Local feature description with invariance against affine projection

While the understanding of the image content is relatively easy for most humans, an automatic analysis is challenging for a computer vision system. To describe the content of an image, local features, which can be understood as mathematical representations of image regions, are used frequently. The image description should be invariant against photo- and geometric image distortions which typically occur when the illumination or the viewing angle during image acquisition changes.This thesis focuses on increasing the robustness of local features under geometric, or more specific, affine image projection, in the context of object detection.To increase robustness of local features against geometric image distortions, the affine invariant coordinate transformation is developed. The affine invariant coordinate transformation is an iterative normalization algorithm which exploits local properties of an image region to normalize it, such that two image regions captured from different viewpoints are identical, up to a rotational transformation after normalization. It can be combined with various feature detection and feature extraction algorithms and used for both globally and locally distorted images.For the detection of objects with various affine projections in different image recordings, the correspondence consensus merging is developed. The correspondence consensus merging uses the normalization matrices of two corresponding features to estimate the projection between the objects. Based on the assumption that features belonging to one semantic object are projected similarly across the images, the correspondence consensus merging groups correspondences whose projection estimates are similar. The algorithm provides reliable object detection results even when the established correspondences between the two images contain plenty false correspondences and can also be used to distinguish between correct and false feature correspondences.The developed algorithms are evaluated on synthetically warped as well as on camera captured image pairs with global and local geometric projection and compared against state of the art methods for affine invariant feature extraction, respectively for feature grouping. It is shown, that especially for images with local geometric projections, the presented algorithm is superior to the state of the art. Furthermore, it is shown that the presented grouping of feature correspondences allows for reliable object detection