Multi-modal motion estimation for video sequences

[0001] The present invention relates generally to the field of video signal processing, and more particularly, a system and method for hierarchical block motion estimation of video sequences.

[0002] Video is typically represented by sequences of two-dimensional image frames or fields. In providing the information representing the sequence of images, however, a bandwidth issue may exist because of the amount of data that needs to be transmitted. In order to provide video information that accommodates potential bandwidth issues, video compression techniques are needed. In a typical compression technique, frames are removed from the sequence of images prior to providing the information, and then, prior to when the images are to be displayed, the images that have been removed need to be reconstructed. One method of reconstructing images is through the use of motion estimation. That is, image frames are constructed based on the estimated motion of objects displayed by the available image frames. More generally, motion estimation can be used for a variety of other video signal processing purposes as well.

[0003] Different techniques have been developed for the purposes of estimating motion using available image frames. One such technique of motion estimation employs a block matching technique. In block matching methods, an image is subdivided into square or rectangular blocks having constant sizes or having a prescribed plurality of picture elements, for example, 16×16 or 8×8 picture elements per block. Motion vectors representing the estimated motion from a previous search image are typically determined for each of the blocks of a reference image. In simple applications of the block matching method, the same motion vector for an entire block is attributed to each of the picture elements in the block. Generally, in determining the motion vector for each block of the reference image, a range of possible motion vectors is determined for a respective block using a least difference calculation. From the determined ranges, the motion vector to a block in the search image frame having the least calculated difference from a block in the present reference image is accepted as the motion vector for the block in the reference image frame. Image frames can the be constructed using the motion vectors to fill in those images that were removed during video compression.

[0004] In addition to the basic block matching technique previously described, refinements to the basic process have been developed to provide more accurate motion estimations. For example, one refinement provides an iterative block matching technique where the matching process is repeated for blocks of decreasing size. Following the determination of motion vectors for blocks of a first size, the blocks are then sub-divided into smaller child blocks. The child blocks then inherit the motion vector of the parent block in which they are located as a starting point from which a more precise motion vector for each of the child blocks may be calculated. In another refinement, block matching is based on hierarchical processing with different block sizes, starting with coarse resolution blocks at a higher level of the pyramid, and advancing through finer resolution layers for at each layer of the pyramid. A predicted motion vector that is calculated for each block is used to point to a search region in which a matching search block is likely to be located. However, as the resolution gets finer, and the effective block size gets smaller, there are objects for which the parent motion vectors no longer apply and new motion vectors need to be found.

[0005] In another refinement of the block matching technique, a quad tree-like sub-division process is used to refine the motion vector field from a coarse to a fine resolution. In this process, the motion vector for the parent block of the reference frame are initially determined. Each parent block is then sub-divided into child blocks, which inherit the motion vector of its parent block. A refinement process is performed where a new motion vector may be attributed to a child block if a local block search around the vicinity of the block to which the inherited motion vector points yields a better matching search block. Thus, for each of the child blocks, a motion vector search is performed. However, the search is restricted to a set of candidate motion vectors consisting of the motion vectors associated with the parent and child blocks adjacent to the child block for which a motion vector is being determined. Although this block matching technique maintains smoothness of the motion field, it may be the case that the true motion vector is not found in the immediate surrounding neighborhood. That is, the candidates from which a motion vector is searched may be too constrained for accurate motion estimation for multiple non-planar motion surfaces, such as for scrolling text credits overlay, and multiple moving foreground and background objects in video scenes. Simply extending the neighborhood increases computational costs as well as the possibility of false matches.

[0006] Therefore, there is a need for a system and method that facilitates motion estimation that can escape typical neighborhood constraints in determining motion vectors.

[0007] The present invention is directed to a method and system for estimating motion between a reference image and a search image of a video image sequence. Motion vectors are determined for pixel blocks of a reference image of the video image sequence. The major modes of the motion vectors for the reference image are identified and used for determining each pixel block of the reference image, if a better match can be made based on the major modes of the motion vectors. A hierarchical process may be employed where the method is performed for a first image resolution, and then repeated for continually finer image resolutions until the method is completed for an original image resolution.

[0008] The determination of the respective motion vectors for the pixel blocks of the reference may be accomplished by attributing a motion vector from a related lower resolution pixel block and refining the respective attributed motion vector for each pixel block. Refinement of the attributed motion vectors may be performed by calculating an error measure for a pixel block in a reference image with respect to pixel blocks in a search image in the vicinity of the pixel block in the search image to which the attributed motion vector points, and if a better match is determined, associating the vector pointing to the better matching pixel block in the search image with the pixel block in the reference image.

[0009] The major modes of the motion vectors may be performed by compiling a list of the motion vectors for the image in descending order of the frequency of occurrence, and defining the motion vectors occurring more times than a threshold value as the major modes.

[0010] A better match can be determined by calculating for a pixel block of a reference image an error value for every pixel block of a search image to which the major modes of the motion vectors point, and comparing the calculated error values to an initial error value, the initial error value representing an error measure between the pixel block of the reference image and the pixel block of the search image identified by the motion vector determined initially for the pixel block.

[0011] The method can also incorporate a looping routine where a respective predicted motion vector for pixel blocks a subsequent reference image of the video image sequence is calculated from the motion vectors determined for the pixel blocks of a previous reference image. A list of the motion vectors for the pixel blocks of the subsequent reference image are compiled in descending order of the frequency of occurrence and then the predicted motion vectors of the subsequent image are refined. Based on the major modes of the motion vectors for the subsequent image for each pixel block of the subsequent image, a determination of whether a better match exists is then made.

[0014] Embodiments of the present invention provide a multi-modality motion estimation system and method which facilitates the estimation of motion based on major modes of the motion field of a reference video frame. As a result, the determination of motion vectors can be more dynamic and can escape neighborhood constraints when necessary. The motion estimation can be performed in a hierarchical fashion where the method is repeated for continually higher levels of image resolution until the process is completed for the original image resolution. Certain details are set forth below to provide a sufficient understanding of the invention. However, it will be clear to one skilled in the art that the invention may be practiced without these particular details. In other instances, well-known circuits, control signals, timing protocols, and software operations have not been shown in detail in order to avoid unnecessarily obscuring the invention.

[0015] Illustrated in FIG. 1 is a flow diagram of an embodiment of the present invention. It will be appreciated that the following description is being provided as an example for illustrative purposes only, and should not interpreted as limiting the scope of the present invention to the details discussed.

[0016] The motion estimation typically involves two images: a reference or new image and a search or previous image. For each video frame, a multi-resolution pyramid is constructed. The pyramid is constructed with a predetermined number of resolution layers. By low pass filtering or down-sampling of a finer resolution layer, coarser resolution layers may be obtained. For example, in the case where the input frames are in a high-definition (HD) video format, typically, 1920 pixels/scanline×1080 scanlines, the input video can be down-sampled five times, each time down-sampling by a ratio of two in both the horizontal and vertical directions. Thus, the 1920×1080 HD video format is reduced to approximately 60×34 at the coarsest resolution (i.e., top-level) layer. As will be discussed in greater detail below, the multi-resolution pyramid from a previous image frame is buffered in a memory for later use.

[0017] For each of the resolution layers of the pyramid, the reference and search images are divided into an array of blocks for which motion vectors will be determined. The blocks of the reference image will be referred to herein as reference blocks and the blocks of the search image will be referred to as search blocks. The dimension of the block themselves remains the same for each of the different resolution layers, but because the resolution at each layer is different, the effective resolution of the blocks are different per layer as well. An example of a typical pixel block dimension is 16×16 pixels.

[0018] For every block in the reference image, the objective is to find the displacement or motion vector pointing to a search block in the search image such that the displaced search block is the “best match.” It will be appreciated that the determination of the quality of the match can be made by applying a variety of calculations. A well-known technique that may be used to measure block correlation is the sum of absolute differences (SAD) of a block of pixels in a reference frame F(t) with respect to a displaced block of pixels in a search frame F(t−T). Higher powers of the measure may also be used. Calculation of the block SAD is defined as follows: $\mathrm{SAD}\left(x_0,y_0,\mathrm{dx},\mathrm{dy}\right)=\sum _{j=0}^{M-1}\sum _{i=0}^{N-1}\mathrm{DFD}\left(x_0+i,y_0+j,\mathrm{dx},\mathrm{dy}\right)$
where SAD is the block sum of the displaced frame pixel difference (DFD), (x₀, y₀) refers to the upper left corner of the block of pixels with a dimension of M×N pixels, (dx, dy) is the displacement of the pixel in the search frame F(t−T). The DFD is the absolute difference of the pixel in frame time (t) with displaced pixel in frame time (t−T). The definition for the DFD is as follows:
DFD(x,y,dx,dy)=|F(x,y,t)−F(x+dx,y+dy,t−T)|
where T is the temporal distance between reference frame F(t) and search frame F(t−T), (x,y) is the coordinate of the pixel, and (dx, dy) is the displacement of the pixel in the search frame F(t−T).

[0019] Operation of embodiments of the present invention will be described by way of the following example. However, it will be appreciated that the details of the example are provided for illustrative purposes, and should not be interpreted as limiting the scope of the present invention.

[0020] An input video frame including a luminance signal is provided for motion estimation. The luminance signal of the input video frame or field is received along with a scene change flag (step 104). When the scene change flag is set, it indicates that the current input video frame has started a new scene. Such flags are well-known in the art and will not be discussed in further detail.

[0021] In the following example, the image received has the change of scene flag set, thus indicating that the image is the first of a new scene and the motion search will begin without the use of any prior motion tracking information. That is, the image signaling the scene change will be the search image for the following received image, which will become the reference image. Following receipt of the reference (following) image, the initial motion estimation is performed using the coarsest resolution image of the reference and search pyramids (step 108). A conventional method of determining the motion vectors of the blocks of the first or top-level resolution layer will be used for the first reference image after a scene change is indicated. Many such techniques are well-known and may be used to establish the motion vectors for the top level blocks. For example, a pre-defined fixed search radius could be used in the block matching process.

[0022] After the completion of the initial top-level block search following a scene change, each of the blocks will have an associated motion vector. At the next highest resolution layer of the pyramid, each child block of the finer resolution layer inherits the motion vector of its parent block (step 112). That is, the motion vector of a parent block of a coarser layer is attributed to each of its child blocks of the next highest resolution layer. The inherited motion vector is properly scaled to account for the difference in sampling factor between the parent and child layer in the pyramid.

[0023] The respective motion vector that is inherited by each of the blocks of the next highest resolution layer provides a starting point from which the respective motion vectors are further refined (step 112). An example of a suitable refinement technique is to search around the search block to which the inherited motion vector is directed for a predefined small radius and determine if a better match search block can be found. For example, a radius for refinement may be set at ±1 pixel. Consequently, the SAD will be calculated for each of the eight search blocks surrounding the search block pointed to by the inherited motion vector and, if a better match is determined based on the SAD, a new motion vector will be associated with the reference block. It will be appreciated, however, that other refinement techniques may be applied in embodiments of the present invention as well, and consequently, the previous method should not be interpreted as limiting the scope of the present invention.

[0024] Following the local refinement searches, a histogram of the motion vectors for the current resolution layer is constructed. The major modes, up to a certain number of modes, is extracted from the histogram for a subsequent tournament stage, which will be described in more detail below (step 116). For example, the major modes that may be candidates can be limited to the five or six modes having the greatest frequency of occurrence. In the alternative, to minimize spurious matches, modes having fewer occurrences than a minimum threshold are rejected as candidates. The resulting list is arranged in the order of the frequency of occurrences, such that in the tournament stage, modes having higher frequencies have priority.

[0025] During a tournament stage (step 120), the error measure, for example, the SAD, associated with every motion vector in the list of the major modes is calculated. In selecting the final motion vector for a reference block, the process begins by using its refined motion vector with its associated error measure as the initial setting. The tournament process is carried out with the higher frequency mode first, such that the outcome favors the majority mode. Traversing down the mode list, if the error measure of the current mode is less than the current setting by a prescribed margin, the current setting is replaced by the values of the current mode. That is, a new motion vector and its associated error measure are assigned to the current reference block. As a result, if the inherited motion vector is an outlier, one of the motion vectors in the mode list will most likely be selected and replace the inherited motion vector. The prescribed margin by which the difference between the error measure of the current mode and the current setting must exceed is specified such that hysteresis is presented to favor the first selection.

[0026] As illustrated in FIG. 1, the steps inheriting a scaled motion vector and the further refinement of the inherited vectors, calculation of the major modes of the current resolution layer, and the tournament process, are carried out for the finer resolution layers of the pyramid until the full resolution layer is completed (step 124).

[0027] As previously mentioned, in the present example a scene change was indicated by the input video image. However, where the new input frame or field is a continuation of the same scene, the previously derived motion field can provide motion-tracking information to help the derivation of the motion field for the new reference image. That is, motion tracking of the previous motion field is incorporated in the initial top level search for the new reference image (step 128). Through this process, the top-level blocks of the new reference image inherit the motion-tracking information from the previous full resolution motion field for the former reference image. The following information is derived from the previous full resolution motion field with the proper scaling (step 132):

• • [0028] a predicted motion vector for each reference block in the top-level reference frame is calculated from the corresponding local extent of the previous full resolution motion field; and
  • [0029] a motion vector histogram is constructed from the full resolution motion field, from which a list of the major motion vector modes is extracted.
    For every block of the reference image, the search involves a refinement search of a very small radius around the predicted and scaled motion vector inherited from the previous full resolution image. After the refinement, a tournament process similar to that previously described is carried out for each of the top-level blocks of the reference image. The resulting output motion vector field for the top-level resolution layer is then provided into the iteration loop for determination of the full resolution motion field of the new reference image.

[0030] FIG. 2 illustrates a video processor 200 in which embodiments of the present invention may be implemented. Included in the video processor 200 is a motion estimator 204 for estimating motion in sequences of video images according to aspects of the present invention. It will be appreciated that there are many well-known techniques for implementing the video processor 200 and the motion estimator 204 described herein, and, although a more detailed description of the various techniques has not been provided, those of ordinary skill in the art have sufficient understanding to practice embodiments of the present invention. It will be further appreciated that embodiments of the present invention may be practiced using conventional software language, conventional hardware circuitry, or combinations thereof, well understood by those of ordinary skill in the art. In either case, the description provided herein is sufficient to allow those of ordinary skill in the art to practice the present invention without the need for undue experimentation.

[0031] It will be appreciated that embodiments of the present invention may be practiced using conventional software language and conventional hardware circuitry well understood by those of ordinary skill in the art. In either case, the description provided herein is sufficient to allow those of ordinary skill in the art to practice the present invention without the need for undue experimentation. Moreover, as previously mentioned, certain details have been set forth above to provide a sufficient understanding of the invention. However, it will be clear to one skilled in the art that the invention may be practiced without these particular details. In other instances, well-known circuits, control signals, timing protocols, and software operations have not been shown in detail in order to avoid unnecessarily obscuring the invention.

[0032] From the foregoing it will further be appreciated that, although specific embodiments of the invention have been described herein for purposes of illustration, various modifications may be made without deviating from the spirit and scope of the invention. For example, embodiments of the present invention have been described as being implemented in a video processor. However, embodiments of the present invention may be implemented in other systems where motion estimation is desirable, such as a computer system, video displays, televisions, and the like. More generally, embodiments of the present invention may be implemented in general processing environments, where the motion estimation described herein are carried out by a general purpose computer. Accordingly, the invention is not limited except as by the appended claims.