Six degree of freedom pose estimation using color and depth feature descriptors for an industrial application

Manufacturing in the United States may be on the verge of a revolution. The US government is investing heavily in manufacturing innovations and new technologies are enabling improvements in capabilities of robotics and automation. One of the recent developments is the availability of new cost-effective 3D sensors. With these 3D sensors, better, more descriptive features can be extracted from the data. These more descriptive features should perform better in object recognition algorithms. The purpose of this work is to explore kernel descriptor performance with respect to object pose estimation. These kernel descriptors are derived from images acquired with an Aus Xtion Pro 3D sensor. Four kernel descriptors are used to create a predictive model using a support vector machine. This model is cross validated by splitting the set of collected images into a test set and train set. The training set is used to create the SVM model and the model is tested against the test set resulting in accuracy based upon number of correct predictions versus total test samples. Using the four kernel descriptors resulted in 83% accuracy for estimation of the rotation of the industrial part. Using the trained SVM model along with algorithms from the Point Cloud Library, a six degree of freedom pose was determined. This pose estimation was compared to a measured pose (measured with a coordinate measuring machine) to obtain the accuracy of the pose estimation software pipeline.