초록 ▼

This invention provides a system and method for determining position of a viewed object in three dimensions by employing 2D machine vision processes on each of a plurality of planar faces of the object, and thereby refining the location of the object. First a rough pose estimate of the object is der

This invention provides a system and method for determining position of a viewed object in three dimensions by employing 2D machine vision processes on each of a plurality of planar faces of the object, and thereby refining the location of the object. First a rough pose estimate of the object is derived. This rough pose estimate can be based upon predetermined pose data, or can be derived by acquiring a plurality of planar face poses of the object (using, for example multiple cameras) and correlating the corners of the trained image pattern, which have known coordinates relative to the origin, to the acquired patterns. Once the rough pose is achieved, this is refined by defining the pose as a quaternion (a, b, c and d) for rotation and a three variables (x, y, z) for translation and employing an iterative weighted, least squares error calculation to minimize the error between the edgelets of trained model image and the acquired runtime edgelets. The overall, refined/optimized pose estimate incorporates data from each of the cameras' acquired images. Thereby, the estimate minimizes the total error between the edgelets of each camera's/view's trained model image and the associated camera's/view's acquired runtime edgelets. A final transformation of trained features relative to the runtime features is derived from the iterative error computation.

대표청구항 ▼

1. A method for registering an object in three dimensions using machine vision comprising the steps of: at training time, acquiring training images of an object used for training with one or more cameras;at runtime, acquiring runtime images of an object to be registered at runtime with the one or mo

1. A method for registering an object in three dimensions using machine vision comprising the steps of: at training time, acquiring training images of an object used for training with one or more cameras;at runtime, acquiring runtime images of an object to be registered at runtime with the one or more cameras; anddetermining a three dimensional pose transformation between the pose of the object used at training time and the pose of the object to be registered at runtime by(a) defining features in each of the runtime images as three-dimensional rays through an origin of each of the one or more camera's, respectively,(b) associating the three-dimensional rays with corresponding runtime features from the training images, and(c) computing an optimal pose estimate which maps the training features onto the corresponding three-dimensional rays of runtime features using iterative, reweighted least squares analysis. 2. The method as set forth in claim 1 wherein the step of acquiring training images includes acquiring at least two poses of at least one training planar face of the object used for training, the at least one training planar face being located within a defined training subwindow of a field of view of the one or more cameras, the training subwindow including spatial coordinates. 3. The method as set forth in claim 1 wherein the step of computing the optimal pose comprises for each of the runtime features, establishing a plane normal to a plane passing through each of the runtime features and a corresponding ray from an origin of the one or more cameras, and minimizing a sum squared error of a dot product between the plane normal and a center point of the one of the closest training features associated with the plane passing through the respective of the runtime features. 4. The method as set forth claim 3 wherein the step using the iterative, reweighted least squares analysis includes changing a weighting with respect to each of the runtime features based upon an error for that runtime feature versus an averaged error for all runtime features. 5. The method as set forth in claim 4 wherein the averaged area is based upon a root-mean-square error. 6. A method for registering an object in three dimensions using machine vision comprising the steps of: at training time, acquiring training images of an object used for training with one or more cameras;at runtime, acquiring runtime images of an object to be registered at runtime with the one or more cameras; anddetermining a three dimensional pose transformation between the pose of the object used training time and the pose of the object to be registered at runtime by(a) defining features in each of the runtime images as three-dimensional rays through an origin of each of the one or more camera's, respectively,(b) associating the three-dimensional rays with corresponding runtime features from the training images, and(c) computing an optimal pose estimate which maps the training features onto the corresponding three-dimensional rays of runtime features using iterative, reweighted least squares analysis;wherein the step of acquiring training images includes acquiring at least two poses of at least one training planar face of the object used for training, the at least one training planar face being located within a defined training subwindow of a field of view of the one or more cameras, the training subwindow including spatial coordinates;wherein the step of acquiring runtime images includes, determining candidate features representative of at least one runtime planar face in the runtime image data that correspond to the at least one training planar face and associating positions in three-dimensions of the spatial coordinates of the training subwindow with a spatially reoriented version of each of the at least one runtime planar face that corresponds with the at least one training planar face. 7. A system for registering an object in three dimensions using machine vision comprising: one or more cameras constructed and arranged so that, at training time, the one or more cameras each acquire training images of an object used for training, and at runtime, the one or more cameras acquire runtime images of an object to be registered at runtime with the one or more cameras; anda pose transformation determination process that computes a three-dimensional pose transformation between the pose of the object used training time and the pose of the object to be registered at runtime by(a) defining features in each of the runtime images as three-dimensional rays through an origin of each of the one or more camera's, respectively,(b) associating the three-dimensional rays with corresponding runtime features from the training images, and(c) computing an optimal pose estimate which maps the training features onto the corresponding three-dimensional rays of runtime features using iterative, reweighted least squares analysis. 8. The system as set forth in claim 7 wherein the pose transformation determination process is constructed and arranged to acquire at least two poses of at least one training planar face of the object used for training, the at least one training planar face being located within a defined training subwindow of a field of view of the one or more cameras, the training subwindow including spatial coordinates. 9. The system as set forth in claim 7 wherein the pose transformation determination process is constructed and arranged to compute the optimal pose by, for each of the runtime features, establishing a plane normal to a plane passing through each of the runtime features and a corresponding ray from an origin of the one or more cameras, and minimizing a sum squared error of a dot product between the plane normal and a center point of the one of the closest training features associated with the plane passing through the respective of the runtime features. 10. The system as set forth claim 9 wherein the pose transformation determination process is constructed and arranged to change a weighting with respect to each of the runtime features based upon an error for that runtime feature versus an averaged error for all runtime features. 11. The system as set forth in claim 10 wherein the averaged area is based upon a root-mean-square error. 12. A system for registering an object in three dimensions using machine vision comprising: one or more cameras constructed and arranged so that, at training time, the one or more cameras each acquire training images of an object used for training, and at runtime, the one or more cameras acquire runtime images of an object to be registered at runtime with the one or more cameras; anda pose transformation determination process that computes a three-dimensional pose transformation between the pose of the object used training time and the pose of the object to be registered at runtime by(a) defining features in each of the runtime images as three-dimensional rays through an origin of each of the one or more camera's, respectively,(b) associating the three-dimensional rays with corresponding runtime features from the training images, and(c) computing an optimal pose estimate which maps the training features onto the corresponding three-dimensional rays of runtime features using iterative, reweighted least squares analysis;wherein the pose transformation determination process is constructed and arranged to acquire at least two poses of at least one training planar face of the object used for training, the at least one training planar face being located within a defined training subwindow of a field of view of the one or more cameras, the training subwindow including spatial coordinates;wherein the pose transformation determination process is constructed and arranged to determine candidate features representative of at least one runtime planar face in the runtime image data that correspond to the at least one training planar face and associating positions in three-dimensions of the spatial coordinates of the training subwindow with a spatially reoriented version of each of the at least one runtime planar face that corresponds with the at least one training planar face. 13. A non-transitory computer-readable medium containing executable program instructions, which when executed by the computer perform the steps of registering an object in three dimensions using machine vision, the executable program instructions comprising program instructions for: at training time, acquiring training images of an object used for training with one or more cameras;at runtime, acquiring runtime images of an object to be registered at runtime with the one or more cameras; anddetermining a three dimensional pose transformation between the pose of the object used training time and the pose of the object to be registered at runtime by(a) defining features in each of the runtime images as three-dimensional rays through an origin of each of the one or more camera's, respectively,(b) associating the three-dimensional rays with corresponding runtime features from the training images, and(c) computing an optimal pose estimate which maps the training features onto the corresponding three-dimensional rays of runtime features using iterative, reweighted least squares analysis.