Publication Details

SELECT * FROM publications WHERE Record_Number=11296
Reference TypeThesis
Author(s)Baig, I.
TitleDeep End-to-End Calibration Thesis
Journal/Conference/Book TitleMaster Thesis
AbstractHand-eye calibration is an essential operation for intelligent robots to effectively utilize their sensor data, however,traditional calibration techniques are laborious and time consuming, as they require careful setup of calibration markers[1] in the environment. The repetitive nature of the task and human administration also introduces the possibility oferrors in the calibration process, which may be eliminated if the calibration and environment is memorized by a system.This thesis investigates a deep learning based approach to memorize an environment using RGBD image and uncalibratedlink pose data, and perform hand-eye calibration by comparing object and pose transformations in the learnt scene. Apredicted pose is considered calibrated camera pose if it is aligned with the camera axis, in which case, any transformationin the pose should also transform the image objects by a calculable amount; by comparing the consistency of the twoentities we propose to find true camera pose in world frame. In this work, we use geometric transformations to calculatetransform of the objects, by first projecting pixels to a point cloud, using depth information, transforming each point, andreprojecting it back to the image plane. However, our experiments show that geometric transformation of images is not avery reliable signal to learn calibration parameters. This thesis lays the foundation for future work of more sophisticateddeep learning based hand-eye calibration methods, as the proposed method requires significant improvements to beapplicable in real world situations.
Link to PDF


zum Seitenanfang