlled by and are interfaced to computer. The calibration process and basic applications to obtain measurements of the external environment are described.
A solution to the problem of automatic location of objects in digital pictures by computer is presented. A self-scaling local edge detector which can be applied in parallel on a picture is described. Clustering algorithms and boundary following algorithms which are sequential in nature process the edge data to locate images of objects.
An approach to simultaneous interpretation of objects in complex structures so as to maximize a combined utility function is presented. Results of the application of a computer software system to assign meaning to regions in a segmented image based on the principles described in this paper and on a special interactive sequential classification learning system, which is referenced, are demonstrated.
An approach to extracting depth information out of a stereo pair of images is described. Special camera pair configurations are shown to be effective in reducing complexity. A simple algorithm which pairs objects in the two images and allows verification is described.