3D pose estimation by a 2D camera
A system and method for obtaining a 3D pose of an object using 2D images from a 2D camera and a learned-based neural network. The neural network extracts a plurality of features on the object from the 2D images and generates a heatmap for each of the extracted features that identify the probability of a location of a feature point on the object by a color representation. The method provides a feature point image that includes the feature points from the heatmaps on the 2D images, and estimates the 3D pose of the object by comparing the feature point image and a 3D virtual CAD model of the object.
1. A method for estimating a 3D pose of an actual object, said method comprising:
obtaining a 2D image of the actual object using a 2D camera;
extracting a plurality of features on the actual object from the 2D image using a neural network;
generating an actual heatmap for each of the extracted features that identify the probability of a location of a feature point on the actual object;
providing a feature point image that combines the feature points from the heatmaps and the 2D image; and
estimating the 3D pose of the actual object using the feature point image.
2. The method according to claim 1 wherein estimating the 3D pose of the actual object includes comparing the feature point image to a 3D virtual model of the object.
3. The method according to claim 2 wherein estimating the 3D pose of the actual object includes using a perspective-n-point algorithm.
4. The method according to claim 1 wherein the probability of a location of the feature point in the heatmap is shown as color on the heatmap.
5. The method according to claim 1 wherein the probability of a location of the feature point on the actual object is assigned a confidence value that it is a feature point.
6. The method according to claim 1 further comprising training the neural network by obtaining a data set including placing a training object representative of the actual object on a marker board including a marker, measuring an offset between the marker and the training object on the board, taking at least one training 2D image of the marker board, identifying a 3D pose of the marker in the training 2D image, adding the offset to the 3D pose of the marker to obtain feature points on the training object, generating a separate training heatmap identifying the probability of each feature point location, and using the training heatmaps to train the neural network.
7. The method according to claim 6 wherein taking at least one training 2D image includes taking a plurality of training 2D images at different orientations.
8. The method according to claim 6 wherein generating a separate training heatmap uses a Gaussian distribution process.
9. The method according to claim 1 wherein extracting a plurality of features on the actual object includes extracting at least four features.
10. The method according to claim 1 wherein the method is employed in a robotic system and the actual object is being picked up a robot.
11. A method for estimating a 3D pose of an object, said object being picked up by a robot, said method comprising:
obtaining 2D images of the object using a 2D camera;
extracting a plurality of feature points on the object from the 2D images using a neural network, where weights of nodes in the neural network are trained from a collected training set using a representative object of the object;
generating a heatmap for each of the extracted features that identify the probability of a location of a feature point on the actual object, wherein the probability of a location of the feature point in the heatmap is shown as color on the heatmap;
providing a feature point image that combines the feature points from the heatmaps and the 2D image; and
estimating the 3D pose of the actual object by comparing the feature point image to a 3D virtual model of the object using a perspective-n-point algorithm.
12. The method according to claim 11 wherein the probability of a location of the feature on the actual object is assigned a confidence value that it is a feature point.
13. The method according to claim 11 wherein extracting a plurality of features on the actual object includes extracting at least four features.
14. A system for estimating a 3D pose of an object, said system comprising:
means for obtaining a 2D image of the actual object using a 2D camera;
means for extracting a plurality of features on the actual object from the 2D image using a neural network;
means for generating an actual heatmap for each of the extracted features that identify the probability of a location of a feature point on the actual object;
means for providing a feature point image that combines the feature points from the heatmaps and the 2D image; and
means for estimating the 3D pose of the actual object using the feature point image.
15. The system according to claim 14 wherein the means for estimating the 3D pose of the actual object compares the feature point image to a 3D virtual model of the object.
16. The system according to claim 15 wherein the means for estimating the 3D pose of the actual object uses a perspective-n-point algorithm.
17. The system according to claim 14 wherein the probability of a location of the feature point in the heatmap is shown as color on the heatmap.
18. The system according to claim 14 wherein the probability of a location of the feature on the object is assigned a confidence value that it is a feature.
19. The system according to claim 14 further comprising means for training nodes in the neural network from a collected training set using a representative object of the object.
20. The system according to claim 14 wherein the system is employed in a robotic control and the object is being picked up a robot.