On Feature Tracking in X-Ray Images

Abstract

Abstract. Feature point tracking and detection of X-ray images is challenging due to overlapping anatomical structures of different depths, which lead to low-contrast images. Tracking of motion in X-ray sequences can support many clinical applications like motion compensation or 2D / 3D registration algorithms. This paper is the first to evaluate the performance of several feature tracking and detection algorithms on artificial and real X-ray image sequences, which involve rigid motion as well as external disturbances. A stand-alone application has been developed to provide an overall test bench for all algorithms, realized by OpenCV implementations. Experiments show that the Kanade-Lucas-Tomasi tracker is the most consistent and effective tracking algorithm. Considering external disturbances, template matching provides the most sufficient results. Furthermore, the influence of feature point detection methods on tracking results is shown.

Multi-view Object Detection Using Epipolar Constraints within Cluttered X-ray Security Imagery

“…where ∆x ∼ N (0, σ 2 ) is the error associated with the measurement process of each point coordinate. Under our conditions, where we have a set of uncalibrated X-ray viewpoints, traditional feature points based matching will fail [21]. The only available information is the ground truth bounding boxes of the threat items used for training the object detection model.…”

“…When the geometry is unknown (uncalibrated cameras) and point correspondences are not provided, the common methodology is to use feature detectors and descriptors to find matches between the different image views and then proceed to solve for F via least-squares minimization of the geometric inter-image feature projection error [20]. However, the prior work from Kluppel et al [21] demonstrates that conventional feature detection and matching is not suitable for transmission imagery such as X-ray due to the transparent nature of the object projections which vary with perspective view. Moreover, prior object detection work using multiple view X-ray imagery, with consideration for epipolar constraints, is limited and primarily focuses on 3D bounding box reconstruction [22], where three views are needed [16].…”

“…-A novel approach for recovery of the fundamental matrix from uncalibrated views based on the use of ground truth object-level annotations, applied to transmission (X-ray) imagery where conventional feature point matching fails [21]. -Formulation of a multi-view detection approach that cross correlates detections from multiple views by considering the inter-view epipolar constraint as an additional measure of confidence with NMS post-processing.…”