In this paper, we present a novel method to detect and resolve motion conflicts in visual-inertial odometry. Recently, it has been common to integrate an IMU sensor with visual odometry in order to improve localization accuracy and robustness. However, when a disagreement between the two sensor modalities occurs, the localization accuracy reduces drastically and leads to irreversible errors. In such conditions, multiple motion estimates based on the set of observations used are possible. This creates a conflict (motion conflict) in determining which observations to use for accurate ego-motion estimation. Therefore, we present a method to detect motion conflicts based on per-frame positional estimate discrepancy and per-landmark reprojection errors. Additionally, we also present a method to resolve motion conflicts by eliminating inconsistent IMU and landmark measurements. Finally, we implement Motion Conflict aware Visual Inertial Odometry (MC-VIO) by combining both detection and resolution of motion conflicts. We perform quantitative and qualitative evaluation of MC-VIO on visually and inertially challenging datasets. Experimental results indicate that the MC-VIO algorithm reduces the increase in absolute trajectory error by 80% and the relative pose error by 60% for scenes with motion conflict, in comparison to the state-of-the-art reference VIO algorithm.

We present an FPGA-based co-processor for accelerating computations associated with Singular Value Array Reconciliation Tomography (SART), a recently developed method for RF source localization. The co-processor allows this relatively complex computational task to be performed using less hardware and less power than would be required by a microprocessor-based computing cluster with comparable throughput and accuracy. The architecture exploits parallelism of the SART algorithm at many levels in order to efficiently map it into the FPGA platform. The system has been developed in VHDL, and implemented on a Virtex-4 SX55 FPGA. Compared to a Pentium 4 CPU running at 3 GHz, use of the co-processor system provides a speed-up of about 6 times for the current signal matrix size of 128-by-16. Greater speed-ups will be obtained when larger matrices are processed because the co-processor reduces the complexity of a m-by-n SVD from O(mn2), to O(mn). Even larger speed-ups may be obtained by using multiple devices in parallel. The system is capable of computing the SART metric to an accuracy of about -145 dB (0.2 ppm) with respect to its true value. This level of accuracy, which is better than that obtained using single precision floating point arithmetic, allows even relatively weak signals to make a meaningful contribution to the final SART solution.

Asynchronous transfer mode (ATM) technology performance is evaluated for Class C (variable bit rate, asynchronous) data transfer, as would be typically found when migrating legacy applications to an ATM environment. Of particular interest are the end-to-end latencies obtainable using ATM networks to transfer time critical data. Latencies obtained using classical IP over ATM and raw ATM using the Fore ATM API are compared to performance of traditional TCP/IP on equivalent FDDI networks. It is shown that at OC-3, in its simplest configuration of back to back nodes, classical IP over ATM latencies are greater than those obtained using FDDI, particularly at small message sizes. However, the results show that IP over ATM via ATM switches can exhibit better performance than FDDI switches or routers, due to the optimization of the ATM switch for fixed size cells. Thus, these benchmarks confirm that ATM provides a good foundation for building high performance IP internetworks. A discussion of trade-offs associated with using FDDI versus ATM for mission critical, real time data transfer applications is also included.

This paper introduces a new method for adaptively building a multivariate C/spl deg/ spline approximation from scattered samples of an unknown function. The central feature of the method is a means for adaptively tesselating an approximation space to form a multidimensional mesh over which the spline fitting then occurs. The mesh used is a Delaunay tesselation of the approximation space whose vertices lie at a subset of the scattered sample locations. The specific subset of sample locations used is adaptively determined by repeated overfitting and simplification of the resulting spline approximation. Overfitting and simplification is an attractive paradigm for high-dimensional approximation problems because it provides a means for forming an approximation that is complex only in regions where the scattered sample data provide sufficient evidence of complexity in the underlying unknown function. Overfitting and simplification is effectively exploited in this new approach as the function representation used is not subject to certain recursive dependencies. The properties of the new technique are demonstrated in the context of an easily visualized bivariate classification problem, The technique is then applied to a 10-dimensional clinical ECG classification problem, and the results are compared to those obtained with a perceptron based neural network.>

A method for finding the affine transformation which relates two projections of a planar curve, arbitrarily oriented in 3-space, is presented. The method is based on a canonic parameterization of curves and the decomposition of these curves into component ellipses such that an arbitrary linear transformation applied to the whole scene also applies to the component ellipses. The relationship between the ellipses allows for a simple solution for the affine transformation relating views of a contour without point correspondence information. Experimental results indicated the need for contour smoothing in the presence of noise; hence filtering in the Fourier domain was applied to the contours prior to parameterization. The theoretical development of this technique and experimental results are presented.

A method for determination of the affine transform relating projections of planar contour features is developed and tested experimentally. The method does not require knowledge of any point correspondences and is applicable to general affine transforms including those resulting from rotations of contours out of the plane of projection. This algorithm, together with a feature extraction procedure to derive contours (open or closed loops) from images of the complete object could form the basis of a system for determination of orientation and identity of a wide class of objects from 2-D views. The technique is based on the relation of moment tensors derived from two views of the planar contours, and determines the affine transform by solution of a set of linear equations.

A method is developed by which images resulting from orthogonal projection of rigid planar-patch objects arbitrarily oriented in three-dimensional (3-D) space may be used to form systems of linear equations which are solved for the affine transform relating the images. The technique is applicable to complete images and to unlabeled feature sets derived from images, and with small modification may be used to transform images of unknown objects such that they represent images of those objects from a known orientation, for use in object identification. No knowledge of point correspondence between images is required. Theoretical development of the method and experimental results are presented. The method is shown to be computationally efficient, requiring O(N) multiplications and additions where, depending on the computation algorithm, N may equal the number of object or edge picture elements.

The problem of identification of rigid planar-patch objects from images projected from random orientations in 3-D space is approached using tensor representation of image measures. This allows standardization of object orientation without knowledge of object identity, and hence permits a simple one-step comparison operation between the standardized unknown and each library image. Geometric transformations due to oblique viewing angles, as well as translation, rotation, and scale change are handled. Experimental results using a library of camera-acquired images demonstrate correct performance of the algorithm.

An often encountered pattern recognition problem is that of identifying a given image as related to one of a set of prototypes even though they may differ by virtue of the action of a non-linear mapping of image points. Certain parameterized distortions are such that a transform may be found that has the property of isolating the distortion effect as a factor in the transform space allowing subsequent identification and removal of the distortion. Herein the most general form of distortion is found that allows a related transform with such properties to be constructed. Furthermore a test for the admissibility of a distortion is given as well as the general form of the required parameter isolating transform.