Summary of the invention
Propose a kind of new depth map fusion and process the depth map comprising noise, the method base with some cloud filter method
In hypothesis below: estimated that the body surface obtained local should be with the thing obtained by multi-amplitude deepness image fusion by individual depth map
Surface is similar, but the geometric properties of a noise spot cloud is unstable, and therefore, the similarity of surface character is deep for merging
Degree figure provides clue.For measuring similarity, the present invention is extracted the surface of a kind of brand-new design from local with global point cloud
Geometric feature description;Meanwhile, in order to increase the integrity of reconstruction, the present invention make use of again method based on figure to merge many chis
Degree depth map;The particular content of the present invention can be divided into five steps.
The calculating of depth map, the present invention uses method based on multiple dimensioned piece and calculates depth image, the method master
There are three features: depth map obtains after Multi-Scale Calculation;After selecting figure through the overall situation, each pixel can obtain
Oneself randomly choose partial view, these views can be propagated together with normal vector with the degree of depth;Coupling optimum, and not
It is meansigma methods, as Matching power flow function;This step is that ensuing fusion process provides input depth map.
Local and the generation of global point cloud, local surfaces partial points cloud in other words refers to the point produced by single width depth image
Cloud, and overall situation surface or some cloud refer to all local surfaces or some cloud are fused into single model, owing to depth map is to use
Multi-Scale Calculation, first with arest neighbors interpolation, they should be adjusted to best yardstick;For every width depth map, partial points
Cloud can obtain through back projection, for each some p, utilizes from its k nearest point, calculates covariance matrix, and method is as follows
State formula
Calculate the eigenvalue λ of covariance matrix subsequentlyjWith associated characteristic vector vj(j=0,1,2).Afterwards, it is possible to will
Minimal eigenvalue characteristic of correspondence vector is as the normal vector at p point.And surface curvature c can be estimated by following formula
Afterwards, depth map is filtered by each yardstick, by back projection's every width depth map can to a partial points cloud,
Subsequently, initial global point cloud just can obtain by being merged by all partial points clouds.
Surface geometrical property extracts, and geometrical constraint is based on the similarity between local surfaces and overall situation surface.But, existing
Feature Descriptor can not meet description local geometric features and the overall situation conforming requirement of geometric properties.Therefore, the present invention proposes
A kind of simple geometric feature description HTI(Histogram of Truncated Icosahedron rapidly), HTI structure
Being made up of 12 positive five faces, limit and 20 positive six sides, this description is based on query point and gives in radius surface method
The result of vector statistical, all normal vectors are divided into 32 groups according to the HTI face of its indication, when all of method in search area
By rectangular histogram normalization after Vector Processing, form final HTI feature.
Geometry is supported, the HTI Feature Descriptor that geometry support can utilize top to mention calculates, for an inquiry
Point p, implements arest neighbors and searches and calculate corresponding local HTI feature and overall situation HTI feature, define local feature hpAnd the overall situation
Feature hgBetween similarity g be
In addition to the similarity of geometric properties, local surface curvature c can also be comprised in geometry and support, in s, to be defined as follows
Herein, α represents that weight parameter, t are followed by the threshold value during optimization based on figure;Curvature is covered several
What mainly has two reasons in supporting: first, owing to curvature can be used to estimate the slickness of local surfaces, big geometry support
(little curvature) means to make surface more to level off to plane, meets infinitesimal facet and assumes;Second, due to Curvature Estimate it
Depending on the minimal eigenvalue of covariance matrix, easily by noise jamming, less curvature generally means that less noise.
Multiscale Fusion based on figure, in order to merge multiple dimensioned depth map, the present invention propose a kind of based on figure
Optimization Framework.Basic thought is, if point can provide extra information, then just retain this in rougher yardstick
Point, if not having extra information just to delete these points.More specifically, a cloud is set up the figure with two kinds of limit types, G=
(P, S, C), herein, P represents vertex set, and corresponding to the point in some cloud, S and C represents support limit and the conflict limit collection of orientation respectively
Close, in order to set up these limits, first to each summit one radius of influence of definition be
Herein, d represents the degree of depth at this point, and f represents the focal length of the corresponding video camera of this point, and n is the yardstick of depth map, x
Represent the inner product between standard shaft opposite direction and normal vector;Now, two distinct types of limit can be set up as follows: if
There is another p in the radius of influence of q on one point, and the radius of influence of p is less than q, p both can support q, it is also possible to q
Conflict.If q is also within the radius of influence of p, (p, q) is added in figure, because p is more accurate, and q is not on conflict limit
It is provided that more information;Otherwise support that (p, q) would be added in figure, because q can provide within the coverage of p on limit
More information;Afterwards, all points are sorted according to their radius of influence size, from the beginning of the point with least radius,
We only need to detect other points within its radius of influence, and therefore this Algorithms T-cbmplexity is O (nlog n).Along with figure
Setting up, can give its energy equation, target energy E is supported ENERGY E by geometryg, point supports ENERGY EsWith a conflict energy
EcConstituting, it is defined as follows:
β and γ is two weight parameter herein, and variable l is the vector of pixel selection label composition, for a p, label l (p)=1,
If p is selected, otherwise l (p)=0.Geometry supports ENERGY EgIt is defined as follows:
For the support energy of a single point p, the geometry of the point of all support p is used to support that the opposite number of sum defines, then
Use the radius of influence of p to supporting that energy is normalized, be defined as follows:
Finally conflict energy definition is as follows:
In order to minimize E (l), the energy of pointwise test point is contributed and changes label l, and based on above-mentioned formula, a single point can
The contribution of amount E can be measured by following formula
In actual applications, generally whole some cloud is carried out 10 scanning, can ensure that this optimization method is close to local best points.