Motion detection of objects for the implementation of a motion compensated tomographic device

In this thesis, we present two computer vision algorithms for object motion detection in the field of Computed Tomography (CT). Traditionally, CT devices require objects to remain stationary while being scanned within a moving gantry. However, there are cases where it is advantageous to have a fixed scanner and allow objects to pass through it while in motion. For such a design, accurate determination of object motion during X-ray scanning is crucial for tomographic reconstruction. To address this challenge, we have developed two Computer Vision algorithms, utilizing one and two cameras respectively. Their purpose is to extract the rotation and translation matrices of the scanned object through a sequence of frames acquired during motion. Our algorithms involves leveraging various computer vision techniques such as feature extraction (AKAZE), feature matching (k-NN) and circle detection (Hough), which is useful in calbrating scanner geometry to establish relationships among reference systems within it. With the information obtained, the required matrices are subsequently estimated (exploiting Kabsch and RANSAC). Calibration of the single-camera or stereo vision system using known patterns, such as checkerboards, is also discussed. Through optimization of design parameters, we improved performance. Our reconstruction results have been promising, contributing to the advancement of a novel device for industrial inline Computed Tomography. Distinguished from most tomographic devices described in literature, this new scanner features a fixed X-ray source and detector, with the object rotating during the scanning process.