Perceptual grouping by multiple hypothesis probabilistic data association

Claim 2

A computer vision process for finding at least one contour in an image scene comprising the steps of, a) storing a frame of the image scene, b) analyzing said frame by an edge detection process for delineating points of said image scene that are points of high intensity gradient that are to be partitioned into disjoint contours, c) scanning a local neighborhood of said points for observing edgels in said neighborhood, d) matching the observed edgels for a neighborhood with predicted edgels for said neighborhood for a hypothesis for a contour based on edgels observed or predicted for said neighborhood, said predicted edgels being based on earlier observed and predicted edgels for adjacent edgels in the local neighborhood, e) generating data association hypotheses based on the hypothesis matrix that assign specific edgels to specific contours and deriving a hypothesis tree, f) by discarding hypotheses that have relatively low probability and retaining hypotheses that have relatively high probability, reducing the total number of hypotheses and partitioning the hypothesis tree into spatially disjoint trees, g) utilizing the saved hypothesis for predicting edgels for the next neighborhood for use in the earlier-mentioned step of matching, and repeating steps d-g until a desired contour has been delineated.

Claim 3

A computer vision process for partitioning an image scene into disjoint sets of points based on a method of probabilistic data association, particularly, a computer vision process for segmenting an image into at least one spatially disjoint region comprising the steps of storing a frame of an image scene, scanning a local neighborhood of said scene for observing edgels in said neighborhood, matching the observed edgels for a neighborhood with predicted edgels values for said neighborhood for developing a hypothesis matrix for a region based on edgels observed or predicted for that neighborhood, said predicted edgels being based on earlier observed and predicted values, assigning newly observed edgels in said hypothesis matrix to regions, generating data association hypotheses based on the hypothesis matrix that assigns specific edgels to the partition of an image scene into spatially disjoint regions, providing probability statistics to each hypothesis generated, summing the probability statistics of each hypothesis and discarding hypotheses with relatively low probability statistics while saving for utilization and for storage the hypotheses with relatively high probability statistics, reducing the total number of hypotheses by partitioning the hypotheses tree into spatially disjoint trees, and utilizing the saved hypotheses for generating predicted edgel values for the next neighborhood for use in the earlier-mentioned matching, repeating the processing until the desired partitioning of the image scene has been delineated.

Claim 4

A computer vision process for partitioning an image scene into disjoint sets of points based on a method of probabilistic data association, particularly, a computer vision process for segmenting an image into at least one spatially disjoint region comprising the steps of storing a frame of an image scene, scanning a local neighborhood of said scene for observing edgels in said neighborhood, matching the observed edgels for a neighborhood with predicted edgels values for said neighborhood for developing a hypothesis matrix for a region based on edgels observed or predicted for that neighborhood, said predicted edgels being based on earlier observed and predicted values, assigning newly observed edgels in said hypothesis matrix to regions, generating data association hypotheses based on the hypothesis matrix that assigns specific edgels to the partition of an image scene into spatially disjoint regions, providing probability statistics to each hypothesis generated, summing the probability statistics of each hypothesis and discarding hypotheses with relatively low probability statistics while saving for utilization and for storage the hypotheses with relatively high probability statistics, reducing the total number of hypotheses by partitioning the hypotheses tree into spatially disjoint trees, and utilizing the saved hypotheses for generating predicted edgel values for the next neighborhood for use in the earlier-mentioned matching, repeating the processing until the desired partitioning of the image scene has been delineated.