L. Matthies, R. Szeliski, T. Kanade, "Kalman Filter-based Algorithms for Estimating Depth from Image Sequences", CMU Tech. Report, CMU-CS-87-185, December 1987.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....temporal accumulation of information through a monocular observer. The relationship between subsequent still images in a video stream provides a wealth of information in the form of spatio temporal change. The temporal integration of such velocity fields is essential for solving shape from motion [6, 16, 13, 42, 46, 54, 70], timeof collision [18] object tracking [51] object recognition [4] and figure ground separation problems. At first glance the problem of 3 D reconstruction from motion images seems trivial as it is intuitively sound to suggest that changes in intensity on an image plane are somewhat coupled ....

....the surface must be known a priori. This approach is in general not realistic for an autonomous robot that cannot afford to perform large motions without attending to the scene and thus running the risk of a critical collision. The second approach to depth estimation is the iconic depth estimator [30, 46, 60, 70] in which all pixels contribute a depth estimate. This approach is more suitable for a navigating robot as it lends itself to small motions between viewpoints. Thus, the image and depth correspondence problems are locally constrained and facilitated, and a dense disparity field is obtained. The ....

[Article contains additional citation context not shown here]

Matthies, L., Kanade, T. and Szeliski, R., Kalman Filter-based Algorithms for Estimating Depth from Image Sequences, International Journal of Computer Vision, 3, 209-236, 1989.

....and, indeed, it opens a new and exciting avenue of research in nonlinear systems theory. Although the first formulations of the visual motion estimation problem date back to the beginning of the century [31, 77] only within recent years tools from control and estimation theory have been applied [3, 8, 9, 28, 29, 35, 52, 56, 60, 66], with rather encouraging results in traditionally difficult applications, such as autonomous vehicle navigation [18, 19, 20] visionbased tracking and servo [12, 21, 42, 44] vision based manipulation [5, 21, 42] docking [19, 37] vision based planning [14] active sensing [69] As the ....

....may be further classified in terms of the data processing technique as 2 frames schemes (see for example [38, 49, 78] multiframe batch methods [70, 75] or recursive algorithms. In the last decade a variety of schemes has been proposed for recursively reconstructing structure for known motion [52], motion for known structure [9, 28, 29] or both structure and motion [3, 35, 56, 60, 66] In general, given either the relative motion or the shape of the object being viewed, the other can be recovered easily, since the problem can be reduced to a linear estimation task. When neither the motion ....

[Article contains additional citation context not shown here]

L. Matthies, R. Szeliski, and T. Kanade. Kalman filter-based algorithms for estimating depth from image sequences. Int. J. of computer vision, 1989.

....di#erent gain and bias. Since our work is novel primarily in terms of the matching cost, we will focus on related work that addresses this. Readers are referred to [14] for a survey and taxonomy of stereo. The most common matching costs include the sum of L di#erences [16] L di#erences [11], or truncated di#erences [1] These costs are sensitive to camera gain and bias. It is also possible to first compute a local quantity that is insensitive to gain and bias and then perform correlation. This has been done using ordering information [21] or properties of the intensity gradient ....

L.H. Matthies, R. Szeliski, and T. Kanade. Kalman filter-based algorithms for estimating depth from image sequences. International Journal of Computer Vision, 3(3):209--238, September 1989.

....emerged. 1.3.3 Prior Motion Estimation Work The literature on determining the motion of a camera and or the geometry of a scene is extensive. Motion can be recovered from a known scene [48] and this problem is related to visual odometry. Scene structure can be determined from camera motion [45][32]. Shape and motion can also be determined simultaneously [44] and all shape and motion assumptions seem ultimately unnecessary. 1.3.4 Prior Visual Tracking Work Once a camera is permitted to move relative to a scene, one can observe correspondence or flow [40] For correspondence, the related ....

L. Matthies, T. Kanade, and R. Szeliski. Kalman filter based algorithms for estimating depth from image sequences. International Journal of Computer Vision, 3(3):209-236, September 1989.

....Spatial grouping allows a significant reduction of complexity, since points need not be detected and tracked individually. 1. 1 Relation to previous work The present work falls within the category of structure from motion, a field that encompasses a vast variety of research efforts, such as [1, 3, 5, 9, 10, 12, 14, 15, 16, 18, 19, 21, 22, 23, 24, 25, 28, 29, 31]. Of all the work in SFM, we consider in particular causal estimation algorithms. A batch approach would obviously perform better, but at the expense of compromising the usability for control actions such as manipulation, navigation or, more in general, real time interaction. Since we integrate ....

L. H. Matthies, R. Szeliski, and T. Kanade, Kalman filter-based algorithms for estimating depth from image sequences, International Journal of Computer Vision 3, no. 3, 209--238, September 1989.

....when executing obstacle avoidance maneuvers during a long traverse. In such cases, 3D registration of successive range maps can be used to estimate rover motion. Previous work in this area includes maximum likelihood estimation of motion coupled with a Kalman filter update of rover position [20, 21]. A recent method using probabilistic map matching based on a maximum likelihood map similarity estimator was developed by Olson [23, and references therein] for rover localization. His studies, which only included translations in the environment, indicated that sub pixel accuracy within a ....

L. Matthies and T. Kanade, "Kalman filter-based algorithms for estimating depth from image sequences," International J. Computer Vision, Vol. 3, pp. 209-236, 1989.

....Base Station Antenna Fig. 2. The experimental setup. provided by the EKF based scheme. III. The terrain mapping algorithm Incremental terrain mapping takes place via four main processes. An incremental dense depth from camera motion algorithm (which is an adaptation of the work reported in [12]) is used to obtain the depth to various features in the environment. The relative pose of the robots at these locations are associated with particular depth information. An elevation gradient of the terrain is determined by fusing GPS altitude information and vertical displacements obtained from ....

....The variance associated with the optical flow is determined by fitting the flow corresponding to the smallest SSD and its two nearest neighbors to a parabola. Then the variance is 2 where a is the coefficient of the quadratic term in a the parabola and a, is the uncertainty in pixel posi tion [12]. Raw disparity is computed from (1) while the predicted disparity and variance are computed as a = H PXH x respectively. Pm is the covariance matrix obtained from the optical flow computation and ax: V(x,y) x,y) U: x,y) The predicted and raw disparities are fused using a Kalman filter, ....

L. Matthies, T. Kemade and R. Szelinski, "Kalman Filter- based Algorithms for Estimating Depth from Image Sequences, " International Journal of Computer Vision, vol. 3, pp. 209-236, 1989.

....that compute optic flow dynamically and refine the flow estimates incrementally over an image sequence. For instance, Singh [39] uses a Kalman filter base approach [42] to estimate optical flow incrementally. There are a number of other analogous approaches for estimating depth from motion [43, 44]. While there are problems with the Kalman filter approach, it brings us closer to the objective of dynamic optical flow. An alternative incremental minimization approach [31] uses a robust formulation and solves the difficult minimization problem incrementally over a sequence of images. This ....

L. Matthies, R. Szeliski, and T. Kanade, Kalman filter-based algorithms for estimating depth from image sequences, Int. J. Cornput. Vision 3(3), Sep. 1989, 209-236.

....that compute optic flow dynamically and refine the flow estimates incrementally over an image sequence. For instance, Singh [39] uses a Kalman filter base approach [42] to estimate optical flow incrementally. There are a number of other analogous approaches for estimating depth from motion [43, 44]. While there are problems with the Kalman filter approach, it brings us closer to the objective of dynamic optical flow. An alternative incremental minimization approach [31] uses a robust formulation and solves the difficult minimization problem incrementally over a sequence of images. This ....

L. Matthies, R. Szeliski, and T. Kanade, Kalman filter-based algorithms for estimating depth from image sequences, lnt. J. Campat. Vision 3(3), Sep. 1989, 209-236.

....likely to be brittle to changes in lighting, point of view, etc. Papanikolopoulos [1991] smoothly tracks pre selected features (image patches) It is not known how to select such features auto matically, although this problem has been and will doubtlessly continue to be a subject of investigation [Matthies et al. 1989; Thorpe, 1983] Waxman et al. 1988] saccadi cally track a set of features whose spatial relationship is pre selected, and the object is therefore identifiable. Clark and Ferrier [1988] present a rare binocular tracking system that saccadically tracks the appropriate conjunction of features to ....

L. Matthies, T. Kanade, and R. Szeliski, "Kalman Filter-Based Algorithms for Estimating Depth From Image Sequences," International Journal of Computer Vision, 3:209-236, 1989.

....estimates depth in an on line manner from image sequences taken by a single, moving camera (monocular vision) the algorithm therefore produces the depth image, introduced in the previous Chapter. In formulating such algorithm we have been mostly inspired by a paper by Matthies, Kanade Szeliski [35], although some major modifications to their method have been made that should make our approach more reliable in terms of error variance. First of all, we will derive a formula that relates motion field with the agent s linear and angular velocities. Then we will explain in great detail our ....

....the components of motion field (f x , f y ) where all quantities (except for focal length) have to be intended as functions of time; a similar expression holds for y(t) Globally we will have that: 2 xy x 2 5 (4. 53) which is also known in the literature [35] as equation of optical flow : in fact vector field [ x, y] represents motion field f , introduced in the last section: f x f y or, in a more rigorous manner: f x (#(P, t) t) f y (#(P, t) t) where # = # 1 , # 2 ] is the projection operator we introduced in Chapter ....

[Article contains additional citation context not shown here]

L. Matthies, T. Kanade, and R. Szeliski. Kalman filter-based algorithms for estimating depth from image sequences. International Journal of Computer Vision, 3(3):209--238, 1989.

.... algorithms resemble such iterative algorithms, but typically operate on an image pyramid, where results from coarser levels are used to constrain a more local search at finer levels [126, 90, 11] The most common pixel based matching costs include squared intensity differences (SD) [51, 1, 77, 107] and absolute intensity differences (AD) 58] In the video processing community, these matching criteria are referred to as the mean squared error (MSE) and mean absolute difference (MAD) measures; the term displaced frame difference is also often used [118] More recently, robust measures, ....

....after the initial discrete correspondence stage. An alternative is to simply start with more discrete disparity levels. Sub pixel disparity estimates can be computed in a variety of ways, including iterative gradient descent and fitting a curve to the matching costs at discrete disparity levels [93, 71, 122, 77, 60]. This provides an easy way to increase the resolution of a stereo algorithm with little additional computation. However, to work well, the intensities being matched must vary smoothly, and the regions over which these estimates are computed must be on the same (correct) surface. Recently, some ....

[Article contains additional citation context not shown here]

L. Matthies, R. Szeliski, and T. Kanade. Kalman filter-based algorithms for estimating depth from image sequences. IJCV, 3:209--236, 1989.

....of pixels or regions across different images. A typical error measure is the RGB or intensity difference between images (these differences can be squared, or robust measures can be used) Some methods compute subpixel disparities by computing the analytic minimum of the local error surface [MSK89] or using gradientbased techniques [LK81, ST94, SC97] Birchfield and Tomasi [BT98] measure pixel dissimilarity by taking the minimum difference between a pixel in one image and the interpolated intensity function in the other image. 1.1.2 Aggregation method The aggregation method refers to the ....

....techniques choose the disparity with the minimum SSSD error, which measures the degree of photoconsistency at a hypothesized depth. The best match can also be assigned a local confidence computed using the variance (across disparity) of the SSSD error function within the vicinity of the best match [MSK89]. While window based techniques work well in textured regions and away from depth discontinuities or occlusions, they run into problems in other cases. Figure 2 shows how a symmetric (centered) window may lead to erroneous matching in such regions. Two ways of dealing with this problem are ....

L. H. Matthies, R. Szeliski, and T. Kanade. Kalman filter-based algorithms for estimating depth from image sequences. International Journal of Computer Vision, 3:209--236, 1989.

....optimization algorithms [46, 63, 83] Hierarchical (coarse to fine) algorithms resemble such iterative algorithms, but typically operate on an image pyramid [80, 58, 7] 3.1. Matching cost computation The most common pixel based matching costs include squared intensity differences (SSD) [34, 1, 48, 68], and absolute intensity differences (SAD) 39] More recently, robust measures, including truncated quadratics and contaminated Gaussians have been proposed [11, 12, 63] These measures are useful because they limit the influence of mismatches during aggregation. Other traditional matching ....

.... view synthesis results (the scene appears to be made up of many thin shearing layers) To remedy this situation, sub pixel disparity estimates can be computed in a variety of ways, including iterative gradient descent and fitting a curve to the matching costs at discrete disparity levels [76, 48, 40]. This provides an easy way to increase the resolution of a stereo algorithm with little additional computation. However, to work well, the intensities being matched must vary smoothly, and the regions over which these estimates are computed must be on the same (correct) surface. Recently, some ....

L. Matthies, R. Szeliski, and T Kanade. Kalman filter-based algorithms for estimating depth from image sequences. IJCV, 3:209-236, 1989.

....well structured environments consisting of man made objects, but lack representative power in complex domains. Pixel based models represent depth at each pixel in the image. Statistical formulations of pixelbased depth estimation, using random field models of the depth map, have been presented in [14, 19, 26]. Pixelbased depth models promise more generality than feature based models; however, much remains to be done on both mathematical and system aspects of this approach. The central system issue is how to find stereo correspondences efficiently and reliably. There are two types of approach: ....

....about surface shape, in particular assumptions about local smoothness [4, 13, 23, 25, 27, 30] and those that obtain constraint by augmenting the sensor, in particular by using redundant images. Redundant images can come from trinocular camera systems [21, 25] fine motion image sequences [3, 19], or the use of fine motion to initialize stereo fusion [8] Redundant sensing is the more effective of the two types of approach; however, questions remain about which sensing strategy is the most effective, about how to formulate the matching problem for a given sensing strategy, and about how ....

L. H. Matthies, R. Szeliski, and T. Kanade. Kalman filter- based algorithms for estimating depth from image sequences. lnternationl Journal of Computer Vision, 3:209-236, 1989.

....techniques choose the disparity with the minimum SSSD error, which measures the degree of photoconsistency at a hypothesized depth. The best match can also be assigned a local confidence computed using the variance (across disparity) of the SSSD error function within the vicinity of the best match [16]. While window based techniques work well in textured regions and away from depth discontinuities or occlusions, they run into problems in other cases. Figure 2 shows how a symmetric (centered) windowmay lead to erroneous matching in such regions. Two ways of dealing with this problem are ....

L. H. Matthies, R. Szeliski, and T. Kanade. Kalman filter-based algorithms for estimating depth from image sequences. Intl. J. Comp. Vision, 3:209--236, 1989.

No context found.

L. Matthies, R. Szeliski, T. Kanade, "Kalman Filter-based Algorithms for Estimating Depth from Image Sequences", CMU Tech. Report, CMU-CS-87-185, December 1987.

No context found.

L. Matthies, R. Szeliski, T. Kanade, "Kalman Filter-based Algorithms for Estimating Depth from Image Sequences", CMU Tech. Report, CMU-CS-87-185, December 1987.

No context found.

Matthies, L., R. Szeliski, and T. Kanade, "Kalman Filter-based Algorithms for Estimating Depth from Image Sequences", CMU Tech. Report, CMU-CS-87-185, December 1987.

No context found.

Matthies, L., Kanade, T., and Szeliski, R., "Kalman Filter-based Algorithms for Estimating Depth from Image Sequences", Intl. J. of Computer Vision, 3, p2092.

No context found.

L. Matthies, R. Szeliski, and T. Kanade. Kalman filterbased algorithms for estimating depth from image sequences. IJCV, 3(3):209--238, 1989.

No context found.

Matthies, L., Kanade T. and Szeliski R., "Kalman Filter-based Algorithms for Estimating Depth from Image Sequences", International Journal of Computer Vision, 3, 1989, pp. 209-238.

No context found.

Matthies, L., Kanade, T. and Szeliski, R., Kalman Filter-based Algorithms for Estimating Depth from Image Sequences, International Journal of Computer Vision, 1989. pp. 209-236.

No context found.

L. Matthies, T. Kanade, and R. Szeliski. Kalman filter-based algorithms for estimating depth from image sequences. International Journal of Computer Vision, 3(3):209--236, September 1989.

No context found.

Matthies, L., Kanade, T. and Szeliski, R.: Kalman Filter-based Algorithms for Estimating Depth from Image Sequences, International Journal of Computer Vision, 3, 209-236, 1989.

First 50 documents  Next 50

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC