The present invention relates to motion estimation techniques which can be used, for example, in systems for processing and transmitting video signals.
Most of the motion estimation techniques developed in recent years can be essentially divided into certain basic categories, namely:
i) fast search techniques, which attempt to mitigate complexity by reducing the number of candidate motion vectors that are tested in the search area. The choice of the vectors is driven in a heuristic way with the aim of locating the absolute minimum of the cost function. Examples of this category are the solutions known as Three Step Search, Fast Search, New Three Step Search and Four Step Search, documented, for example, in the papers by B. Furth, J. Greenberg and R. Westwater, “Motion Estimations Algorithms for Video Compression,” Kluwer Academic Publishers, 1997; R. Li, B. Zeng and M. L. Liou, “A new Three-Step Search Algorithm for Block Motion Estimation,” IEEE Transactions on Circuits and Systems for Video Technology, Vol. 4, No. 4, August 1994, pp. 438-441; L.-M. Po and W.-C. Ma, “A novel Four-Step Search Algorithm for Fast Block Motion Estimation,” IEEE Transactions on Circuits and Systems for Video Technology, Vol. 6, No. 3, June 1996, pp. 313-317;
ii) fast matching techniques, in which the complexity is reduced by decreasing the number of points on which the cost function is calculated. Examples of this category are the techniques of quincunx sub-sampling and one-bit transform techniques described in the papers by K. Lengwehasatit and A. Ortega, “A Novel Computationally Scalable Algorithm for Motion Estimation,” VCIP '98, January 1998; B. Natarajan, V. Bhaskaran and K. Konstantinides, “Low-Complexity Block-Based Motion Estimation via One-Bit Transforms,” IEEE Transactions on Circuits and Systems for Video Technology, Vol. 7, August 1997, pp. 702-706;
iii) techniques using a spatio-temporal correlation. In these, the motion vector is selected from those which have already been calculated in the current end for the preceding frames. This vector is frequently the starting point of a refining process, as documented in the paper by J. Chalidabhongse and C.-C. Kuo, “Fast Motion Vector Estimation Using Multiresolution-Spatio-Temporal Correlations,” IEEE Transactions on Circuits and Systems Technology, Vol. 7, No. 3, June 1997, pp. 477-488, and in the paper by K. Lengwehasatit and A. Ortega cited previously;
iv) hierarchical or multiresolution techniques, in which the motion vectors are first searched for in a low-resolution image and then refined in the normal-resolution image, as documented in the paper by J. Chalidabhongse and C.-C. Kuo cited previously; and
v) rate-distortion optimized techniques, which are not really low-complexity algorithms, since the cost function is somewhat more complicated, taking into account not only the energy of the motion-compensated differential signal but also that derived from the choice of the motion vector (the number of bits required for coding it, the coding modes, etc.), as documented in the paper by F. Kossentini, Y.-W. Lee, Mark J. T. Smith and R. K. Ward, “Predictive RD Optimized Motion Estimation for Very Low Bit-Rate Video Coding,” IEEE Journal on Selected Areas in Communications, Vol. 15, No. 9, December 1997, pp. 1752-1763.
The embodiment of the present invention disclosed herein provides a solution that, while retaining characteristics of low complexity, makes it possible to obtain results essentially comparable with the results obtainable with techniques based on an exhaustive search.
According to the embodiments of the present invention, a process is disclosed having the characteristics that are radically different from the fast search techniques mentioned previously, in which the computational complexity is contained by correspondingly limiting the search area to a restricted number of points.
The disclosed embodiment of the invention also differs from the solutions called Rate-Distortion Optimized to which reference was made previously, in which a Lagrangian function which combines the residual energy with the number of bits consumed is used to evaluate the motion vectors. A motion vector is estimated from a diamond-shaped search area divided into levels on the basis of the existing correlation between adjacent macroblocks.
The motion vectors from all macroblocks on the same level are considered and then combined suitably with a prediction function (mean, median, weighted mean, and statistical mean) to generate a new motion vector on which the Lagrangian function can be evaluated and the decision made as to whether other levels must be tested.
When compared with the solutions described in the previously cited papers by Chalidabhongse and Kuo and by Lengwehasatit and Ortega, the solution according to the invention is essentially different during the refining phase, since the choice of the best candidate, on the basis of the spatio-temporal correlation between the vectors with respect to the adjacent macroblocks, is the same for all the algorithms. However, in these algorithms the refining phase is a recursive process, since the refining grid is moved until no reduction of the cost function is obtained.
On the other hand, in the solution disclosed, the refining phase is conducted on a grid with a fixed number of points. The grid can be amplified to permit recovery of the estimation errors, but the number of points of the grid is always the same.
The embodiment of the invention includes three fundamental advantages over the known solutions:
the disclosed embodiment has constant complexity, since the number of operations is the same for all the macroblocks,
the motion vector range is independent of any search window, and
the embodiment has low complexity, since the number of operations required is negligible by comparison with other algorithms.
The disclosed embodiment will now be described, purely by way of example and without restrictive intent, with reference to the attached drawings, in which:
In the diagram in
In the diagram in
The line 14 leads both to the memory 13 and to the set of output lines of the device 1, indicated as a whole by 15.
The reference 16 indicates the coding control block designed to act on two switching blocks 17a, 17b and also for other purposes.
The switching block 17a is designed to transfer to the block 18, which implements what is commonly called the DCT (acronym for Discrete Cosine Transform) function, either of the following alternatives:
the signal present on the input line 10,
the signal obtained in a difference block 19 as the difference between the signal present on the input line 10 and the signal taken from the output line 12 of the memory 13.
The output of the block 18 is sent to a quantization block 20 which acts under the control of the coding control block 16.
The output of the block 20, indicated by 21, forms another of the outputs 15 of the device 1. The signal present on the line 21 is also sent to two blocks 22, 23 which carry out, in cascade, functions which are the inverse of the functions carried out by the blocks 20 and 18, in such a way as to reconstitute samples which can be transmitted to the memory 13 via a further line 24. An adding node 25, to which the output line of the switch 17b leads, is inserted in the line 24. The operation of this node is coordinated with the operation of the switch 17a in such a way that, when the block 17a transfers the signal present on the input line 10 to the block 18, the signal present on the line 12 is added in the block 25 to the output signal of the block 23.
In addition to the lines 14 and 21 described above, the set of output lines 15 of the device 1 comprises further lines 26, 27 and 28, all leading from the coding control block 16 and designed to carry the intra/inter flag, transmitted or not-transmitted flag, and quantizer indication signals respectively. It will be appreciated that the line 28 effectively coincides with the line via which the block 20 is driven by the block 16.
The whole of the above descriptions relates to criteria that are known and are mainly recapitulated here in order to provide an illustration of a possible field of application of the invention.
In general, any motion estimation process (such as that implemented, for example, by the block 11 of
There are various functions that can estimate the matching and/or the residual error. They are generally functions of the pointwise difference between the two macroblocks.
In a preferred embodiment of the invention, the function called SAD (acronym for Sum of Absolute Differences) is used. This is a function capable of providing a good trade-off between precision and calculation complexity. As a possible alternative, use could be made, for example, of the SSD (acronym of Sum of Squared Differences) function. This alternative solution has some advantages in terms of reliability, since it calculates the real energy associated with the difference. However, it has a drawback in that it requires a multiplication.
The SAD function is defined as follows:
SAD(m,n)=Σ|B(i,j)−B′(i+m,j+n)| (I)
where the summation extends to all the values i,j belonging to the macroblock and B(i,j) is the pixel intensity in the position i,j of the current frame and B′i+m,j+n) is the corresponding pixel intensity in the preceding frame, shifted by a vector with components x,y equal to m and n respectively.
The SAD function is a measure of the motion vector quality, since a correct motion vector produces a low SAD, while an incorrect motion vector produces a high SAD.
The calculation of the SAD is therefore executed for all candidates which can possibly be selected as motion vectors for a given macroblock. The motion vector producing the lowest SAD is selected as the motion vector for the macroblock in question, and is used for the corresponding prediction function.
The motion estimation process according to the disclosed embodiment of the invention belongs to the category of processes based on the evaluation of the spatio-temporal correlation existing among motion vectors belonging to adjacent blocks.
If the motion field varies slowly both locally and from frame to frame is true, then it may be possible to check only a few sets of candidate motion vectors. The candidate motion vector that produces the minimum SAD is selected as the predictor motion vector which, after a refining phase, yields the final motion vector.
Since it provides for the use of a motion vector acting as a predictor, this solution is classed as a prediction process.
Like all processes of this type, it is essentially based on two phases, namely:
the identification of the candidate predictors, and
the refining of the best predictor.
At the end of the whole process, the motion vector that produces the lowest SAD is associated with each macroblock.
The first phase mentioned above therefore consists in identifying, from a set of candidates, the predictor motion vector identified as the best and therefore as that to which the subsequent refining phase is to be applied.
With the objective of achieving a low-complexity solution, the embodiment of the invention is intended to reduce the number of candidates as far as possible and, in the same way, to select “good” candidates, presumably those close to the correct vector.
The criterion for constituting the set of candidates is that of selecting the vectors of the relative macroblock position, in the current frame and in the preceding frame, which is close to the current vector. Naturally, there is a constraint due to the fact that a predictor vector must be calculated in advance. Since the macroblocks are normally subjected to scanning in lexicographical order, only the vectors belonging to macroblocks located above and to the left of the current one are available as vectors usable as predictors.
This solution is shown schematically in
The current macroblock is the one left blank in the right-hand part of FIG. 2 and the macroblock G in the left-hand part of
For clarity, the two spatial predictors taken from the frame t are those indicated by the letters A and B. The two temporal predictors are those indicated by the letters G and H in the frame t-1.
The motion vector, of the set thus defined, which produces the lowest residual error (for example, the lowest SAD function) is used as the starting point for the subsequent refining phase and is indicated below as the “best motion vector predictor.”
When the predictor in question has been determined, the process continues with a refining phase implemented by means of a grid of n points that is applied to it.
An example of such a grid is shown in FIG. 3. The grid in question is formed by four points I to IV at the vertices of a square and four further points V to VIII located at the vertices of a square of smaller size with its sides inclined at 45° to the sides of the square at whose vertices the points I to IV are located.
The background grid shown in
The aforesaid grid is considered to be centered on the position to which the best motion vector MV points.
The distance of points I to VIII from the center is defined by a linear function which depends on the matching error found during the preceding step (for example, a distance defined as the product of corresponding coefficients and the corresponding SADs). Additionally, since the process operates with motion vectors with a precision of half a pixel, in a preferred embodiment, the points are at a distance of half a pixel from the center.
After all the points on the grid have been considered, the vector with the lowest SAD is selected as the motion vector for the current macroblock and is used for its motion compensation.
In those cases in which the best predictor may be incorrect, for example in the presence of a change of scene in the environment of a rapidly changing motion, the grid correction can be amplified. The grid is amplified only when the best predictor has a high SAD, which means that it is probably not a good predictor.
Starting from the analysis of the SAD distributions found for some sample sequences, it is possible to determine a linear function of the SAD (for example, the same coefficient×SAD function mentioned above), which can be used to calculate the magnitude of the amplification function and the number of points, while also discriminating the conditions in which amplification is necessary from those in which this step is not necessary.
A measure of the complexity of the solution according to the invention (particularly in respect of the comparison with other known solutions, such as those based on a full search algorithm) is provided by the number of times that the cost function is calculated. In the case in question, the measure can be found in the number of operations of calculating the SAD function per macroblock.
It should be noted that this measure of complexity is completely independent of the format and frame frequency of the sequence on which the coding process is carried out, and is therefore valid for all the coding algorithms.
For a full search algorithm, the number of calculations of the SAD depends on the dimensions of the search area. For example, in the default prediction operating mode of the H.263+ video standard (see the document Image Processing Lab, University of British Columbia, “TMN (H.263+) encoder/decoder, version 3.0,” TMN (H.263+) codec, September 1997) the search area has a size of 32×32 pixels. Additionally, the best motion vector that is found is refined with a precision of half a pixel, by the application of a grid on n points. Therefore, the computational complexity of the full search algorithm is
OPMB=Num.SADinteger-pixel+Num.SADhalf-pixel=32×32+8=1032 (II)
However, the embodiment of the invention requires, in the example of embodiment illustrated here, four calculations of SAD for the selection of the best predictor (see
OPMB=Num.SADhalf-pixel+Num.SADNull-Vector=12+1=13 (III)
The embodiment of the invention can therefore be used to reduce the computational cost by approximately 90% with respect to the full search algorithm. Moreover, the cost is stable because the total number of calculations of the SAD function is the same for each macroblock. This gives rise to a considerable advantage of the solution according to the invention over other motion estimation algorithms with variable complexity deriving from their interactive form.
The embodiment of the invention lends itself to particularly advantageous developments with reference to the H.263+ video standard mentioned above.
This standard provides some optional modes which can be used to improve the subjective quality of the reconstructed sequences. In general, however, the use of these modes produces a more or less significant increase in the computational cost.
In the context of the illustration of the present invention it is advantageous to consider, among these options included in the standard, some options which are closer to the problem of motion estimation, in other words the improved PB-frames mode, abbreviated to “IPB mode,” and the advanced prediction mode (commonly abbreviated to “APM”). The integration of these options into a predictive algorithm makes them more uniform with the standard.
For a description of the IPB mode, reference may usefully be made to the documents ITU Telecom Standardization Sector of ITU, “Video Coding for Low Bitrate Communication,” Draft 21 ITU-T, Recommendation H.263+ Version 2, January 1998 and ITU Telecom Standardization Sector of ITU, “Video Codec Test Model, Near Term, Version 10,” TMN10 ITU-T, April 1998, as well as to the general document on the H.263+ standard cited previously, which are incorporated herein by reference in their entirety.
As a variation from the standard, it is possible to introduce one change only in order to permit its integration in a functional way into the algorithm. This change relates to the order in which the motion estimation is carried out. The standard applies the motion estimation on the P frame of the PB pair first and then on the B frame of the pair (even if the temporal order is exactly the reverse).
However, this solution has the drawback that the temporal predictors of the B frame of the PB pair would be chosen from the reference motion field relating to the successive frame in temporal terms.
In the solution shown in
The ITU documents cited previously provide a detailed description of the APM mode.
In this field of application, the disclosed embodiment of the invention makes it possible, for example, to associate the frame of a macroblock with four vectors instead of only one. All this is done in order to provide a more accurate estimate.
As indicated in the standard, the use of the APM consists in an application of the motion estimation algorithm to the current macroblock (to obtain an associated motion vector as in the base mode) and a successive application of the same algorithm for each 8×8 block forming the 16×16 macroblock to which the previously found vector points (so that four associated vectors are obtained for the current macroblock). By means of special preference rules, the coder chooses the optimal mode for preparing the current macroblock.
In the embodiment shown schematically in
This scheme is shown in
The arrow BV indicates the obtaining of the best motion vector for the current 16×16 macroblock, implemented by means of the predictive process, while the arrows MV1, MV2, MV3 and MV4 illustrate the obtaining of four corresponding best vectors obtained by the refining phase for the corresponding four 8×8 blocks of the PM macroblock obtained by prediction.
In the case of the APM mode, the complexity of the process according to the invention is greater as compared with the base version, owing to the fact that the number of predictor candidates is higher. For example, if the algorithm uses the equivalent of twenty-one predictors for a total of twenty-one SAD calculation operations per macroblock (it is emphasized that this is the maximum number of operations), the gain in terms of complexity is 1.62 times with respect to that of the base version.
Clearly, provided that the principle of the invention is retained, the details of implementation and the forms of embodiment can be widely varied from what has been described and illustrated, without thereby departing from the scope of the present invention as defined by the attached claims.
From the foregoing it will 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. Accordingly, the invention is not limited except as by the appended claims and the equivalents thereof.
Number | Date | Country | Kind |
---|---|---|---|
00830332 | May 2000 | EP | regional |
Number | Name | Date | Kind |
---|---|---|---|
5581308 | Lee | Dec 1996 | A |
5668608 | Lee | Sep 1997 | A |
5745183 | Lam | Apr 1998 | A |
5786860 | Kim et al. | Jul 1998 | A |
6014181 | Sun | Jan 2000 | A |
6512550 | de Garrido et al. | Jan 2003 | B1 |
6563872 | Suzuki | May 2003 | B2 |
Number | Date | Country | |
---|---|---|---|
20020012396 A1 | Jan 2002 | US |