This paper describes a new approach to tracking independently moving objects from a moving camera. A background filtering approach is adapted to non-static backgrounds using a motion compensation scheme and gives reliable tracking at superior frame rates to previous approaches. A filtering and object detection system is then described which converts the object mask into a list of objects present in each frame. This list is then passed to a CONDENSATION based tracking system which forms the final stage of the process. Testing the algorithm reveals it is successful detecting and tracking moving objects against non-static backgrounds

A novel approach to the automatic extraction of geological faults from three-dimensional (3-D) seismic data is described, and qualitative and quantitative comparisons of manually and automatically picked fault geometries interpreted from both high and medium quality 3-D seismic datasets are presented. An algorithm has been developed that allows semiautomated identification, extraction, and modeling of fault surfaces imaged in 3-D seismic datasets. Based on a multistage approach, the algorithm operates initially at a small spatial scale, identifying local discontinuities in the seismic horizons, and then gradually considers larger and larger segments of fault surfaces until a set of complete fault surfaces are identified. A large portion of the work involves merging of segments of fault surfaces, performed using a highest confidence first (HCF) stratagem, taking into consideration the context of the resultant fault geometry. We show that results from the automated fault picker compare favorably with a manually labeled set of faults surfaces interpreted from a high-quality dataset. Last, we present an estimate of the savings in human operator time that can be made by using the automated approach, indicating savings of multiple person-days for the multigigabyte datasets that typify the petroleum industry.

A new method for the compression of angiogram video sequences is presented. The method is based on the philosophy that diagnostically significant areas of the image should be allocated the greatest proportion of the total allocated bit budget. The approach uses a three-dimensional wavelet-coder based on the popular set partitioning in hierarchical trees algorithm. Incorporated into this framework are a region-of-interest (ROI) detection stage and a texture-modeling stage. The combined result is an approach that models the high-frequency wavelet coefficients for some diagnostically unimportant regions of the image in an extremely efficient manner. This allows additional bits to be used within the ROI to improve the quality of the diagnostically significant areas. Results are compared for a number of real data sets and evaluated by trained cardiologists.

This paper presents an approach to the problem of estimating a dense optical flow field. The approach is based on a multiframe, irregularly spaced motion trajectory set, where each trajectory describes the motion of a given point as a function of time. From this motion trajectory set a dense flow field is estimated using a process of interpolation. A set of localized motion models are estimated, with each pixel labeled as belonging to one of the motion models. A Markov random field framework is adopted, allowing the incorporation of contextual constraints to encourage region-like structures. The approach is compared with a number of conventional optical flow estimation algorithms taken over a number of real and synthetic sequences. Results indicate that the method produces more accurate results for sequences with known ground truth flow. Also, applying the method to real sequences with unknown flow results in lower DFD, for all of the sequences tested.

Summary form only given. A novel method for the compression of angiogram video sequences is presented. The approach consists of a 3D wavelet encoding algorithm, incorporating a region of interest (ROI) estimation model to provide higher-quality image reconstruction in areas considered to be diagnostically significant. This is coupled with a texture modelling procedure, which provides a model for some of the high frequency wavelet coefficients corresponding to diagnostically insignificant regions. This allows some of the wavelet coefficients to be represented extremely efficiently. The results are evaluated by trained cardiologists and show some promise.

This paper presents a region-based motion estimation algorithm for use in video compression systems. It is based on overlapping block motion estimation with a least-median of squares (LMedS) error measure to estimate the initial motion field. The overlapping blocks improve the motion estimate for each pixel. The block-based motion field is subsequently segmented into various regions of coherent motion and the segmentation is refined in a rate-distortion framework. Region boundaries are then estimated and refined in an iterative procedure. The results show significantly lower prediction errors compared to other block-based approaches.

This paper describes an unsupervised region merging technique based on a novel robust statistical test. The merging decision is derived from the mutual inlier ratio (MIR) of adjacent regions. This ratio is computed using robust regression techniques and a novel method to estimate the robust scale of the Gaussian distribution. A discrimination value to recognize identical Gaussian distributions with the MIR is derived theoretically as a function of the sizes of the compared sets. The presented method to test distributions is compared with the established Kolmogorov-Smirnov test and implemented into a segmentation algorithm for planar range images. The iterative region growing technique is evaluated using an established framework for range image segmentation comparison involving 60 real range images. The evaluation incorporates a comparison with four state-of-the-art algorithms and gives an experimental demonstration of the need for robust methods capable of handling noisy data in real applications.

This paper presents a new approach for the estimation of motion trajectories from image sequences. This approach is seen to be similar to both "token tracking" feature-based methods, and "block based" region-matching approaches. Initially, a parametric motion trajectory model is formulated, and the motion trajectory estimation problem is conveniently cast into a Markov Random Field (MRF) framework which allows a priori motion constraints to be expressed. A three-stage optimization algorithm is presented, and a recursive estimation approach is developed. A method for the robust detection of areas of occlusion is presented, as are comparative results to show the relative success of this motion estimation approach. Although robustly computed motion trajectories show great promise in many motion-related areas of computer vision, this paper illustrates their use in the context of motion compensated prediction as part of a video compression algorithm.

An unsupervised method to extract 2D and 3D inner earth structures from seismic reflection measurements is described. The application is a typical texture segmentation problem, which can be split up into a feature extraction stage and a segmentation stage. As a texture feature, the locally emergent frequency is estimated by a Gabor filter bank. The instantaneous frequency (IF) has already been successfully used for seismic trace analysis21 and will be compared with the results of the filter bank. The second stage of the algorithm involves a region-growing method to compute the final object structure. The extremely flexible segmentation scheme is appropriate for application to 2D and 3D images of arbitrary vectorial dimension. The merging decision is based on the mutual inlier ratio of two adjacent regions. This ratio is computed by robust regression techniques19 to avoid noise artifacts. A mutual inlier ratio discrimination function to recognize identical Gaussian distributions, guaranteeing a 97.5% certainty, is derived. This method is compared with the Kolmogorov–Smirnov test and results of the application in a segmentation algorithm are shown. The segmentation stage is also tested with different benchmark data sets from other computer vision problems to demonstrate its general flexibility.

This paper presents a new method to extract the 3D shape of objects from 3D gray level images using a bicubic B-spline surface model. Extraction of object shape is achieved through a hierarchical surface fitting by exploiting the multi-scale representation of the model. A strategy for converting the surface estimation into curve estimations is devised. The model surface is estimated by successively computing a set of cubic B-spline curves consisting of a coordinate curve net defining the surface. A regularising component is incorporated into the curve estimation to encourage the generation of an orthogonal coordinate curve net, preventing the creation of unwanted creases. Experimental results are presented for extracting the 3D shape of objects from real 3D images. 1 Introduction The interpretation of 3D images often needs the shape information of objects in the image. A set of unmodelled 3D structures derived from local low level operations (e.g. edge detection) [10] can hardly ever ...

This paper presents a method for extracting the 3D shape of objects from 3D gray level images. The object surface is assumed to be described by a bicubic B-spline surface. The detection of the object shape is achieved using a hierarchical approach in which the object surface is successively estimated using the spline model at increased scales. At each scale, the model is iteratively fitted to a set of measurement points which are believed to be on the object surface. The measurement points are detected by searching for voxels with the strongest edge response along the direction normal to the model surface. Searching for the measurement points is controlled within a search radius determined based on the edge strength at the points on the current model surface. The result is a parametric surface approximating the object surface to be detected.

This paper presents an unsupervised segmentation method applicable to both 2D and 3D images. The segmentation is achieved by a bottom-up hierarchical analysis to progressively agglomerate pixels/voxels in the image into non-overlapped homogeneous regions characterised by a linear signal model. A hierarchy of adjacency graphs is used to describe agglomeration results from the hierarchical analysis, and is constructed by successively performing a clustering operation which produces an optimal classification by merging each region with its nearest neighbours determined under the framework of statistical inference. The top level of the hierarchy then describes the segmentation result.

This paper presents an algorithm that provides good estimates of the optical flow at motion discontinuities and occlusion regions based on the least-median-squares robust estimator. The use of a multiresolution scheme enables optical flow to be computed for objects with large motions. A movable overlapping neighbourhood strategy eliminates block-effects and also deals with cases of the local neighbourhood straddling regions of two or three motions. Results for both synthetic and real image sequences are presented. A quantitive comparison with the matching technique of Anandan (1989) and the differential technique of Odobez and Bouthemy (see Proceedings of 7th European Conference on Signal Processing, 1994) shows that the proposed algorithm gives much lower angular and absolute errors.

This paper is concerned with the segmentation of scene objects on the basis of their unique uniform motions. A number of previous approaches have been founded upon greyscale spatio-temporal gradient based estimation of the optic flow; these have shown some success. However optical flow only permits a limited range of recoverable motion displacements and exhibits a relatively low robustness to noise. Multiresolution image data can be used to increase the range of allowed motion displacements but the correct resolution at which to compute motion estimates is difficult to determine. It is postulated that with a priori knowledge of the elementary motions arising from the dynamic scene, the resolution level of a multiresolution support can be automatically set. These elementary motions may be used to increase noise robustness by permitting a relative rather than absolute classification of motion. We present a multi-stage algorithm in which feature correspondences are used to create a dictionary of elementary motions arising from the scene. The scene is initially segmented into small primitive regions using a maximum a posteriori (MAP) criterion in conjunction with a Markov random field (MRF) model and the motion dictionary. An affine motion model and maximum likelihood (ML) criterion are then used to fuse primitive regions of coherent motion into the full scene segmentation. Results for both synthetic and real imagery are given which demonstrate that scene segmentation may be performed across a wide range of motion displacements and at high levels of additive noise.

This paper presents an algorithm to compute optical flow accurately at motion discontinuities and occlusion regions based on a robust estimator (the Least-Median-Squares estimator). The motion constraint equation and the 2-D affine motion model are used to compute the optical flow in a local neighbourhood. The use of the least-median-squares robust estimator enables points where optical flow cannot be computed to be rejected as outliers rather than assigning erroneous flow to such points. In addition, the use of an overlapping neighbourhood strategy eliminates the block-effects that are commonly faced in local differential methods for computing optical flow. The algorithm is also able to deal with cases of the local neighbourhood straddling regions of three motions. Results for both synthetic and real image sequences are presented. 1.

A problem often encountered with multiresolution segmentation algorithms is that small or thin features of an object become lost. This is particularly evident in linked-pyramid structures and a method is required to restore these features. This paper shows that simple adaptive isotropic and non-isotropic filtering based on the inter-region signal-to-noise ratio can be used to iteratively re-establish the lost features. It is also shown that boundary placement is also improved and a balance between class certainty and boundary placement is achieved. A number of results are presented for synthetic and real images.

For pt.I see ibid., vol.9, no.6, p.787 (1987). The problem of uncertainty in image feature description is discussed, and it is shown how finite prolate spheroidal sequences can be used in the construction of feature descriptions that combine spatial and frequency-domain locality in an optimal way. Methods of constructing such optimal feature sets, which are suitable for graphical implementation, are described, and some generalizations of the quadtree concept are presented. These methods are illustrated by examples from image processing applications, including feature extraction and texture description. The problem of image segmentation is discussed, and the importance of scale invariance in overcoming the limitations imposed by uncertainty is demonstrated. A novel texture segmentation algorithm that is based on a combination of the new feature description and multiresolution techniques is described and shown to give accurate segmentations on a range of synthetic and natural textures.>