The primary goal of this dissertation is to develop point-based rigid and non-rigid image registration methods that have better accuracy than existing methods. We first present point-based PoIRe, which provides the framework for point-based global rigid registrations. It allows a choice of different search strategies including (a) branch-and-bound, (b) probabilistic hill-climbing, and (c) a novel hybrid method that takes advantage of the best characteristics of the other two methods. We use a robust similarity measure that is insensitive to noise, which is often introduced during feature extraction. We show the robustness of PoIRe using it to register images obtained with an electronic portal imaging device (EPID), which have large amounts of scatter and low contrast. To evaluate PoIRe we used (a) simulated images and (b) images with fiducial markers; PoIRe was extensively tested with 2D EPID images and images generated by 3D Computer Tomography (CT) and Magnetic Resonance (MR) images. PoIRe was also evaluated using benchmark data sets from the blind retrospective evaluation project (RIRE). We show that PoIRe is better than existing methods such as Iterative Closest Point (ICP) and methods based on mutual information. We also present a novel point-based local non-rigid shape registration algorithm. We extend the robust similarity measure used in PoIRe to non-rigid registrations adapting it to a free form deformation (FFD) model and making it robust to local minima, which is a drawback common to existing non-rigid point-based methods. For non-rigid registrations we show that it performs better than existing methods and that is less sensitive to starting conditions. We test our non-rigid registration method using available benchmark data sets for shape registration. Finally, we also explore the extraction of features invariant to changes in perspective and illumination, and explore how they can help improve the accuracy of multi-modal registration. For multimodal registration of EPID-DRR images we present a method based on a local descriptor defined by a vector of complex responses to a circular Gabor filter.

The main result of this thesis is a deep learning model named BearNet, which can be trained to detect an arbitrary amount of objects as a set of points. The model is trained using the Weighted Hausdorff distance as loss function. BearNet has been applied and tested on two problems from the industry. These are: From an intensity image, detect two pocket points of an EU-pallet which an autonomous forklift could utilize when determining where to insert its forks. From a depth image, detect the start, bend and end points of a straw attached to a juice package, in order to help determine if the straw has been attached correctly. In the development process of BearNet I took inspiration from the designs of U-Net, UNet++ and a high resolution network named HRNet. Further, I used a dataset containing RGB-images from a surveillance camera located inside a mall, on which the aim was to detect head positions of all pedestrians. In an attempt to reproduce a result from another study, I found that the mall dataset suffers from training set contamination when a model is trained, validated, and tested on it with random sampling. Hence, I propose that the mall dataset is evaluated with a sequential data split strategy, to limit the problem. I found that the BearNet architecture is well suited for both the EU-pallet and straw datasets, and that it can be successfully used on either RGB,  intensity or depth images. On the EU-pallet and straw datasets, BearNet consistently produces point estimates within five and six pixels of ground truth, respectively. I also show that the straw dataset only constitutes a small subset of all the challenges that exist in the problem domain related to the attachment of a straw to a juice package, and that one therefore cannot train a robust deep learning model on it. As an example of this, models trained on the straw dataset cannot correctly handle samples in which there is no straw visible.

This study examines the efficacy of Sentinel-1 Radial Velocity (RVL) imagery at determining the position of ocean current front features, using the Gulf Stream (GS) and Agulhas Current (AC) as case studies. Fronts derived from RVL imagery are compared to fronts derived from Sea Surface Temperature (SST) imagery, specifically Multi-scale Ultra-high Resolution Sea Surface Temperature Analysis (MURSST) data. In the case of the GS, front locations from the Naval Oceanographic Office (NAVOCEANO) were also used for comparison. Only the northern walls of ocean current features are considered in this study, which is broken into three main steps: Preprocessing, front extraction, and front comparison. First, RVL imagery is selected from Sentinel-1 ocean products, preprocessed to remove antenna mispointing artifacts, and all products from the same orbit are combined into a single swath. Second, front features are extracted from both the RVL and MURSST imagery using a ridge detection algorithm, the main ocean current is chosen from all ridge features using a ranking algorithm, and the northern wall of this current is extracted. Third, the RVL, SST, and in the case of the GS, NAVOCEANO GS locations, features are compared using a symmetric Hausdorff Distance (HD) measure, and Mean Hausdorff Distance (MHD). In some cases, the automatic front extraction failed by either misclassifying an eddy or similar ocean feature as the ocean current in either the RVL or SST image or failed to extract the entire length of the front visible within the image. All the SST and RVL fronts were classified manually to determine the success rate of the automatic front extraction and to exclude failed front extractions from the analysis, as they are not accurate representations of the SST and RVL data’s ability to detect fronts. In special cases, the RVL image itself does not detect the entire ocean current, such that there are noticeable gaps in the ocean current. Similarly, in special cases the MURSST does not detect the entire ocean current. The automatic front extraction succeeded 65% of the time, including the special cases. The results demonstrated that RVL products were effective at determining the location of ocean fronts where the angle of the front's normal vector is within approximately 40° of the sensor’s azimuthal heading. A mean HD of 31.9 km and a mean MHD of 13.2 km was calculated for all front pairs over all study areas. The RVL fronts appeared consistently to the north of the SST fronts, with an average offset of 25.4 km between the centroids of the SST and RVL fronts. Positive correlations were noted between cloud coverage and MURSST error in both study regions. Several RVL images detected ocean currents in regions of high MURSST error where the MURSST did not detect the ocean currents, suggesting that RVL may provide more accuracy than SST-based products in clouded regions where there is no auxiliary data.