Apparatus and method for providing panoramic images

[0002] The present invention relates to an apparatus, algorithm, and method for stitching different pieces of images of a scene into a panoramic environment map.

[0003] The most common way to electronically represent the real world is with image data. Unlike traditional graph-based systems, there are systems which use panoramic images to construct a virtual world. The major advantage of a system which uses panoramic images is that very vivid and photo-realistic rendering results can be obtained even when using PCs. In addition, the cost of constructing the virtual world is independent of scene complexity. In such systems, panoramic images are stitched together into a panoramic map from several individual images which are acquired by rotating a camera horizontally or vertically. This panoramic map can be used in different applications such as movie special effects, the creation of virtual reality, or games. A typical problem is how to stitch the different pieces of a scene into a larger picture or map. One approach to address this problem is to manually establish correspondences between images to solve unknown parameters of their relative transformation. Because manual methods are tedious for large applications, automatic schemes are preferably used for generating a seamless panoramic image from different pieces of images.

[0004] One proposed approach uses a nonlinear minimization algorithm for automatically stitching panoramic images by minimizing the discrepancy in intensities between images. This approach has the advantage of not requiring easily identifiable features. However, this technique does not guarantee finding the global minimum if the selection of starting points is not proper. Further because the optimization process is time-consuming, the approach is inefficient. In this invention, the domain of images under consideration is panoramic images.

[0005] The invention allows users to generate panoramic images from a sequence of images acquired by a camera rotated about its optical center. In general, the invention combines feature extraction, correlation, and relaxation techniques to get a number of reliable and robust matching pairs used to derive registration parameters. Based on the obtained registration parameters, different pieces of consecutive images can-be stitched together to obtain a seamless panoramic image.

[0006] In a first aspect, a method of merging a pair of images to form a seamless panoramic image includes the following steps. A set of feature points along the edges of the images is extracted, each feature point defining an edge orientation. A set of registration parameters is obtained by determining an initial set of feature points from a first one of the images which matches a set of feature points of a second one of the images. A seamless panoramic image is rendered using the first and second images with the set of registration parameters.

[0007] The invention provides a feature-based approach for automatically stitching panoramic images acquired by a rotated camera and obtaining a set of matching pairs from a set of feature points for registration. Since the feature points are extracted along the edges, each feature point specifies an edge orientation. Because the orientation difference between two panoramic images is relatively small, the difference between edge orientations of two feature points is also small if they are good matches. Based on this assumption, edge information of feature points can be used to eliminate in advance many false matches by checking their orientation difference. Moreover, many unnecessary calculations involving cross-correlation can be screened in advance, thereby significantly reducing the search time needed for obtaining correct matching pairs. After checking, by calculating the value of correlation of the remaining matching pairs, a set of possible matches can be selected with a predefined threshold. The set of possible matches are further verified through a relaxation scheme by calculating the quality of their matches. Once all of the correct matching pairs are found, they are then used to derive registration parameters. In this invention, an iterative scheme is applied to increase the reliability in providing matching results. Since only three iteration or fewer are needed and only a few feature points are involved in the matching pairs, the whole procedure can be accomplished very efficiently. Also, as discussed above, because the orientation difference of two feature points is checked in advance (before matching). Many calculations involving cross-correlation are not required and the efficiency of stitching is significantly improved. Compared with conventional algorithms, the proposed scheme offers improved efficiency and reliability for stitching images.

[0008] Embodiments of this aspect of the invention may include one or more of the following features. In one embodiment, a set of feature points are first extracted through wavelet formations. Among other advantages, the invention uses wavelet transforms to obtain a number of feature points with edge orientations. Such edge information can speed up the entire registration process by eliminating many impossible matches in advance and avoiding many unnecessary calculations of correlation. The method determines a number of reliable and robust matching pairs through relaxation. The method also measures the quality of a matching pair, imposes angle consistency constraint for improving the robustness of registration, and uses a voting concept to get the desired solution from the set of final matching results.

[0009] In other embodiments, the method forms the final panoramic image with the help of the registration results. In particular, the method adjusts the intensity differences between consecutive input images and blends the intensities of adjacent images to obtain a seamless panoramic image. The final panoramic images can then be used to build a virtual world.

[0010] Still other embodiments may include one or more of the following features:

[0011] For example, the method selects a number of feature points through wavelet transforms. Each feature point is associated with an edge orientation so that the speed of the registration process is increased.

[0012] The method uses an angle constraint to construct a set of matching pairs, which are used to obtain reliable matching results through a relaxation and a voting technique. The set of matching results are then used to form the final seamless panoramic image.

[0013] Constructing an initial set of matching pairs for registration includes comparing the edge orientation differences of feature points in one image and its corresponding feature points in another, calculating the values of correlation of each possible matching pair, and thresholding them with a predefined threshold.

[0014] Getting reliable matching results through relaxation and a voting technique includes calculating the quality of a matching pair, imposing angle consistency constraint to filter out impossible matching pairs, updating matching results through relaxation, and using the voting technique to obtain the reliable registration parameters. In addition, it refines the final registration results by using the correlation technique with a proper starting point.

[0015] Forming the final panoramic images includes dynamically adjusting and properly blending the intensity differences between adjacent images.

[0016] In another aspect, the invention features a system for merging pairs of images to form a panoramic image. The system includes an imaging device which, in operation, acquires a series of images, a storage for storing a series of images, a memory which stores computer code, and at least one processor which executes computer code to extract a set of feature points along the edges of the images, each feature point defining an edge orientation and to obtain a set of registration parameters by determining an initial set of feature points from a first one of the images which matches a set of feature points of a second one of the images, and to render a seamless panoramic image using the first and second images with the set of registration parameters.

[0017] Other advantages and features of the invention will become apparent from the following description, including the claims and the drawings.

[0018] FIG. 1 illustrates an approach for generating a virtual panoramic world.

[0019] FIG. 2 shows the geometry for cylindrical mapping.

[0020] FIG. 3 is a flowchart illustrating the steps used to extract feature points from images using wavelet transform.

[0021] FIG. 4 is a flowchart illustrating the steps used to obtain an initial set of matches using edge orientations and correlation values.

[0022] FIG. 5 is a flowchart illustrating the relaxation procedure used to obtain a set of reliable matching pairs.

[0023] FIG. 6 illustrates adjusting intensities of I_aand I_bsuch that their intensity differences are small.

[0024] FIG. 7 illustrates an example for a blending technique.

[0025] FIG. 8 shows an architecture for implementing the present invention.

[0026] FIG. 9 shows a series of input images for stitching.

[0027] FIG. 10 shows the stitching result obtained by the proposed stitcher.

[0028] FIG. 11 shows images used to demonstrate the ghost-like effect.

[0029] FIG. 12 shows how the ghostlike effect is removed: (a) with a wider blending width and (b) with a narrower blending width.

[0030] Referring first to the flow diagram of FIG. 1, a method of the general steps of generating a seamless panoramic image from different pieces of an image is shown.

[0031] In general, it is difficult to seamlessly stitch two adjacent images together to form a panoramic image due to perspective distortion introduced by a camera. To remove the effects of this distortion, these images are preferably reprojected onto a simple geometry, e.g., a cube, a cylinder, or a sphere. In many applications, a cylindrical geometry is preferable since its associated geometrical transformation is simple. In this example, the cylindrical geometry is used. FIG. 2 illustrates the relationship between the cylindrical surface □pqrs and film plane □PQRS of the camera. Note that the plane □PQRS is tangent to the cylindrical surface □pqrs. Let O denote the optical center, C the center of image plane, f the focal length and r the radius of the cylinder. Further, assume that P(x, y) is a pixel in the image plane and p(u, v) is its corresponding pixel in the cylindrical map. Using FIG. 2, the coordinates of p (u, v) can be obtained as follows: $\begin{array}{cc}u=r\angle C\mathrm{OB}=r{\tan }^{-1}\frac{\stackrel{\_}{\mathrm{CB}}}{\stackrel{\_}{\mathrm{OC}}}=r{\tan }^{-1}\frac{x}{f},& \left(1\right)\end{array}$

[0032] and $\begin{array}{cc}v=r\frac{\stackrel{\_}{\mathrm{PB}}}{\stackrel{\_}{\mathrm{OB}}}=r\frac{x}{\sqrt{x^2+f^2}}.& \left(2\right)\end{array}$

[0033] Moreover, since the radius r is equal to f, Equations (1) and (2) can be rewritten as follows:

[0034] u=ftan⁻¹x/f  (3)

[0035] and $\begin{array}{cc}v=f\frac{x}{\sqrt{x^2+f^2}}.& \left(4\right)\end{array}$

[0036] Based on Eqs. (3) and (4), input images are provided and then (step 10) warped into a cylindrical map for further registration to construct a complete panoramic scene (step 11).

[0037] Referring again to FIG. 1 and FIG. 3, once the input images have been warped, useful features required for registration are extracted (step 12). This process advantageously uses a wavelet transform for feature extraction (step 31). In particular, all of the feature points are detected along the edges. First, let S(x, y) be a 2-D smoothing function. Two wavelets, ψ¹(x, y) and ψ²(x, y) are the partial derivatives of the smoothing function S(x,y) in the x and y directions, respectively, where ${\psi }^1\left(x,y\right)=\frac{\partial S\left(x,y\right)}{\partial x}$

[0038] and ${\psi }^2\left(x,y\right)=\frac{\partial S\left(x,y\right)}{\partial y}.$

[0039] Let ${\psi }_{2^j}^1=\frac{1}{4^j}{\psi }^1\left(\frac{x}{2^j},\frac{y}{2^j}\right)$

[0040] and ${\psi }_{2^j}^2=\frac{1}{4^j}{\psi }^2\left(\frac{x}{2^j},\frac{y}{2^j}\right).$

[0041] At each scale 2^j, the 2-D wavelet transform of a function I(x,y) in L²(R²) can be decomposed into two independent directions as follows:

[0042] W₂¹I(x, y)=I*ψ₂¹(x, y) and W₂²I(x, y)=I*ψ₂²(x, y).

[0043] Basically, these two components are equivalent to the gradients of I(x,y) smoothed by S(x,y) at scale 2^jin the x and y directions. At a specific scale s=2ⁱ, the modulus of the gradient vector of f(x,y) can be calculated as follows:

[0044] M₂I(x, y)={square root over (|W₂¹I(x, y)|²+|W₂²I(x, y)|²)}.

[0045] If the local maximum of M₂I(x, y) are located and thresholded with a preset value, then the edge points of I(x,y) at scale 2^jcan be detected. Since we are interested in specific feature points for scene registration, additional constraints have to be introduced. In general, noise is the main cause of false detection of edge points. In order to suppress the effect of noise, a criterion called edge correlation is introduced (step 32): $R_n\left(j,x,y\right)=\prod _{i=0}^{n-1}{M}_{2^{j+1}}I\left(x,y\right),$

[0046] where n is a positive integer indicating the number of scales involved in the multiplication, and j represents the initial scale for edge correlation. R_nreveals a peak whenever a true edge exists and is suppressed by the multiplication process if a point at location (x,y) is not a true edge. Thus, using the relationship for R_n(j, x, y) above, the noise in an image can be suppressed while the true edges can be retained. In one embodiment of the invention, the number of scales for multiplication is chosen to be 2. In order to conserve the energy level, R_n(j, x, y) should be normalized as follows: ${\stackrel{\_}{R}}_n\left(j,x,y\right)=R_n\left(j,x,y\right)\sqrt{\frac{\mathrm{MP}\left(j\right)}{{\mathrm{RP}}_n\left(j\right)}},$

[0047] where $\mathrm{MP}\left(j\right)=\sum _{x,y}^{}{\left|M_{2^j}I\left(x,y\right)\right|}^2$

[0048] and ${\mathrm{RP}}_n\left(j\right)=\sum _{x,y}^{}{\left|R_n\left(j,x,y\right)\right|}^2.$

[0049] During the feature point selection process, an edge point is recognized as a candidate if its corresponding normalized edge correlation R₂(j, x, y) is larger than its corresponding modulus value. Basically, the above mentioned process is equivalent to detecting an edge point with the strongest edge response in a local area. The three conditions which will be used to judge whether a point P(x,y) is a feature point or not are as follows:

[0050] Condition 1: P(x,y) must be an edge point of the image I(x,y). This means that P(x,y) is a local maxima of M₂I(x, y), and M₂I(x, y)>a threshold (step 34),

[0051] Condition 2: {overscore (R)}₂(1, x, y)>M₂I(x, y) (step 36),

[0052] Condition 3: $M_{2^j}I\left(x,y\right)=\underset{\left(x,y\right)\in N_p}{\max }\left\{M_{2^j}I\left(x,y\right)\right\},$

[0053] where N_pis the neighborhood of P(x,y) (step 34).

[0054] Referring to FIG. 3 and 4, as part of step 12 of FIG. 1, a set of feature points is extracted using wavelet transform (step 31). Let FP_I={p_i=(p_xⁱ, p_yⁱ)} and FP_I={q_i=(q_xⁱ, q_yⁱ)} represent two sets of feature points extracted from two partially overlapping images I_aand I_brespectively. Assume that N_Iand N_Irepresent the number of elements in the set of points FP_Iand FP_I, respectively. Let. A(u) be the angle of an edge point u. A standard method for estimating the orientation of an edge-based feature point u at scale 2^jcan be expressed as follows:

[0055] Arg(W₂¹I(x, y)+W₂²I(x, y)).

[0056] However, the above representation can be very sensitive to noise. Therefore, an edge tracking technique plus a line-fitting model is used to solve the noise problem (step 41).

[0057] Let P be a feature point and N_ebe its neighborhood. Since P is an edge point, there should exist an edge line passing through it. By considering P as a bridge point, an edge line passing through P can be determined by searching in all the directions from P. All the edge points on the edge are then used as candidates for determining the orientation of the edge line. During the searching process, the edge connection constraint and the direction consistency constraint are used to restrict the searching domain. The edge connection constraint means that if N_econtains another edge l but l does not pass P, all edge points in l will not be included in estimating the orientation of P. In certain cases, however, there will exist more than one edge line passing through P. In these cases, the first line detected is adopted to estimate the orientation. Let l₁denote this line. The direction consistency constraint means all the edge points along other edge lines whose orientations are inconsistent with l₁are not included to estimate the orientation of P. In this way, a set of edge points can be selected and then used to estimate the orientation of P using a line-fitting model. In other embodiments, other edge tracking technique can also be applied to provide a better estimation.

[0058] After the orientation is estimated, all feature points u will associate with an edge orientation A(u). For a feature point p_iin the set of points FP_Iand q_jin the set of points FP_I, the orientation difference between them is calculated as follows (step 42):

[0059] θ_ij=A(q_j)−A(p_i).

[0060] In fact, if p_iand q_iprovide a good match, the value of θ_ijwill be small since the orientation of image I_ais similar to that of I_b. Assuming this is the case, step 14 (FIG. 1) is performed to eliminate impossible false matches in advance, avoiding many unnecessary correlation calculations.

[0061] For a feature point p_iin FP_Iand a feature point q_jin FP_I, if they form a good match (step 44), the following condition will be satisfied:

[0062] |A(p_i)−A(q_j)|<10°.  Condition 1:

[0063] Adding this criterion will significantly speed up the search time. On the other hand, if p_iand q_jform a good match, the similarity degree between p_iand q_jshould be larger. A cross-correlation which can be used to measure the similarity degree between p_iand q_j(step 46) and is defined as follows (step 32): $\begin{array}{cc}C\left(p_i;q_j\right)=\frac{1}{{\sigma }_i{\sigma }_j{\left(2M+1\right)}^2}\sum _{x,y=-M}^{x,y=M}\left[I_a\left(x+p_x^i,y+p_y^i\right)-u_i\right]\left[I_b\left(x+q_x^i,y+q_y^i\right)-u_j\right],& \left(5\right)\end{array}$

[0064] where u_iand u_iu_jare the local means of p_iand q_j, respectively, σ_iand σ_iare the local variances of p_iand q_j, respectively; and (2M+1)²represents the area of matching window. Based on this correlation measure, a pair {p_i⇄q_j} is qualified as a possible matching pair if the following conditions are satisfied (step 33): $\mathrm{Condition}2\text{:}{C}_{I_a{I}_b}\left(p_i;q_j\right)=\underset{q_k\in {\mathrm{FP}}_{I_b}}{\mathrm{Max}}C\left(p_i;q_k\right),\mathrm{Condition}3\text{:}{C}_{I_a{I}_b}\left(p_i;q_j\right)=\underset{p_k\in {\mathrm{FP}}_{I_a}}{\mathrm{Max}}C\left(p_k;q_i\right),$

[0065] and

[0066] C_II(p_i;q_j)>T_c, where T_c=0.65 (step 35).  Condition 4:

[0067] Condition 2 means that given a feature point p_iit is desired to find a point q_iεFP_Isuch that a the value of C_II(p_i;q_j) is maximized for all points q_kεFP_I. Condition 3 means that given a feature points q_j, it is desired to find a point p_iεFP_Isuch that the value of C_II(p_i;q_j) is maximized. If only Condition 2 is used, it is possible that several points p_imatch with single point q_j. Conversely, if only Condition 3 is used, several points q_jpossibly will match with single point p_i. As for Condition 3, it forces the value of C_IIof a matching pair to be larger than a threshold. In a preferred implementation, the orientation constraint will be checked first. If the constraint is not satisfied, it is not necessary to check Conditions 2, 3, and 4. In this way, only a few pairs are needed to calculate the cross-correlation measure C_II, which is considered a time bottleneck of the whole process.

[0068] Referring to FIGS. 1 and 5, a set of reliable matching pairs is obtained through relaxation. Let MP_I,I={p_i⇄q_i}_{i=1,2 . . .}be the set of matching pairs which satisfy Conditions 1, 2, 3, and 4 above, where p_iis a point in image I_a(x, y) and q_iis another point in image I_b(x, y). Let Ne_I(p_k) and Ne_I(q_k) denote the neighbors of p_kand q_kwithin an area of radius R, respectively. Assume that NP_pq={n_k¹⇄n_k²}_{k=1,2 . . .}is the set of matching pairs, where n_k¹εNe_I(p_i), n_k²εNe_I(q_j), and all elements of NP_pqbelong to MP_I,I. The proposed method is based on a concept that if {p_i⇄q_i} and {p_j⇄q_j} provide a pair of good matches, the distance between p_iand p_jshould be similar to the one between q_iand q_j. Based on this assumption, we can measure the quality of a matching pair {p_i⇄q_i} according to how many matches {n_j¹⇄n_j²} in NP_pqwhose distance d(p_i, n_j¹) is similar to the distance d(q_i, n_j²), where d(u_i,u_j)=∥u_i−u_j∥, the Euclidean distance between two points u_iand u_j. With this concept, the measure of the quality of a match {p_i⇄q_i} is defined as follows: $\begin{array}{cc}{G}_{I_a{I}_b}\left(i\right)=\sum _{\left\{n_k^i\iff n_k^2\right\}\in {\mathrm{NP}}_{\mathrm{Pi4i}}}^{}\frac{r\left(i,k\right)}{1+\mathrm{dist}\left(i,k\right)},& \left(6\right)\end{array}$

[0069] where $\mathrm{dist}\left(i,k\right)=\left[d\left(p_i,n_k^1\right)+d\left(q_i,n_k^2\right)\right]/2,r\left(i,k\right)=\{\begin{array}{cc}{e}^{-u\left(i,k\right)/T_1}& \mathrm{if}\mu \left(i,k\right)<T_2\\ 0,& \mathrm{otherwise}\end{array}$

[0070] with the two predefined thresholds $T_1\mathrm{and}{T}_2,\mathrm{and}u\left(i,k\right)=\frac{\left|d\left(p_i,r_k^1\right)-d\left(q_i,r_k^2\right)\right|}{\mathrm{dist}\left(i,k\right)}.$

[0071] The contribution of a pair {n_k¹⇄n_k²} in NP_pqmonotonically decreases based on the value of dist(i,k). Further, if the value of u(i,k) is larger than the threshold T₂, the contribution of {n_k¹⇄n_k²} is set to zero.

[0072] Referring to FIG. 5, after the quality of match of each pair {p_i⇄q_i} in MP_I,I(step 51), relative quality value of each pair is obtained to G_II(i) to further eliminate false matches (step 51). Now, based on the quality value of each candidate match, a relaxation technique is used to eliminate false candidates for further registration. If we define the energy function as follow: $F=\sum _{\{p_i\iff q_i❘❘\in {\mathrm{MP}}_{I_a{I}_b}}{G}_{I_a{I}_b\left(i\right),}$

[0073] then the relaxation procedure can be formulated as follow: $\mathrm{Iterate}\{-\mathrm{Compute}\mathrm{the}\mathrm{quality}\mathrm{for}\mathrm{each}\mathrm{candidate}\mathrm{match}-\mathrm{Choose}\mathrm{the}\mathrm{best}\mathrm{possible}\mathrm{candidates}\mathrm{for}\mathrm{minimizing}F\mathrm{according}\mathrm{to}\mathrm{the}\mathrm{quality}\mathrm{value}{G}_{I_a{I}_b}\left(i\right).\}\mathrm{until}F\mathrm{converges}.$

[0074] There are several strategies for updating the matching candidates. In one application an update strategy, referred to here as “some-looser-take-nothing” is used to update the matching candidates (step 52). First, according to the quality value of G_II(i), we sort each element of MP_I,Iis sorted in increasing order (step 52). Then, a predetermined percentage Q% of matches in MP_I,Iare eliminated as the impossible matches. Thus, the remaining (100−Q%) of matches are selected as potential matches for further relaxation (step 54). In our implementation, Q is set to 25. Three iterations are generally sufficient for achieving relaxation (step 53).

[0075] On the other hand, in order to make the matching results more reliable, the method includes an “angle consistency” constraint (step 53) within the first iteration of the relaxation process to further eliminate impossible matching pairs. That is, if {p_i⇄q_i} and {p_j⇄q_j} are well matched, the angle between {right arrow over (p_ip_j)} and {right arrow over (q_iq_j)} must be close to zero. In this case, during this first iteration, for each element {p_i⇄q_i} in MP_I,I, a counter CA_iis used to record the number of matches {p_k⇄q_k} in MP_I,Iwhere the angle between {right arrow over (p_ip_k)} and {right arrow over (q_iq_k)} is less than a predefined threshold θ. According to the value of CA_i, the elements of MP_I,Iare sorted in increasing order. The first Q% of matches in MP_I,Iare considered to be impossible matches.

[0076] After applying the relaxation process, a set of reliable matches is obtained. Referring to FIG. 1, the method uses a “voting” concept to derive a desired offset from the set of reliable matching pairs. For example, assume that this set is CP_I,I={u_i⇄v_i}_{i=1,2 . . . Ne}, where N_eis the total number of elements in CP_I,I. In general, the 2-D point sets {u_i} and {v_i} satisfy the following relation:

[0077] v_i=u_i+T, for i=1,2,3 . . . , N_e,

[0078] where T is the desired solution. However, in real cases, different pairs {u_i⇄v_i} will lead to different offsets T_i. Therefore, a voting technique is used to measure the quality of different solutions T_i. Let S(i) denote a counter which records the number of solutions T_kconsistent with T_i. Two solutions T_iand T_kare said to be consistent if the distance between T_iand T_kis less than a predefined threshold. Since there are N_eelements in CP_I,I, the total number of consistency tests will be N_e(N_e−1)/2. After applying the consistency test, the offset {overscore (T)} associated with the maximum value of S(i) is chosen as the desired solution.

[0079] Note that due to noise and image quality, the positions of feature points will not be precisely located and the accuracy of {overscore (T)} is affected. In practical implementation, the y-component {overscore (T_y)} of {overscore (T)} is generally affected more then {overscore (T_x)}. Therefore, the following method is proposed to refine and correct the solution of {overscore (T_y)} for the y-component. Let {u⇄v} be the matching pair of CP_I,Ihaving the highest quality value. Given point u and the offset {overscore (T)}, the refined offset can be found by searching the local neighborhood of the point (u+{overscore (T)}) in another image I_husing the correlation technique. However, if there is little texture information within the local neighborhood of the point u, the above approach will not necessarily provide a satisfactory solution. Let $g_x\left(k\right)=\sum _{i=1}^5\left|I_a\left(u_x+i,k\right)-I_b\left(u_x-i,k\right)\right|$

[0080] be the horizontal gradient of the point (u_x,k) in I_h, where u_xis the x-coordinate of u. Instead of using the starting point u directly, we use another starting point {overscore (u)} for refining the desired offset {overscore (T)} by searching the point {overscore (u)} whose a horizontal gradient g_xis largest along a column of pixels with the same x-coordinate u_x. Based on the starting point {overscore (u)}, the final offset can then be accurately obtained with the correlation technique.

[0081] In general, when stitching two adjacent images, discontinuities of intensity exist between their common areas. Therefore, step 17 of FIG. 1 is used to eliminate such intensity discontinuities. The scheme used in step 17 can be divided into two stages. The first stage is used to adjust the intensities of two adjacent images such that their intensities are similar. The second stage is used to blend their image intensities according to a distance measure such that the final composite image appears smooth. Assume I_aand I_bare two adjacent images with their widths in w_aand w_b, respectively. Let ΔI be the average intensity difference between the overlapping area

[0082] of I_aand I_b, that is, $\begin{array}{cc}\Delta I=\frac{1}{\left|A\right|}\sum _{i\in A}\left(I_b\left(q\left(i\right)\right)-I_a\left(p\left(i\right)\right)\right),& \left(7\right)\end{array}$

[0083] where A is the overlapping area of I_aand I_b, |A| is the number of pixels in A, p(i) is a pixel in I_a, and q(i) is its corresponding pixel in I_b.

[0084] In particular, referring to FIGS. 6A and 6B, the gap of average intensity between I_aand I_bis about ΔI. According to ΔI, w_a, and w_b, the first stage is used to adjust the intensities of I_aand I_bas follows: $\begin{array}{cc}{I}_a\left(p\left(x,y\right)\right)=I_a\left(p\left(x,y\right)\right)+x\frac{\Delta I}{2w_1},& \left(8\right)\end{array}$

[0085] and $\begin{array}{cc}{I}_b\left(q\left(x,y\right)\right)=I_b\left(q\left(x,y\right)\right)+\left(x-w_2\right)\frac{\Delta I}{2w_2}.& \left(9\right)\end{array}$

[0086] After this adjusting step, the intensities of I_aand I_bin FIG. 6(a) will be gradually changed to approach the intensity line {overscore (EF)} shown in FIG. 6(b), thereby bringing the intensities between I_aand I_bcloser. As will be described below, in order to further smooth the intensity discontinuity between I_aand I_b, a blending technique is then applied.

[0087] The second stage uses a ray-casting method to blend different pixel intensities together. Referring to FIG. 7, p_iis a pixel in I_a, q_iis its corresponding pixel in I_h, and I_aand I_hare two boundary lines in I_aand I_h, respectively. With p_iand q_i, the intensity of the corresponding pixel r_iin the composite image I can be obtained as follows: $\begin{array}{cc}I\left(r_i\right)=\frac{d_b^t{I}_a\left(p_i\right)+d_a^t{I}_b\left(q_i\right)}{d_a^t+d_b^t}& \left(10\right)\end{array}$

[0088] Where d_ais the distance between p_iand I_a, d_hthe distance between q_iand I_h, and t is an adjustable parameter. Using Equation (10), the intensities in I_aare gradually changed to approach the intensities of pixels in I_hsuch that the final composite image I looks very smooth. In fact, if the blending area is chosen too large, a “ghostlike” effect will occur, particularly when moving objects in the common overlapping area exist between I_aand I_b. However, since the intensities of I_aand I_bhave been adjusted, the blending width can be chosen small such that the so-called ghostlike effect is significantly reduced. In one preferred embodiment, the blending width is chosen as one-third of the original width of the overlapping area.

[0089] Referring to FIG. 8, one preferred architecture for implementing the real-time stitcher apparatus, described above, includes input devices 60 (e.g., digital cameras or scanners) to acquire a series of panoramic images. Then, the panoramic images are stored into external storage 62 such as hard disks for further processing or being directly provided to one or more microprocessors 64 for stitching. The microprocessors 64 perform the stitching including warping, feature extraction, edge orientation estimation, correlation calculation, relaxation, rendering and blending the final panoramic images, etc. All the temporary data are stored in the system RAM memory to speed up stitching. Finally, the stitching result is sent to the display engine for displaying. In many applications this architecture can be implemented using a general personal computer.

[0090] Referring to FIG. 9, to analyze the performance of the real-time stitcher apparatus, described above, a series of original panoramic images 9a-9ecaptured by a rotated camera were provided.

[0091] Referring to FIG. 10, the stitched panoramic image obtained by our proposed method is shown. In addition, if moving objects in two adjacent images exist, the proposed method can significantly minimize ghost-like effects. FIG. 11 shows two adjacent images with moving objects. The posture of the man in (a) is clearly different from the one in (b). Besides, large intensity differences exist between them. In order to smooth such large intensity differences, the blending width should be chosen to be large. However, with a large area, the ghost-like effect will appear in the final composite image (see FIG. 12(a)). Contrary to the previous technique, with our proposed blending technique, since the intensities of input images have been adjusted before blending, the blending width can be chosen smaller. Therefore, the ghostlike effect is significantly lessened, the quality of (b) is clearly improved over the quality shown in FIG. 12(a).