System and method for aligning video sequences

Abstract

Methods and Systems for aligning multiple video sequences of a similar scene. It is determined which video sequences should be aligned with each other using linear dynamic system (LDS) modeling. The video sequences are then spatially aligned with each other.

Description

BRIEF DESCRIPTION OF THE FIGURES

FIG. 1 illustrates a system for registering video sequences, according to one embodiment.

FIG. 2 illustrates a method for registering video sequences, according to one embodiment.

FIG. 3 sets forth details of determining which images from multiple (e.g., two or more) video sequences should be associated with one another for comparison purposes (205 of FIG. 2), according to one embodiment.

FIG. 4 illustrates sample images from two video sequences, according to one embodiment.

FIGS. 5-6 illustrate sample comparison results, according to several embodiments.

DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION

FIG. 1 illustrates a system for registering video sequences, according to one embodiment. The videos may be from synchronous and/or asynchronous cameras. In addition, in some embodiments, the videos may be of a rigid and/or a non-rigid scene. FIG. 1 illustrates a system 100 for presenting information, according to one embodiment. In one embodiment, server computer 110 can communicate over a network 105 (e.g., Intranet, Internet) with client computer 115. A comparing videos application 120 can run on the server computer 110. In some embodiments, a user can access the comparing videos application 120 utilizing the server computer 110. In other embodiments, the user can access the comparing videos application 120 utilizing the client computer 115, which can communicate with the server computer 110 for this purpose.

The comparing videos application 120 can comprise: model module 125, identify parameters module 135, canonical form module 130, register module 140, and compare module 145. The model module 125 can model each video sequence. In one embodiment, each video sequence can be modeled as the output of a Linear Dynamical System (LDS). The identify parameters module 135 can identify the parameters of the models (e.g., the LDSs from the multiple video sequences). The canonical form module 130 can transform the parameters into a canonical form so that the all parameters from all video sequences can be expressed with respect to the same basis. The associate parameters module 140 can associate scenes from the video sequences using the transformed parameters and register the associated scenes with each other. The compare module 145 can compare the associated images to each other using various image-comparing mechanisms.

FIG. 2 illustrates a method for registering video sequences, according to one embodiment. In 205, it is determined which images from multiple video sequences (e.g., two or more) should be associated with one another. In 210, the associated images can be compared to each other using various image-comparing mechanisms known to those of ordinary skill in the art such as feature based methods or intensity based methods. Further details on such comparison mechanisms can be found in Szeliski, R. , “Image Alignment and Stitching: A Tutorial”, Fundamental Trends in Computer Graphics and Vision, 2(1) (2006) 1-104.

FIG. 3 sets forth details of determining which parameters extracted from multiple (e.g., two or more) video sequences should be associated with one another for comparison purposes (205 of FIG. 2), according to one embodiment.

In 305, each video sequence can be modeled, for example, as the output of a LDS. For example, the video sequence {I(t)}_t=1^F (where I(t) is a p-dimensional vector representing the image frame of the video at time t, p is the number of pixels in each image , and F is the number of frames in the video) can be modeled as the output of a LDS as follows:z(t+1)=Az(t)+Bv(t)I(t)=C⁰+Cz(t)+w(t)

In the above formulas, z(t) is an n-dimensional vector, with n much smaller than p, representing a compressed version of the image at time t. Specifically, z(t) are the coefficients needed to express the image I(t) in terms of the appearance images (columns of C) which represent the appearance of the video, and the mean image C⁰. Together, C⁰ and C form a basis for all the images in a video. This LDS model can decouple the appearance of the video (represented by C⁰ and C) from the temporal evolution of the video (represented by A). Since C⁰ and C are the only model parameters that depend on the image pixels, the spatial registration can be recovered independently from the temporal lag between the video sequences. In addition, Bv(t) and w(t) model the error between the output of the LDS and the image I(t) due to noise. These errors can be ignored (if there is no noise) or approximated (when there is noise).

Note that FIG. 4 illustrates sample columns of the C matrix displayed as images from (a) a first video sequence of a parking set and (b) a second video sequence of the same parking set taken from a different viewpoint.

Referring back to FIG. 3, in 310, parameters of the LDSs from the multiple video sequences can be identified. Given I_i(t) from above, C_i⁰ (which is the temporal image mean of the images in the video) can be calculated. Then, set Ĩ_i(t)=I_i(t)−C_i⁰. If the noise terms (e.g., B, ν, ω) are ignored as explained above, the LDS output model is z(t)=A¹z₀, where z₀ is the initial state of the system. If there is a temporal lag for the ith video sequence of τ_i frames, then the evolution of the hidden state of the ith sequence is z_i(t)=A^τiz(t), and the mean subtracted images are Ĩ_i(t)=C_iA^τiz(t). At this point, the parameters C and Z=[z(1)L z(F)] can be calculated using the singular value decomposition of the matrix W (which represents the image frames from all of the video sequences stacked into a single matrix), which can be factorized using the rank n singular value decomposition (SVD) of the matrix W as follows:

$\begin{array}{c}W=\lfloor \begin{array}{ccc}{\tilde{I}}_l\left(l\right)& L& {\tilde{I}}_l\left(F\right)\\ M& O& M\\ {\tilde{I}}_m\left(l\right)& L& {\tilde{I}}_m\left(F\right)\end{array}\rfloor \\ =\lfloor \begin{array}{ccc}{C}_l{A}^{{\tau }_l}z\left(l\right)& L& C_l{A}^{{\tau }_l}z\left(F\right)\\ M& & M\\ C_m{A}^{{\tau }_m}z\left(l\right)& L& C_m{A}^{{\tau }_m}z\left(F\right)\end{array}\rfloor \\ =\lfloor \begin{array}{c}{C}_l{A}^{{\tau }_l}\\ M\\ C_m{A}^{{\tau }_m}\end{array}\rfloor \left[z\left(l\right)Lz\left(F\right)\right]={\mathrm{USV}}^T\end{array}$ where Z=SV^T and C=U . The parameter A can then be computed as follows:A=[z(2), . . . , z(F)][z(1), . . . , z(F−1)]^t

From above, the parameters A and C_i were found. Then, if the p_i by n matrix C_i is in the matrix formed by rows Σ_j=1ⁱ⁻¹p_j+1 to Σ_j=1ⁱp_i of C, the pair (A, C_i) can be converted to a canonical form, as explained below with respect to 315.

Referring back to FIG. 3, in 315, the parameters A and C_i are transformed with respect to a common (canonical) form so that the all parameters A and C_i from all video sequences are expressed with respect to the same basis. In one embodiment, a Jordan Canonical Form (JCF) can be utilized. When A has 2q complex eigenvalues and n−2q real eigenvalues, the JCF is given by:

$A_c=\lfloor \begin{array}{cccccccc}{\sigma }_1& {\omega }_1& 0& & & & & 0\\ -{\omega }_1& {\sigma }_1& & & & & & \\ 0& & O& & & & & \\ & & & {\sigma }_1& {\omega }_1& & & \\ & & & -{\omega }_1& {\sigma }_1& & & \\ & & & & & {\lambda }_1& & \\ & & & & & & O& \\ 0& & & & & & & {\lambda }_{n-2q}\end{array}\rfloor$ $C_c=\left[10L101L1\right]$ where the eigenvalues of A_c are given by {σ₁±√{square root over (−1)}ω₁, L , σ_q±√{square root over (−1)}ω_q, λ₁, L , λ_n-2q}, where σ_i and ω_i are parameters capturing the oscillatory behavior of image intensities in the video, λ_i captures transient behaviors, and q is the number of complex eigenvalues in each image. Additional information on the JCF can be found at W. J. Rugh. Linear System Theory (Prentice Hall, 2d. ed. 1996). It should be noted other canonical forms, which use a different basis (e.g., a Reachability Canonical Form (RCF), an Observability Canonical Form (OCF) can be used in other embodiments. Additional information on the RCF and the OCF can be found at W. J. Rugh, Linear System Theory (Prentice Hall, 2d. ed. 1996).

The procedure for converting the LDS parameters (A, C_i) into any canonical form involves finding an n by n invertible matrix M_I such that M_iAM_i⁻¹, γC_iM_i⁻¹)=(A_c, C_c), where the subscript c represents any canonical form, and the p-dimensional vector γ (where p is the number of pixels) is an arbitrary vector. In one embodiment, γ can be chosen to be [1 1 L 1], so that all rows of C are weighted equally. Once M is found, M can be used to convert LDS (A, C_i) into the canonical using the formula (M_iAM_i⁻¹, C_iM_i⁻¹).

It should be noted that the JCF is unique only up to a permutation of the eigenvalues. However, if the eigenvalues are different, a predefined way can be used to sort the eigenvalues to obtain a unique JCF.

Referring back to 320 of FIG. 3, once the parameters A and C_i from the multiple video sequences are identified and transformed into a canonical form, associated scenes (e.g., static or non-static) can be spatially registered with each other. In the case of two videos, this can be done as follows: The temporal mean images C₁⁰ and C₂⁰ of the two videos were calculated above. Then, the system parameters (A, C₁) and (A, C₂) were converted into a canonical form. At this point, every column of the temporal matrix C_i is converted into its image form. (The notation C_jⁱ can be used to denote the ith column of the jth video sequence represented as an image.) A feature-based approach can then be used to spatially register the two sets of images {C₁⁰, C₁^l, . . . , C₁ⁿ} and {C₂⁰, C₂¹ . . . , C₂ⁿ}. Scale Invariant Feature Transform (SIFT) features and a feature descriptor can be extracted around every feature point in the two sets of n+1 images. SIFT is a feature point that can be obtained by analyzing the gradiants (e.g., spatial) in a small image neighborhood. More information on SIFT can be found in D. Lowe, “Distinctive Image Features from Scale-Invariant Keypoints”, International Journal for Computer Vision, Volume 20 (2003) 91-110. The features extracted from image C₁ⁱ can be matched with the features extracted from image C₂ⁱ, and vice-versa, where i ε{0, . . . , n}. Details about how such features can be extracted and matched can be found at D. Lowe, “Distinctive Image Features from Scale-Invariant Keypoints”, International Journal of Computer Vision, Volume 20 (2003) 91-110. The matches that are consistent across both directions can be retained and the correspondences can be concatenated into the matrices X₁ ε^3×M and X₂ε^3×M.The corresponding columns of X₁ and X₂ are the location of the matched features in homogenous co-ordinates and M is the total matches from the n+1 image pairs. The homography H (such that X₂˜HX₁) is recovered. This can be done by running Random Sampling And Consensus (RANSAC) to obtain the inliers from the matches. RANSAC is a method that can be used to remove outliers from a set. It can be based on randomly sampling points and analyzing how consistent the model obtained with the sampled point is with the rest of the data. More information on RANSAC can be found in M. A. Fischler et al., “RANSAC Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography”, Communications of the ACM, 26 (1981) 381-395. A homography can then be fit using a non-linear method. More information about such a non-linear method can be found in R. Hartley et al., Multiple View Geometry in Computer Vision (Cambridge 2000). This allows the contribution of the correspondences from every image pair to be weighted equally, because the best matches given by RANSAC could arise from the mean image or the dynamic appearance images or both.

As set forth above in FIG. 2, once it is determined which images from the multiple video sequences should be associated with one another, the associated images can be compared to each other using various image-comparing mechanisms known to those of ordinary skill in the art.

FIGS. 5-6 illustrate sample comparison results, according to several embodiments. FIG. 5 illustrates a comparison of results using the initial alignment, a Ravichandran method, a Caspi method, and the method described herein. Additional information on the Ravichandran method can be found in A. Ravichandran et al., “Mosaicing Non-rigid Dynamical Scenes”, Workshop on Dynamic Vision (2007). Additional information on the Caspi method can be found in Y. Caspi et al., “Feature-Based Sequence-to-Sequence Matching”, International Journal of Computer Vision, 68 (1) (2006) 53-64.

FIG. 6 illustrates additional comparisons of the results using our method against other methods. Rows (a-c) are a comparison with the method in Y. Caspi et al., “Spatio-Temporal Alignment of Sequences”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(11) (2002) 1409-1424. Row (d) is a comparison with the method in Y. Caspi et al., Feature-Based Sequence-to-Sequence Matching”, International Journal of Computer Vision, 68(1) (2006) 53-64. Row (e) is a comparison with the method in Y. Ukrainitz et al., “Aligning Sequences and Actions by Maximizing Space-Time Correlations”, European Conference on Computer Vision (2006) 538-550.

While various embodiments of the present invention have been described above, it should be understood that they have been presented by way of example, and not limitation. It will be apparent to persons skilled in the relevant art(s) that various changes in form and detail can be made therein without departing from the spirit and scope of the present invention. Thus, the present invention should not be limited by any of the above-described exemplary embodiments.

For example, multiple examples of different types of video sequences are illustrated in FIGS. 4-6. However, those of ordinary skill in the art will understand that any type of video sequences can be used. For example, medical image video sequences (e.g., a moving heart) can be utilized in one embodiment.

In addition, it should be understood that the figures described above, which highlight the functionality and advantages of the present invention, are presented for example purposes only. The architecture of the present invention is sufficiently flexible and configurable, such that it may be utilized in ways other than that shown in the figures.

Further, the purpose of the Abstract of the Disclosure is to enable the U.S. Patent and Trademark Office and the public generally, and especially the scientists, engineers and practitioners in the art who are not familiar with patent or legal terms or phraseology, to determine quickly from a cursory inspection the nature and essence of the technical disclosure of the application. The Abstract of the Disclosure is not intended to be limiting as to the scope of the present invention in any way.

Finally, it is the applicant's intent that only claims that include the express language “means for” or “step for” be interpreted under 35 U.S.C. 112, paragraph 6. Claims that do not expressly include the phrase “means for” or “step for” are not to be interpreted under 35 U.S.C. 112, paragraph 6.

Claims

1. A system for aligning at least two video sequences of at least one similar non-rigid scene, including, but not limited to: at least one application operable on at least one computer, the at least one application configured for:determining, utilizing at least one comparing videos application operable on at least one computer, which of the at least two video sequences should be aligned with each other, the determining including, but not limited to:modeling each video sequence as output of at least one linear dynamic system (LDS);jointly identifying parameters from at least two LDSs associated to the at least two video sequences of the at least one similar non-rigid scene stacked into a single matrix;transforming the parameters of all LDSs to a canonical form where the parameters of different video sequences are expressed with respect to a same basis;extracting features from the parameters of the LDSs; andusing an image alignment algorithm to find a spatial transformation relating the features in order to spatially align the at least two video sequences with each other, wherein spatial alignment is found independent of temporal alignment.

2. The system of claim 1, wherein any image alignment technique can be utilized.

3. The system of claim 1, wherein the at least two video sequences are from synchronous and/or asynchronous cameras.

4. The system of claim 1, wherein the parameters are transformed into canonical form utilizing the Jordan Canonical Form (JCF).

5. A method for aligning at least two video sequences of at least one similar non-rigid scene, including, but not limited to: determining, utilizing at least one comparing videos application operable on at least one computer, which of the at least two video sequences should be aligned with each other, the determining including, but not limited to:modeling each video sequence as output of at least one linear dynamic system (LDS);identifying parameters from at least two LDSs associated to the at least two video sequences of the at least one similar non-rigid scene;jointly transforming the parameters of all LDSs to a canonical form where the parameters of different video sequences are expressed with respect to a same basis stacked into a single matrix;extracting features from the parameters of the LDSs; andusing an image alignment algorithm to find a spatial transformation relating the features in order to spatially align the at least two video sequences with each other, wherein spatial alignment is found independent of temporal alignment.

6. The method of claim 5, wherein any image alignment technique can be utilized.

7. The method of claim 5, wherein the at least two video sequences are from synchronous and/or asynchronous cameras.

8. The method of claim 5, wherein the parameters are transformed into canonical form utilizing the Jordan Canonical Form (JCF).