| Popis: |
This thesis presents an investigation into the integration of information extracted from co-aligned range and intensity images to achieve pose invariant object recognition. Local feature matching is a fundamental technique in image analysis that underpins many computer vision-based applications; the approach comprises identifying a collection of interest points in an image, characterising the local image region surrounding the interest point by means of a descriptor, and matching these descriptors between example images. Such local feature descriptors are formed from a measure of the local image statistics in the region surrounding the interest point. The interest point locations and the means of measuring local image statistics should be chosen such that resultant descriptor remains stable across a range of common image transformations. Recently the availability of low cost, high quality range imaging devices has motivated an interest in local feature extraction from range images. It has been widely assumed in the vision community that the range imaging domain has properties which remain quasi-invariant through a wide range of changes in illumination and pose. Accordingly, it has been suggested that local feature extraction in the range domain should allow the calculation of local feature descriptors that are potentially more robust than those calculated from the intensity imaging domain alone. However, range images represent differing characteristics from those represented within intensity images which are frequently used, independently from range images, to create robust local features. Therefore, this work attempts to establish the best means of combining information from these two imaging modalities to further increase the reliability of matching local features. Local feature extraction comprises a series of processes applied to an image location such that a collection of repeatable descriptors can be established. By using co-aligned range and intensity images this work investigates the choice of modality and method for each step in the extraction process as an approach to optimising the resulting descriptor. Additionally, multimodal features are formed by combining information from both domains in a single stage in the extraction process. To further improve the quality of feature descriptors, a calculation of the surface normals and a use of the 3D structure from the range image are applied to correct the 3D appearance of a local sample patch, thereby increasing the similarity between observations. The matching performance of local features is evaluated using an experimental setup comprising a turntable and stereo pair of cameras. This experimental setup is used to create a database of intensity and range images for 5 objects imaged at 72 calibrated viewpoints, creating a database of 360 object observations. The use of a calibrated turntable in combination with the 3D object surface coordiantes, supplied by the range image allow location correspondences between object observations to be established; and therefore descriptor matches to be labelled as either true positive or false positive. Applying this methodology to the formulated local features show that two approaches demonstrate state-of-the-art performance, with a ~40% increase in area under ROC curve at a False Positive Rate of 10% when compared with standard SIFT. These approaches are range affine corrected intensity SIFT and element corrected surface gradients SIFT. Furthermore,this work uses the 3D structure encoded in the range image to organise collections of interest points from a series of observations into a collection of canonical views in a new model local feature. The canonical views for a interest point are stored in a view compartmentalised structure which allows the appearance of a local interest point to be characterised across the view sphere. Each canonical view is assigned a confidence measure based on the 3D pose of the interest point at observation, this confidence measure is then used to match similar canonical views of model and query interest points thereby achieving a pose invariant interest point description. This approach does not produce a statistically significant performance increase. However, does contribute a validated methodology for combining multiple descriptors with differing confidence weightings into a single keypoint. |
