Back to results
US 7,613,357 B2Granted
Method for warped image object recognition
AssigneeGM Global Technology Operations LLC
InventorsYuri Owechko, Swarup Medasani
Life of this patent
Estimating the term…
Abstract
A computer vision system includes distorting optics producing distorted or warped input images. The system includes an integral feature object classifier trained using an undistorted image space. Undistorted integral feature values are calculated directly from distorted input images without undistorting or dewarping the distorted input image.
Claims9
Method for calculating undistorted integral features without generating an undistorted image for use in a computer vision object recognition system, comprising:
- within a microprocessor, calculating a distorted integral image delimited by a predetermined region by integrating over the predetermined region the product of a distorted input image captured in a distorted image space and a predetermined vector transformation between an undistorted image space and the distorted image space;
- generating the undistorted integral features by adding and subtracting values of the distorted integral image without generating an undistorted image;
- and
- utilizing the undistorted integral features in an object classifier-based computer vision object recognition system;
- and
- wherein the undistorted integral features describe a restricted set of object features for recognition in the input image in the undistorted image space;
- and
- wherein the distorted integral image comprises a numerical sum of image pixel values in the predetermined region of the distorted input image.
5 more elements
Method for calculating undistorted integral features from a distorted input image for use in a computer vision object recognition system, comprising:
- providing a Jacobian determinant relating a distorted image space to an undistorted image space;
- within a microprocessor, calculating and storing a distorted integral image based on the distorted input image and the provided Jacobian determinant;
- and
- calculating undistorted integral features using distorted integral image values comprising:
- referencing distorted integral image values of the integral image delimited by a predetermined region;
- and
- algebraically combining the distorted integral image values;
- and
- utilizing the undistorted integral features in an object classifier-based computer vision object recognition system;
- and
- wherein the calculating the distorted integral image is based on the distorted input image without generating an undistorted image;
- wherein the undistorted integral features describe a restricted set of object features for recognition in the input image in the undistorted image space;
- and
- wherein the distorted integral image comprises a numerical sum of image pixel values in the predetermined region of the distorted input image.
11 more elements
Method for object recognition in a computer vision system having a distorted input image, comprising:
- providing an integral feature object classifier trained using an undistorted image space;
- providing a transformation mapping of a distorted image space to the undistorted image space;
- providing a distorted input image captured in the distorted image space;
- within a microprocessor, calculating undistorted integral features based on the distorted input image and the transformation mapping;
- and
- recognizing objects in the distorted input image with an object classifier-based computer vision object recognition system based on the calculated undistorted integral features and the provided integral feature object classifier;
- and
- wherein calculating undistorted integral features from the distorted input image based on the transformation mapping comprises integrating over a predetermined region the product of the distorted image and the transformation mapping;
- wherein the undistorted integral features describe a restricted set of object features for recognition in the input image in the undistorted image space;
- and
- wherein the distorted integral image comprises a numerical sum of image pixel values in the predetermined region of the distorted input image.
8 more elements
Method for calculating undistorted integral features from a distorted input image for use in a computer vision object recognition system, comprising:
- providing a Jacobian determinant relating a distorted image space to an undistorted image space;
- within a microprocessor, integrating over a predetermined region the product of a distorted input image in the distorted image space and the Jacobian determinant, the integrating determining a distorted integral image;
- calculating the undistorted integral features based upon the determined distorted integral image;
- and
- utilizing the undistorted integral features in an object classifier-based computer vision object recognition system;
- and
- wherein the undistorted integral features describe a restricted set of object features for recognition in the input image in the undistorted image space;
- and
- wherein the distorted integral image comprises a numerical sum of image pixel values in the predetermined region of the distorted input image.
6 more elements
Method for calculating undistorted integral features from a distorted input image for use in a pattern recognition system, comprising:
- monitoring the distorted input image;
- within a microprocessor, determining a distorted integral image in warped space comprising an array of values for every pixel in the monitored distorted input image, wherein each value of the array can be described by a distorted image pixel column number and a distorted image pixel row number and wherein each value of the array comprises a summation of image values of each of the pixels with a lower distorted image pixel column number and a lower distorted image pixel row number;
- defining a rectangular region in unwarped space, wherein the unwarped space is defined by an unwarped space pixel column number and an unwarped space pixel row number and wherein the rectangular region is defined by a high unwarped space pixel row number, a low unwarped space pixel row number, a high unwarped space pixel column number, and a low unwarped space pixel column number;
- determining the integral of the rectangular region in unwarped space by integrating the sum of values of the array in a region in warped space corresponding to the rectangular region;
- and
- utilizing the integral of the rectangular region to perform pattern recognition;
- and
- wherein the utilizing the integral of the rectangular region to perform pattern recognition comprises:
- iteratively defining rectangular regions in unwarped space;
- determining integrals of each of the rectangular regions in unwarped space;
- determining unwarped integral features for the distorted input image based upon the determined integrals of the rectangular regions;
- and
- performing pattern recognition upon the determined unwarped integral features with an object classifier-based computer vision object recognition system;
- and
- wherein the undistorted integral features describe a restricted set of object features for recognition in the input image in the undistorted image space.
12 more elements
Dependencies are inferred from the claim text.