Invention content
In view of this, the present invention provides a kind of face three-dimensional point cloud super-resolution fusion method and applies its data processing
Device, reduces the calculation amount of fusion, while improving the efficiency of fusion and putting cloud resolution ratio, and high-resolution is provided for three-dimensional face identification
Rate, low noise, imperforate point cloud data.
In a first aspect, a kind of face three-dimensional point cloud super-resolution fusion method is provided, including:
Multiframe 3 D human body point cloud data is obtained with scan mode, the face of people is included at least per frame 3 D human body point cloud data
The point cloud data in portion;
Three-dimensional face detection is carried out respectively to the multiframe 3 D human body point cloud data, interception obtains multiple corresponding to difference
The Initial Face three-dimensional point cloud of frame;
Attitude updating iteratively is carried out to each Initial Face three-dimensional point cloud, until before current iteration and after iteration
The spin matrix of face three-dimensional point cloud is unit matrix;
Face three-dimensional point cloud sequence for all consecutive frames corresponding two Jing Guo attitude updating carries out coarse alignment operation,
The coarse alignment operation, which calculates to obtain, makes the corresponding points that face three-dimensional point cloud dimensionality reduction samples to obtain in data set most to quantity
Coarse alignment transformation matrix, wherein for corresponding points to referring in two datasets, distance is less than the point pair of 2 data resolutions;
Face three-dimensional point cloud after two coarse alignments operation corresponding to all consecutive frames carries out fine alignment operation, the essence
Alignment function to face three-dimensional point cloud by iteratively carrying out point coordinates conversion until the error of face three-dimensional point cloud meets
Predetermined condition;
Respectively from multiple and different visual angles respectively by it is all by fine alignment operate face three-dimensional point clouds visual point set into
Row mixing operation obtains super-resolution merging point cloud, and merges super-resolution merging point cloud that multiple mixing operation obtains to obtain
Super-resolution face three-dimensional point cloud.
Preferably, iteratively to each Initial Face three-dimensional point cloud carry out attitude updating, until current iteration before and
The spin matrix of face three-dimensional point cloud after iteration is that unit matrix includes:
Mean value and covariance matrix are calculated according to the point cloud vector of the face three-dimensional point cloud before iteration, wherein before iteration
The point cloud vector of face three-dimensional point cloud is P=[P1,P2,...,Pk,...,Pn], PkFor k-th point of coordinate vector, n is face
The point quantity of three-dimensional point cloud;
To covariance matrix carry out SVD decompose obtain feature vector composition matrix and eigenvalue cluster at diagonal matrix;
The point cloud vector of the face three-dimensional point cloud after attitude updating acquisition iteration is carried out based on following formula:
P'=V (P-m)
Wherein, P' is the face three-dimensional point cloud after iteration, and P is the face three-dimensional point cloud before iteration, and V forms for feature vector
Matrix, m be the mean value;
Above-mentioned iterative step is repeated until the matrix of feature vector composition is unit matrix.
Preferably, the coarse alignment, which operates, includes:
Using the corresponding face three-dimensional point cloud by attitude updating of previous frame as references object, with after frame it is corresponding without
The face three-dimensional point cloud for crossing attitude updating is that regulating object carries out characteristic matching, obtains multiple matched key points pair;
For each pair of key point to calculating corresponding point transfer matrix;
For the point transfer matrix of each key point pair, calculate between the point transfer matrix of other all key points pair
Matrix distance, the point transfer matrix progress least square fitting for choosing multiple key points pair that matrix distance meets thresholding constraint obtain
Take alternative point transfer matrix;
Point transformation is carried out to regulating object with each alternative point transfer matrix and calculates corresponding points quantity;
The corresponding points of the selection alternative point transfer matrix most to quantity is as coarse alignment transformation matrix.
Preferably, the fine alignment, which operates, includes:
To be used as with reference to object, with corresponding in rear frame in the corresponding face three-dimensional point cloud by coarse alignment operation of previous frame
Face three-dimensional point cloud by coarse alignment operation is regulating object, regulating object is projected to references object, projecting direction is pre-
Fixed direction of illumination;
For a subset of the projection point set of regulating object, is concentrated in the subpoint of references object and search for nearest point set;
It is calculated according to the subset and nearest point set and obtains fine alignment transformation matrix;
The all the points of regulating object are converted according to fine alignment transformation matrix, and calculate the point cloud after transformation and reference
The error of object, the error be transformation after point to reference point clouds on comprising its subpoint tangent plane square distance and;
Above-mentioned iterative step is repeated until the error is less than predetermined threshold.
Preferably, the mixing operation includes:
The visual point set at corresponding visual angle is projected into the corresponding coordinate plane in the visual angle;
By the region rasterizing of the coordinate plane, by the point in fallen with same grid merge into a merging point with
Obtain raster data, the merging point perpendicular to the coordinate plane coordinate for fallen with grid point the direction coordinate
Mean value;
To the raster data in the plane into row interpolation and filtering;
The raster data is mapped to three dimensions to obtain super-resolution merging point cloud.
Preferably, the multiple different visual angles include LOOK LEFT, LOOK RIGHT and preceding visual angle.
Second aspect, provides a kind of data processing equipment, including processor, and the processor is adapted for carrying out as described above
Method.
Four steps are merged by Face datection, human face posture correction, the alignment of point cloud, three visual angle surfaces, without to each
Direction of illumination is merged, and the method merged by projecting to two dimensional surface greatly improves the efficiency.Due to using first appearance
State corrects the strategy merged again, takes full advantage of the geometrical properties of face, has not only maintained the precision of data, but also reduces and ask
The complexity of topic.Simultaneously as not directly in three dimensions to cloud into row interpolation and filtering, but first project to two dimensional surface
Again into row interpolation and filtering, efficiency of algorithm is substantially increased.It, to the point cloud Pointing strategy of essence, both can avoid being absorbed in part most by slightly
It is excellent, and convergence rate can be accelerated.
Specific implementation mode
Below based on embodiment, present invention is described, but the present invention is not restricted to these embodiments.Under
Text to the present invention datail description in, it is detailed to describe some specific detail sections.Do not have for a person skilled in the art
The description of these detail sections can also understand the present invention completely.In order to avoid obscuring the essence of the present invention, well known method, mistake
There is no narrations in detail for journey, flow, element and circuit.
In addition, it should be understood by one skilled in the art that provided herein attached drawing be provided to explanation purpose, and
What attached drawing was not necessarily drawn to scale.
Unless the context clearly requires otherwise, "include", "comprise" otherwise throughout the specification and claims etc. are similar
Word should be construed as the meaning for including rather than exclusive or exhaustive meaning;That is, being containing for " including but not limited to "
Justice.
In the description of the present invention, it is to be understood that, term " first ", " second " etc. are used for description purposes only, without
It can be interpreted as indicating or implying relative importance.In addition, in the description of the present invention, unless otherwise indicated, the meaning of " multiple "
It is two or more.
Fig. 1 is the flow chart of the face three-dimensional point cloud super-resolution fusion method of the embodiment of the present invention.
As shown in Figure 1, the method includes:
Step S100, multiframe 3 D human body point cloud data is obtained with scan mode, at least per frame 3 D human body point cloud data
The point cloud data of face including people.
Step S200, three-dimensional face detection is carried out respectively to the multiframe 3 D human body point cloud data, interception obtains multiple
Corresponding to the Initial Face three-dimensional point cloud of different frame.
Step S300, attitude updating iteratively is carried out to each Initial Face three-dimensional point cloud, until before current iteration
Spin matrix with the face three-dimensional point cloud after iteration is unit matrix.
Step S400, the face three-dimensional point cloud sequence for all consecutive frames corresponding two Jing Guo attitude updating carries out thick
Alignment function, the coarse alignment operation, which calculates, to be obtained so that face three-dimensional point cloud dimensionality reduction samples to obtain the corresponding points pair in data set
The most coarse alignment transformation matrix of quantity, wherein for corresponding points to referring in two datasets, distance is less than 2 data resolutions
Point pair.
Step S500, the face three-dimensional point cloud after two coarse alignments operation corresponding to all consecutive frames carries out fine alignment behaviour
Make, the fine alignment operation carries out the error of face three-dimensional point cloud of the multiple point coordinates conversion after iteration by iterative manner
Meet predetermined condition.
Step S600, respectively from multiple and different visual angles respectively by it is all by fine alignment operate face three-dimensional point clouds can
Viewpoint collection carries out mixing operation, and the merging super-resolution merging point cloud that mixing operation obtains three times is to obtain super-resolution face
Three-dimensional point cloud, the mixing operation are merged with the resolution ratio higher than 3 D human body point cloud data in different face three-dimensional point clouds
Point is to obtain super-resolution merging point cloud.
Fig. 2 is the three-dimensional face point cloud carried out in the embodiment of the present invention in face three-dimensional point cloud super-resolution fusion process
Schematic diagram.As can be seen from FIG. 2, in the face three-dimensional point cloud super-resolution fusion process, multiframe three-dimensional point cloud passes through face
Detection, human face posture correct, by that slightly can obtain the resolution beyond sensor to smart two beans-and bullets shooter clouds alignment, the fusion of various visual angles surface
The super-resolution face three-dimensional point cloud of rate.
By Face datection, human face posture correction, by slightly merging four steps to smart two beans-and bullets shooter clouds alignment, various visual angles surface
Suddenly, without being merged to each direction of illumination, the method merged by projecting to two dimensional surface substantially increases effect
Rate.Due to the strategy merged again using first attitude updating, the geometrical properties of face are taken full advantage of, data are both maintained
Precision, and reduce the complexity of problem.Simultaneously as not directly in three dimensions to cloud into row interpolation and filtering, but
Two dimensional surface is first projected to again into row interpolation and filtering, substantially increases efficiency of algorithm.Meanwhile plan is directed at by the point cloud slightly to essence
Slightly, it not only can avoid being absorbed in local optimum, but also convergence rate can be accelerated.
For step S100, can be obtained by moving three dimension image by any existing acquiring three-dimensional images equipment
Equipment shooting obtains the multiframe 3 D human body point cloud data for embodying the multiple angle characters of face.
For step S200, existing various face recognition algorithms may be used and come from each frame 3 D human body point cloud data
Middle interception face part makes subsequent cloud fusion be absorbed in human face region to obtain face three-dimensional point cloud.Due to only with
Regional area in scene is merged, and the memory overhead and computation complexity needed for method substantially reduce, and are expected in part
Carry out the super-resolution fusion of higher precision.
In a preferred embodiment, document H.Wang, C.Wang, H.Luo et al..3-D Point may be used
Cloud Object Detection based on Supervoxel Neighborhood with Hough Forest
Framework[J].IEEE Journal of Selected Topics in Applied Earth Observations
and Remote Sensing.2015,8(4):The side of the Hough forest model test point cloud target proposed in 1570-1581.
Method carries out three-dimensional face detection.
For step S300, existing various face recognition algorithms may be used to carry out human face posture correction.
Preferably, document A.S.Mian, M.Bennamoun and R.Owens.An efficient may be used
multimodal 2D-3D hybrid approach to automatic face recognition[J].IEEE
Trans.Pattern Anal.Mach.Intell..2007,29(11):1927-1943. the PCA methods of middle proposition carry out face
Attitude updating.Specifically, the flow chart of human face posture correction is as shown in figure 3, step S300 includes:
Step S310, mean value m and covariance matrix C is calculated according to the point cloud vector of the face three-dimensional point cloud before iteration,
In, P=[P1,P2,...,Pk,...,Pn], PkFor k-th point of coordinate vector, n is the point quantity of face three-dimensional point cloud.
That is,With
Step S320, to covariance matrix carry out SVD decompose obtain feature vector composition matrix and eigenvalue cluster at
Covariance matrix is also decomposed by diagonal matrix:CV=DV, wherein V be feature vector composition matrix (it is spin moment
Battle array).
Step S330, the point cloud vector of the face three-dimensional point cloud after attitude updating acquisition iteration is carried out based on following formula:
P'=V (P-m)
Wherein, P' is the face three-dimensional point cloud after iteration, and P is the face three-dimensional point cloud before iteration, m is the mean value.
Step S340, judge whether spin matrix is unit matrix, if it is not, then return to step S310 is carried out next time
Otherwise iteration completes human face posture correction.
For step S400 and step S500, in two steps using from thick to thin the step of carry out.
Specifically, in step S400, using the face three-dimensional point cloud of the 1st frame as object is referred to, with the face three of the 2nd frame
Dimension point cloud is as regulating object, and adjustment regulating object is so that it is aligned with references object, then, then with the 2nd frame after coarse alignment
Face three-dimensional point cloud be references object, using the face three-dimensional point cloud of the 3rd frame as regulating object, progress the 3rd frame alignment, according to
It is secondary to analogize, until the corresponding face three-dimensional point cloud of all frames is aligned.
Preferably, consistent corresponding proof method (Consistent Correspondences may be used in S400
Verification, CCV) (Guo Yulan, point cloud local feature description and automatic object reconstruction identification technology research [D], 2015,
The National University of Defense technology, Changsha) carry out coarse alignment, it specifically includes:
Step S410, using the corresponding face three-dimensional point cloud by attitude updating of previous frame as references object, with the frame after
The corresponding face three-dimensional point cloud by attitude updating is that regulating object carries out characteristic matching, obtains multiple matched key points pair
Step S420, for each pair of key point to calculating corresponding point transfer matrix.
For key point pairFor kth to key point to direct transformation matrix.
Step S430, for the point transfer matrix of each key point pair, the point transformation with other all key points pair is calculated
Matrix distance between matrix, the point transfer matrix for choosing multiple key points pair that matrix distance meets thresholding constraint carry out minimum
Two, which multiply fitting, obtains alternative point transfer matrix.
The thresholding constraint is actually whether detection matrix distance is sufficiently small, that is, obtaining multiple points similar enough
It is average to seek it by least square fitting for transformation matrix.It will seek the average alternately point transfer matrix obtained.
Step S440, point transformation is carried out to regulating object with each alternative point transfer matrix and calculates corresponding points quantity.Such as
Upper described, corresponding points are in two datasets, distance is less than the point pair of 2 data resolutions.
Step S450, the corresponding points of the selection alternative point transfer matrix most to quantity is as coarse alignment transformation matrix.
That is, selection one is so that the maximum alternative point transfer matrix of corresponding points quantity comes after whole progress coarse alignment operation
Carry out coarse alignment.
Preferably, the present invention improves the fine alignment of step S500, using improved iteration nearest neighbor algorithm, tool
Body, step S500 includes:
Step S510, to be used as with reference to object in the corresponding face three-dimensional point cloud by coarse alignment operation of previous frame, with
The corresponding face three-dimensional point cloud by coarse alignment operation of frame is regulating object afterwards, and regulating object is projected to reference pair image point
Cloud, projecting direction are predetermined direction of illumination.
Step S520, for a subset of the projection point set of regulating object, search is concentrated in the subpoint of references object
Nearest point set.
Step S530, it is calculated according to the subset and nearest point set and obtains fine alignment transformation matrix.
Step S540, all the points of regulating object are converted according to fine alignment transformation matrix, and calculate transformation after
The error of point cloud and references object, the error are cutting flat with comprising its subpoint on the point to reference point clouds of the point cloud after transformation
The square distance in face and.
Step S550, whether error in judgement is less than predetermined threshold, if it is, return to step S510 is changed next time
Otherwise in generation, terminates smart calibration operation.
Above-mentioned iterative step is repeated until the error is less than predetermined threshold.
Since the spot projection in space has been arrived camera plane, thus, it is possible to accelerate to find the speed of nearest point set.Together
When, optimization object function is the point of all corresponding points pair to the sum of the distance of plane, rather than the distance of point-to-point.It is arrived using point
The distance of plane can accelerate convergence rate.It, can be linear by nonlinear object function when camera movement speed is less fast
Change, finally can obtain transformation matrix with the form of regular equation.Calculation amount is greatly decreased in terms of two as a result,.
For step S600, the embodiment of the present invention carries out fusion from three different visual angles and obtains three super-resolution fusions
Point cloud, then three super-resolution merging points are converged as super-resolution face three-dimensional point cloud.It only need to be enterprising on boundary when integrated
Row consistency treatment.Specifically, as shown in figure 4, step S600 is merged from preceding visual angle, LOOK LEFT and LOOK RIGHT.Every
Visible dots cloud is only merged in a visual angle, and for the face point cloud of regularization, left view point is actually left half of face
Point cloud data fusion, corresponding right viewpoint and preceding viewpoint it is corresponding be right half of face and positive face.In the fusion of a certain viewing angles
Operation specifically includes:
Step S610, the visual point set at corresponding visual angle is projected into the corresponding coordinate plane in the visual angle.
By taking LOOK LEFT as an example, in this step, visual point set is projected into yoz planes.
Step S620, by the region rasterizing of the coordinate plane, the point in fallen with same grid is merged into one
A merging point to obtain raster data, the merging point perpendicular to the coordinate plane coordinate for fallen with grid point at this
The mean value of the coordinate in direction.
In the example merged from LOOK LEFT, all points are merged by the way that yoz planar gates are formatted, meanwhile, quilt
The x coordinate of combined point is the mean value of the x coordinate of all points in the grid.
The operation for carrying out same place merging in other directions is similar with aforesaid way.
Step S630, to the raster data in the plane into row interpolation and filtering.
Specifically, in the example merged from LOOK LEFT, cubic algorithms may be used to the grid in yoz planes
Data carry out smothing filtering to eliminate hole and reduce noise into row interpolation.Not directly in three-dimensional space in the embodiment of the present invention
Between to cloud into row interpolation and filtering, but first project to two dimensional surface again into row interpolation and filtering, substantially increase algorithm effect
Rate.
Step S640, the raster data is mapped to three dimensions to obtain super-resolution merging point cloud.
It is merged as a result, from multiple visual angles, without being merged to each direction of illumination, by projecting to two dimensional surface
The method merged, greatly improves the efficiency.
Above-mentioned method and apparatus can be applied to data processing system, be executed by its processor.In specific embodiment party
Data structure and code described in formula are generally stored inside on computer readable storage medium, can store for calculating
Any equipment or medium for the code and/or data that machine system uses.Computer readable storage medium is including but not limited to volatile
Property memory, nonvolatile memory, magnetism and optical storage apparatus, such as disc driver, tape, CD (CD), DVD
(digital versatile disc or digital video disk) or currently known or that develops later be capable of store code and/or data
Other media.
Can by specific embodiment part describe method and process be embodied as code and/or data, the code and/
Or data are storable in computer readable storage medium as described above.When computer system is read and is executed computer-readable
When the code and/or data stored on storage medium, computer system execution is embodied as data structure and code and is stored in
Method and process in computer readable storage medium.
Furthermore, it is possible to include in hardware module or device by method described herein and process.These modules or device
It can include but is not limited to application-specific integrated circuit (ASIC) chip, field programmable gate array (FPGA), executed in specific time
The special or shared processor of specific software module or one section of code and/or other are currently known or that develops later programmable patrols
Collect equipment.When activating hardware module or device, they execute the method and process being included therein.
The foregoing is merely the preferred embodiment of the present invention, are not intended to restrict the invention, for those skilled in the art
For, the present invention can have various modifications and changes.It is all within spirit and principles of the present invention made by any modification, equivalent
Replace, improve etc., it should all be included in the protection scope of the present invention.