No description available.

No description available.

This thesis provides a review of the state-of-the-art in vision systems and methodologies and an introduction of important surface attributes and representations. Then three novel methods for the dynamic inspection of specular freeform surfaces are presented. These comprise two novel machine vision systems as well as a novel high-speed, multi-scale line tracing algorithm. Both of the novel systems employ a reciprocal deflectometric arrangement. The reflection of a laser line from a surface is monitored on a translucent screen. Here, a complex curve, known as the 'specular signature', is formed that contains all the information on the surface. Methods for extracting and interpreting this information are presented and incorporated into the two vision systems. Prototype demonstrators were designed and assembled to verify the presented methodologies. Extensive experimental validations of all three contributions are shown and the results are compared to ground truth data. Statistical validations of the systems are also presented. Also, the optical and angular resolutions as well as the limitations and the allowable ranges of surface characteristics for both systems, were calculated and presented. It is shown that they are applicable to a range of surface geometries and roughnesses that is comparable to those of existing techniques. The first of the two novel systems is designed for the robust and qualitative detection, classification and localisation of surface defects. It was validated using various real defects on specular freeform surfaces. It is shown that any discontinuity on a surface will be detected and can be classified as long as one criterion regarding the smallest radius of concave curvature on the surface is fulfilled. It is known that this criterion will be fulfilled for a very wide range of common surfaces. The second proposed vision system serves the purpose of a complete, quantitative reconstruction and digitalisation of moving, specular freeform surfaces. While the first system only requires the information from the specular signature, the second system also uses traditional and highly inaccurate surface height data, gathered through laser triangulation. These two data sets, computed from the diffuse as well as the specular reflection, are fused together to generate highly accurate surface bump (gradient) maps. Through the reverse engineering of several real specular specimens and the comparison to ground truth, it is shown that the standard deviation of the error of the height map reaches micrometer levels while that of the angular accuracy reaches levels below one degree. As a third original contribution to knowledge, a novel, high speed, multi-scale line extraction algorithm was developed. Intended for the rapid extraction of the specular signature from the screen images, it combines the processing speed of crude edge detectors with the versatility and accuracy of complex differential geometrical line extractors. It is also multi- scale, with best match scale space being chosen fully automatically. By combining the formerly separated steps of line point detection and line point linkage, the new algorithm is able to increase the processing speed of existing line extractors by up to 50 times. The time requirement is of the same order of magnitude as for crude edge detection algorithms such as Canny. The novel algorithm can also be implemented without the need for any global thresholds as it defines itself a variable local threshold, thereby increasing the sensitivity drastically.

This thesis is concerned with the development of a computer-aided autonomous navigation system for a visually-impaired person. The system is intended to work in both indoor and outdoor locations and is based around the use of camera systems and computer vision. Following a review of the literature to identify previous work in navigation systems for the blind, the location of accurate image features is shown to be a vital importance for a vision based navigation system. There are many operators that identify image features and it is shown that existing methods for identifying which has the best performance are inconsistent. A statistically valid evaluation and comparison methodology is established, centered around the use of McNemar's test and ANOVA. It is shown that these statistical tests require a larger number of test images than is commonly used in the literature to establish which feature operators perform best. A ranking of feature operators is produced based on this rigorous statistical approach and compared with similar rankings in the literature. Corner detectors are especially useful for a navigation system because they identify the boundaries of obstacles. However, the results from our testing suggest that the internal angle of a corner is one factor in determining whether a corner is detected correctly. Hence an in-depth study of angular sensitivity of corners is presented. This leads to the development of a pair of descriptors, known as CMIE and AMIE, which describe corners. Experiments show that these descriptors are able to be computed at video rate and are effective at matching corners in successive frames of video sequences. Finally, a complete navigation system is presented. This makes use of both a conventional colour camera and a depth sensor combined in a device known as the Microsoft Kinect. It is shown that the system performs robustly in both indoor and outdoor environments, giving audio feedback to the user when an obstacle is detected. Audio instructions for obstacle avoidance are also given. Testing of the system by both blindfolded and blind users demonstrates its effectiveness.

Abstract This thesis presents a novel method of audio-visual fusion, known as multi- modal optimal feature fusion (MOFF), for person identification where both the speech and facial modalities may be corrupted, and there is a lack of prior knowl- edge about the corruption. Furthermore, it is assumed there is a limited amount of training data for each modality (e.g., a short training speech segment and a single training facial image for each person). A new multimodal feature rep- resentation and a modified cosine similarity are introduced for combining and comparing bimodal features with limited training data, as well as vastly differing data rates and feature sizes. Similarity-based optimal feature selection and multi- condition training are used to reduce the mismatch between training and testing, thereby making the system robust to unknown bimodal corruption. Low-level feature fusion is performed using optimal feature selection, which automatically changes the weighting given to each modality based on the level of corruption. The framework for robust person identification is also applied to noise robust speaker identification, given very limited training data. Experiments have been carried out on a bimodal data set created from the SPIDRE speaker recogni- tion database and AR face recognition database, with variable noise corruption of speech and occlusion in the face images. Combining both modalities using MOFF, leads to significantly improved identification accuracy compared to the component unimodal systems, even with simultaneous corruption of both modal- ities. A novel piecewise-constant illumination model (PCIlVI) is then introduced for illumination invariant facial recognition. This method can be used given a single training facial image for each person, and assuming no prior knowledge of the illumination conditions of both the training and testing images. Small areas of the face are represented using magnitude Fourier features, which takes advan- tage of the shift-invariance of the magnitude Fourier representation, to increase robustness to small misalignment errors and small facial expression changes. Fi- nally, cosine similarity is used as an illumination invariant similarity measure, to compare small facial areas. Experiments have been carried out on the YaleB, ex- tended YaleB and eMU-PIE facial illumination databases. Facial identification accuracy using PCIlVI is comparable to or exceeds that of the literature.

Texture segmentation and classification are major issues in computer vision that have not yet been fully explored in the framework of irregularly sample data. Unlike well known image restoration techniques, many analysis methods are mainly concerned with obtaining data representation in a feature space and developing effective distance measures for image discrimination, with no interest in reconstructing back the image from the feature space. On avoiding the later, simpler approaches to image analysis may be developed. This thesis constitutes a research on texture analysis for feature extraction, classification and segmentation of irregularly sampled images. In a real scenario, irregularity in the sampling pattern may be a matter of either an inherent problem property, such as in gathering data in geosciences, or a deliberate design, such as retinomorphic sampling. To extend our results to either case, we introduced irregular sampling by investigating the spatial distributions of three sampling patterns. The first pattern is generated from the uniform distribution. The other two sampling patterns consist of inhomogeniously distributed data, with denser concentration towards the middle, to imitate the biological vision paradigm. One follows the Gaussian distribution and the other the log-pollar distribution. In addition, we extend two of the major approaches in image analysis to irregularly sampled data. The first, co-occurrence matrices, is a statistical approach, which is applied to texture classification. The second approach, Gabor analysis, is extended for unsupervised texture segmentation by using the Fourier transform for non-uniformly sampled data. Following a new trend which looks to enhance computer vision with the functionality of human vision, biologically inspired processing was progressively incorporated into our algorithms to the point of proposing a biological paradigm for image segmentation. Finally, we investigate the use of Gabor analysis for 3D irregularly sampled data, and in particular for the segmentation of volumetric seismic data obtained by the oil industry. The results, however, of this study are rather disappointing.

Vision processing is a topic traditionally associated with neurobiology; known to encode, process and interpret visual data most effectively. For example, the human retina; an exquisite sheet of neurobiological wetware, is amongst the most powerful and efficient vision processors known to mankind. With improving integrated technologies, this has generated considerable research interest in the microelectronics community in a quest to develop effective, efficient and robust vision processing hardware with real-time capability. This thesis describes the design of a novel biologically-inspired hybrid analogue/digital vision chip ORASIS1 for centroiding, sizing and counting of enclosed objects. This chip is the first two-dimensional silicon retina capable of centroiding and sizing multiple objects2 in true parallel fashion. Based on a novel distributed architecture, this system achieves ultra-fast and ultra-low power operation in comparison to conventional techniques. Although specifically applied to centroid detection, the generalised architecture in fact presents a new biologically-inspired processing paradigm entitled: distributed asynchronous mixed-signal logic processing. This is applicable to vision and sensory processing applications in general that require processing of large numbers of parallel inputs, normally presenting a computational bottleneck. Apart from the distributed architecture, the specific centroiding algorithm and vision chip other original contributions include: an ultra-low power tunable edge-detection circuit, an adjustable threshold local/global smoothing network and an ON/OFF-adaptive spiking photoreceptor circuit. Finally, a concise yet comprehensive overview of photodiode design methodology is provided for standard CMOS technologies. This aims to form a basic reference from an engineering perspective, bridging together theory with measured results. Furthermore, an approximate photodiode expression is presented, aiming to provide vision chip designers with a basic tool for pre-fabrication calculations.

Image segmentation is an important branch of computer vision. It aims at extracting meaningM objects lying in images either by dividing images 5 into contiguous semantic regions, or by extracting one or more specific objects in images such as left kidney in CT image. The image segmentation task is, in general, very difficult to achieve since natural images are diverse and complex, and the way we perceive them varies according to individuals.

No description available.

This thesis examines the feasibility, development and deployment of visually tracked flashlights as interaction devices. Flashlights are cheap, robust and fun. Most people from adults to children of an early age are familiar with flashlights and can use them to search for, select and illuminate objects and features of interest. Flashlights are available in many shapes, sizes, weights and mountings. Flashlights are particularly appropriate to situations where visitors explore dark places such as the caves, tunnels, cellars and dungeons that can be found in museums, theme parks and other visitor attractions. Techniques are developed by which the location and identity of flashlight projections are recovered from the image sequence supplied by a fixed camera monitoring a target surface. The information recovered is used to trigger audiovisual events in response to users' actions. Early trials with three prototype systems, each built using existing techniques in computer vision, show flashlight interfaces to be feasible both technically and from a usability point of view. Novel methods are developed which allow extraction of descriptions of flashlight projections that are independent of the reflectance of the underlying physical surface. Those descriptions are used to locate and recognise individual flashlights and support a multi-user interface technology. The methods developed form the basis of Enlighten, a software product marketed by the University of Nottingham spinoff company Visible Interactions Ltd. Enlighten is currently is daily use at four sites across the UK. Two patents have been filed (UK Patent Publication Number GB2411957 and US Patent Application Number 10/540,498). The UK patent has been granted, and the US application is under review.