This thesis deals with the problem of obtaining an accurate registration for objects in multi-modal imagery within the range of 1 -- 4 m. The work focuses on providing a registration for all objects inside the range. An acquisition platform is built in both hardware and software that enables the synchronized capture of visual, thermal, and depth images. The depth image is by default registered to the visual image.
Three scenes within the range are defined. For each scene, a custom-built thermal-visible calibration rig is used to provide point correspondences between the views. Two rectification algorithms have been investigated and tested by the use of the generated point correspondences; stereo rectification and rectification by multiple homographies. The stereo rectification is however shown to provide poor results. The rectification by multiple homographies is based on training data to generate $k$ homographies which are used to map points between the thermal and visible modalities. This method is shown to provide a better accuracy of mapping visible points to the thermal modality than the baseline of a single homography. However, the method falls short on providing the opposite transfer.