CN103714574A - GPU acceleration-based sea scene modeling and real-time interactive rendering method - Google Patents

Abstract

The invention discloses a GPU acceleration-based sea scene modeling and real-time interactive rendering method. The method comprises the following steps that: (1) preprocessing such as water body and stratum separation, target detection and denoising enhancement, is performed on sea scene sonar data; (2) ray casting method-based interactive sea scene three-dimensional body data visualization modeling can be realized through using GPU acceleration; and (3) sea target sonar data surface extraction and real-time rendering are realized based on the GPU acceleration. With the GPU acceleration-based sea scene modeling and real-time interactive rendering method of the invention adopted, the problem that existing visualization algorithms are not suitable for scene sonar data rapid modeling and interactive visualization analysis can be solved; and the influence of subjective factors can be reduced, such that initial visualization results of the data can be more accurate. According to the GPU acceleration-based sea scene modeling and real-time interactive rendering method of the invention, a curved surface subdivision method is utilized to perform optimization processing on target surface extraction results obtained through a marchingcube algorithm, such that smooth rendering results can be obtained. Finally, sea scene modeling and real-time interactive rendering can be realized.

Description

A kind of marine scene modeling and real-time, interactive method for drafting accelerating based on GPU

Technical field

The present invention relates to that three-dimensional data is field visualized to be obtained and field of drawing with three-dimensional geometry surface, relate in particular to a kind of marine scene modeling and real-time, interactive method for drafting accelerating based on GPU.

Background technology

Marine sonar data visualization technique is mainly to utilize marine sonar data as information source, adopts certain data processing and three-dimensional modeling method, sets up three-dimensional visualization and the interaction analysis system of underwater sonar data.The development and utilization of current ocean resources is focuses of new century economic growth, and the detection means of ocean is also being enriched constantly.In the last few years, for the modeling and rendering work of Ocean Scenes, obtained many progress, and in the sense of reality simulation of scene, also obtained considerable success, corresponding fluid simulation method, has also obtained application widely.Yet on the other side: drafting and the modeling work progress of marine (under extra large surface) scene are on the low side, one of its reason is huge due to calculated amount, and former two is therefore because the inexactness in sonar data source.This technology has been carried out pre-service and optimization for marine scene sonar data, the accuracy of the data that reduced the noise in data and improved, has adopted morphologic image Segmentation Technology to process respectively to improve accuracy of identification to the water body in seabed and Lands Department.Used GPU speed technology to make to draw with target signature interaction analysis and reached real-time.

First introduce existing volume rendering method below:

1) ray cast (RayCasting) algorithm

RayCasting algorithm is with the pattern space carrier of final rendering result not, and from each pixel of carrier space, it is long that contrary viewpoint direction divergent-ray passes through three-dimensional data.Along ray, equidistantly sample, and according to the data in data fields, obtain the optical property of this sampled point.The optical property of all sampled points on every ray is superposeed according to optical model, calculate the pixel color space attribute of starting point, thereby obtain the grain distribution of visual carrier.This algorithm need to travel through 3 d data field, and when viewpoint changes, sampled point need recalculate, and has caused huge calculated amount.

2) snow throwing ball method

This algorithm and RayCasting algorithm difference be that each long voxel of its traversal three-dimensional data calculates.It utilizes a function to represent that each voxel projects the intensity that affects of visible space, and we are generally referred to as Footprint.Calculate the coverage of each voxel projection, thereby calculate the overall contribution to image, stack is to generate last image.This algorithm just look like each voxel as throwing away snowball on screen, hit residual certain vestige on screen, thereby gain the name.Such computing method, although let us only consider and show relevant voxel, thereby greatly reduced data volume, the conversion of viewpoint causes all stacks to be recalculated, and still cannot show in real time.

Conventional algorithm also has a lot: wrong contact transformation method, three-D grain reflection method etc.But traditional ray cast (RayCasting) algorithm has three large shortcomings: the one, and transport function (transfer function) need to regulate by manual intervention, wants to obtain a reasonable result, often needs the longer time of cost; The 2nd, traditional RayCasting algorithm is realized at CPU end often, and interaction is poor; The 3rd, traditional algorithm is used a tranfer function to whole volume data, and for classifying as this class of sonar, the data of processing cannot be processed.

Equally, traditional Marching Cube method is not suitable for sonar data yet, and traditional marching cube method threshold value when extracting surface is more single, and for these class data of similar sonar, often target value is within the scope of one, and traditional method cannot be processed; If need to obtain in addition a smoother result, traditional method often needs the cubical resolution of increase, and this will increase the data volume of processing greatly.

In addition, sonar 3 d data field is larger owing to being subject to the impact of environment.Be subject to reverberation, defocus, the factor such as noise affects, the direct visual visual image accurately that cannot obtain enough of general data field.Sum up sonar data and mainly comprise following characteristics:

1. the edge contour in data fields is very fuzzy, and presents comparatively serious disappearance or deformation;

2. be subject to the impact that acoustic signals wavelength is long, frequency is lower, result lacks detail of the high frequency, causes a lot of edge recognizers to realize;

3. sonar signal noise profile is wider, although the form of expression is single, is mainly high-frequency pulse signal, because fluctuating range is relatively large, still very large on the impact of result images.

4. the echo signal in sonar data field is relatively single to the response intensity of sound wave, so gray level is relatively less.But noise data is various owing to producing reason, so all occur in multi-grey level.

5. the propagation of sonar signal in the middle of water body and stratum is different, can produce different signal attributes,

In sum, the visual of sonar 3 d data field is difficult to use existing method to draw and rebuild.

Due to this type of sonar data, the direct application of traditional three-dimensional data visualized algorithm obtains effect and is had a strong impact on.And traditional denoise algorithm (as medium filtering, bilateral filtering, histogram equalization etc.) is not all suitable for sonar data.And disposal route based on small echo, the Processing Algorithm based on fuzzy theory, neural network algorithm etc. are also difficult to deal with excessive data volume.

Reason based on above, it is visual that method is in the past difficult to obtain perfect real-time sonar data.

Invention Inner holds

The object of the invention is to overcome the deficiencies in the prior art, a kind of marine scene modeling and real-time, interactive method for drafting accelerating based on GPU is provided.

The marine scene modeling and the real-time, interactive method for drafting that based on GPU, accelerate comprise the steps:

1) marine scene sonar data adopts morphological method and filtering to carry out pre-service to raw data for sonar data feature, is divided into the separation of water body stratum, the target detection based on gradient based on binaryzation and strengthens based on gradient statistics and the denoising of filtering;

2) utilize GPU to accelerate Ray Tracing Algorithm, the statistics based on data generates transport function, and adds real-time, interactive function, the marine scene three-dimensional data visual modeling that realization can be mutual;

3) use marchingCube method and the GPU accelerated method based on range detection, realize marine target sonar data surface extraction and real-time rendering.

Step 1) described water body stratum is separated into:

2.1) by LOG, convert and improve data area, widen active sonar data area;

2.2) adopt the quick denoising of gaussian filtering to weaken the impact of noise in water body, and retained the marginal original character in water body stratum;

2.3) by the opening operation in the morphological operation on gray scale territory, remove target and noise object, only retain water body and formation data;

2.4) that based on water body, from stratum, sonar signal is reacted is different, adopts robotization Threshold segmentation water body stratum, and threshold calculations formula is:

${\mathrm{\σ}}_{\mathrm{within}}^2\left(T\right)=\frac{N_{(\mathrm{Fgrnd})\left(T\right)}}{N}{\mathrm{\σ}}_{\mathrm{Fgrnd}}^2\left(T\right)+\frac{N_{\mathrm{Bgrnd}}\left(T\right)}{N}{\mathrm{\σ}}_{\mathrm{grnd}}^2\left(T\right)$

${\mathrm{\σ}}_{\mathrm{between}}^2\left(T\right)={\mathrm{\σ}}^2-{\mathrm{\σ}}_{\mathrm{within}}^2\left(T\right)$

$=\left(\underset{x,y}{\mathrm{\Σ}}{f}^2\right[x,y\left]\right)-\frac{N_{\mathrm{Fgrnd}}}{N}\left(\underset{x,y\∈\mathrm{Fgrnd}}{\mathrm{\Σ}}{f}^2\right[x,y]-{\mathrm{\μ}}_{\mathrm{Fgrnd}}^2)-\frac{N_{\mathrm{Bgrnd}}}{N}(\underset{x,y\∈B_{\mathrm{grnd}}}{\mathrm{\Σ}}-{\mathrm{\μ}}^2)$

$=-{\mathrm{\&mu;}}^2+\frac{N_{\mathrm{Fgrnd}}}{N}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="Fgrnd"\; filenames="131219143114.png"\; state="\{\{state\}\}">Fgrnd</figure-callout>}}^2+\frac{N_{\mathrm{Bgrnd}}}{N}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="Bgrnd"\; filenames="131219143114.png"\; state="\{\{state\}\}">Bgrnd</figure-callout>}}^2$

$=\frac{N_{\mathrm{Fgrnd}}}{N}{({\mathrm{\&mu;}}_{\mathrm{Fgrnd}}^2-{\mathrm{\&mu;}}^2)}^2+\frac{N_{\mathrm{Bgrnd}}}{N}{({\mathrm{\&mu;}}_{\mathrm{Bgrnd}}^2-{\mathrm{\&mu;}}^2)}^2$

$=\frac{N_{\mathrm{Fgrnd}}\left(T\right)N_{\mathrm{Brnd}}\left(T\right)}{N}{({\mathrm{\&mu;}}_{\mathrm{Fgrnd}}\left(T\right)-{\mathrm{\&mu;}}_{\mathrm{Bgrnd}}\left(T\right))}^2$

In formula, σ represents variance, and T is for band table data fields, σ _within(T) rear prospect variance and the summation of background variance weighted, σ are cut apart in representative data field _between(T) represent the variance of prospect and background, the difference of prospect and background is measured, and N representative data point number, as N _fgrnd(T) foreground data point number in the T of representative data field, N _fgrnd(T) number of background data point in the T of representative data field, N representative data field whole-sample point number.F[x, y] be positioned at the intensity level that point [x, y] is located in representative data field, three-dimensional widenable to f[x, y, z], μ represents all data mean value, μ _fgrndand μ _bgrndrespectively with the just mean value of prospect and background data point;

Step 1) described target detection and denoising enhancing are:

3.1) calculate the gradient data field of 3 d data field;

3.2) statistical gradient histogram data field, finds that data fields has double-hump characteristics or multimodal characteristic;

3.3) based on histogram of gradients, carry out the crest that thresholding is partitioned into intensity maximum, obtain target object;

3.4) other crest data are carried out to gaussian filtering Fuzzy Processing, obtain denoising effect.

Described step 2) be: use GPU to optimize following steps:

4.1) the sonar data field based on after gradient data field and denoising is calculated weighted sum and is obtained new data field

C _trans＝αC _volumedata+βC _gradientdata

C wherein _volumedatarepresent source data field, C _gradientdatarepresent gradient data field, α and β are weighting coefficient, C _transthe synthetic result of representative;

4.2) the gradient field statistical computation based on new data field obtains the transport function that can change alternately;

4.3) take viewpoint as starting point, to RayCasting carrier cube face projection radiation, equidistantly sample and carry out color addition mixing;

c _o＝α _sc _s+(1-α _s)c _d

α wherein _srepresent voxel transparency, by transport function, provided.C _sfor voxel intensities, C _drepresent the intensity level that superposeed, C _othe intensity level that representative newly obtains, until the opacity of sampled intensity value surpasses 1, or sampling depth stops calculating over data fields.

The marchingCube method based on range detection and GPU accelerated method described in step 3) are:

5.1) cubical eight points are encoded to 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴, 2 ⁵, 2 ⁶, 2 ⁷, the pattern that cube is all is carried out to code index, and then all patterns is placed on to processing and the inquiry of carrying out GPU end in pattern texture texture_Pattern;

5.2) utilize separate characteristic between the parallel processing capability of GPU and cube, all cubes of data fields are classified, systematic searching formula is:

Wherein index is the index of current cube classification, vertex[i] be the value of the data fields of cube summit i, [isovalue1, isovalue2] be the possible threshold value dynamic range of target, utilize index in texture_Pattern, to search the number of vertices numofvertex. that cube can produce and then upgrade corresponding state variable:

cubeVertexNumber[indexofcube]＝numofvertex

$\mathrm{cubeOccupied}\left[\mathrm{indexofcube}\right]=\left\{\begin{array}{cc}0& \mathrm{numbofvertex}<=0\\ 1& \mathrm{numberofvertex}>0\end{array}\right.$

Wherein indexofcube is the coding of current cube on GPU;

5.3) utilize Prefix-Sum algorithm to process cubeOccupid array, in cubeOccupidScan array, preserve the current cube that comprises summit in all cubical sequences that comprise summit;

5.4) result of utilizing Prefix-Sum to process is compressed processing to cube, and the cube that does not contain summit is rejected away, and compression method is:

compactCube[cubeOccupidScan[i]]＝i if(cubeOccupid[i]＞0)

Wherein i is current cubical index; After processing finishes, what in compactCube array, preserve is all cubical index that comprise tri patch;

5.5) utilize cube 256 in the pattern cube that respectively each contained to tri patch process, produce final summit, normal direction and tri patch, computing formula is:

$\mathrm{vertex}=v_1+(\mathrm{isovalue}-\mathrm{value}1)\frac{v_2-v_1}{{\mathrm{value}}_2-{\mathrm{value}}_1}$

Wherein vertex is the apex coordinate of generation, v ₁, v ₂be the coordinate of two end points, value ₁, value ₂be the volume data value of two end points, the mean value of the threshold range that Isovalue is volume data:

isovalue＝(isovalue1+isovalue2)/2

Normal is the normal direction of tri patch:

normal＝(v2-v1)×(v3-v2)

5.6) utilize openGL to show final result, added Tessellation Shader, the grade of surface subdivision is set, obtain level and smooth final demonstration result;

5.7) result marching cube being obtained is preserved with obj form.

The present invention has adopted the pre-service work of data, and sonar data is identified and revised to improve identity.Based on morphologic operation, carry out the separation of water body stratum, avoided the impact of HF noise signal, more rapidly, accurately, and morphologic operation is suitable for GPU and accelerates in the processing that makes final algorithm, and having solved classic method cannot be in real time, be subject to sonar data low resolution to affect large difficult point.And separating resulting is recorded in to 2 d texture the inside, makes follow-up can use different transport functions to regulate to different regions herein.The visual key of 3 d data field is to find a good transport function, nowadays has the transport function generating modes such as genetic algorithm.But generally belong to the better effect design of subjective searching, but not analysis based on data.The transport function producing may not necessarily science be expressed target object accurately.

The transport function generating mode based on gradient statistics that the present invention proposes, affected by sonar pulse signal noise little, and the completely analysis based on data, human assistance to be to give sampled data set optical property, can produce more science information visualization image accurately.

The detection method based on scope that the present invention proposes makes that marchingCube's is on probation wider; The method of the increase volume data resolution of this paper abandoning tradition.With putting of surface subdivision, the result of marching cube is processed.Make final result of drawing smoother.

Accompanying drawing explanation

Fig. 1 is used original image directly to use morphology binarization segmentation water body stratum to obtain result;

Fig. 2 is the binarization segmentation result of using after binaryzation opening operation is processed;

Fig. 3 is the binarization segmentation result after repeatedly opening operation processing;

Fig. 4 is based on morphologic edge detection results;

Fig. 5 is that segmentation result and source data compare;

Fig. 6 is source data process invention process, but omits denoising process, the result obtaining;

Fig. 7 is source data process invention process, but through denoising process, the result obtaining;

Fig. 8 is histogram of gradients statistics;

Fig. 9 is the visable representation of intensity histogram statistics and transport function;

Figure 10 is light projecting algorithm drawing result;

Figure 11 is the initial results of marchcube.

Embodiment

The present invention solves the problem that general three-dimensional data technology cannot be processed in real time magnanimity marine scene sonar data visual modeling and the quick man-machine interaction of realization of goal is analyzed, proposition is the Data Preprocessing Technology with morphological image operation based on gradient statistics, with the light projective technique alternately in real time based on GPU, and the sonar data surface rendering technology based on Marching Cube.

The marine scene modeling and the real-time, interactive method for drafting that based on GPU, accelerate comprise the steps:

1) marine scene sonar data adopts morphological method and filtering to carry out pre-service to raw data for sonar data feature, is divided into the separation of water body stratum, the target detection based on gradient based on binaryzation and strengthens based on gradient statistics and the denoising of filtering;

2) utilize GPU to accelerate Ray Tracing Algorithm, the statistics based on data generates transport function, and adds real-time, interactive function, the marine scene three-dimensional data visual modeling that realization can be mutual;

3) use marchingCube method and the GPU accelerated method based on range detection, realize marine target sonar data surface extraction and real-time rendering.

Step 1) described water body stratum is separated into:

2.1) by LOG, convert and improve data area, widen active sonar data area;

2.2) adopt the quick denoising of gaussian filtering to weaken the impact of noise in water body, and retained the marginal original character in water body stratum;

2.3) by the opening operation in the morphological operation on gray scale territory, remove target and noise object, only retain water body and formation data;

2.4) that based on water body, from stratum, sonar signal is reacted is different, adopts robotization Threshold segmentation water body stratum, and threshold calculations formula is:

${\mathrm{\&sigma;}}_{\mathrm{within}}^2\left(T\right)=\frac{N_{\mathrm{Fgrnd}\left(T\right)}}{N}{\mathrm{\&sigma;}}_{\mathrm{Fgrnd}}^2\left(T\right)+\frac{N_{\mathrm{Bgrnd}}\left(T\right)}{N}{\mathrm{\&sigma;}}_{\mathrm{grnd}}^2\left(T\right)$

${\mathrm{\&sigma;}}_{\mathrm{between}}^2\left(T\right)={\mathrm{\&sigma;}}^2-{\mathrm{\&sigma;}}_{\mathrm{within}}^2\left(T\right)$

$=\left(\underset{x,y}{\mathrm{\&Sigma;}}{f}^2\right[x,y\left]\right)-\frac{N_{\mathrm{Fgrnd}}}{N}\left(\underset{x,y\&Element;\mathrm{Fgrnd}}{\mathrm{\&Sigma;}}{f}^2\right[x,y]-{\mathrm{\&mu;}}_{\mathrm{Fgrnd}}^2)-\frac{N_{\mathrm{Bgrnd}}}{N}(\underset{x,y\&Element;B_{\mathrm{grnd}}}{\mathrm{\&Sigma;}}-{\mathrm{\&mu;}}^2)$

$=-{\mathrm{\&mu;}}^2+\frac{N_{\mathrm{Fgrnd}}}{N}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="Fgrnd"\; filenames="131219143114.png"\; state="\{\{state\}\}">Fgrnd</figure-callout>}}^2+\frac{N_{\mathrm{Bgrnd}}}{N}{\mathrm{\&mu;}}_{\mathrm{<figure-callout\; id="2"\; label="Bgrnd"\; filenames="131219143114.png"\; state="\{\{state\}\}">Bgrnd</figure-callout>}}^2$

$=\frac{N_{\mathrm{Fgrnd}}}{N}{({\mathrm{\&mu;}}_{\mathrm{Fgrnd}}^2-{\mathrm{\&mu;}}^2)}^2+\frac{N_{\mathrm{Bgrnd}}}{N}{({\mathrm{\&mu;}}_{\mathrm{Bgrnd}}^2-{\mathrm{\&mu;}}^2)}^2$

$=\frac{N_{\mathrm{Fgrmd}}\left(T\right)N_{\mathrm{Brnd}}\left(T\right)}{N}{({\mathrm{\&mu;}}_{\mathrm{Fgrnd}}\left(T\right)-{\mathrm{\&mu;}}_{\mathrm{Bgrnd}}\left(T\right))}^2$

In formula, σ represents variance, and T is for band table data fields, σ _within(T) rear prospect variance and the summation of background variance weighted, σ are cut apart in representative data field _between(T) represent the variance of prospect and background, the difference of prospect and background is measured, and N representative data point number, as N _fgrnd(T) foreground data point number in the T of representative data field, N _fgrnd(T) number of background data point in the T of representative data field, N representative data field whole-sample point number.F[x, y] be positioned at the intensity level that point [x, y] is located in representative data field, three-dimensional widenable to f[x, y, z], μ represents all data mean value, μ _fgrndand μ _bgrndrespectively with the just mean value of prospect and background data point;

Use the shape of the M under water cable sonar data of actual acquisition carry out step 1) described water body stratum separation, obtaining following result: Fig. 1 is not pass through the direct water body of any processing stratum separating resulting, water body represents with black, stratum represents by white, and in figure, M shaped objects is the test cable target of arranging under water.Fig. 2 is the design sketch obtaining after opening operation is processed, noise data with should not affect the target object of cutting apart and well by algorithm, eliminated, continue to use opening operation, obtain data as shown in Figure 3.Use obtains the partitioning boundary as Fig. 4 based on morphologic edge detecting technology, and figure compares as shown in Figure 5 with source data.The method that the present invention uses well adapts to the feature of sonar data, obtains the good image of cutting apart.

Step 1) described target detection and denoising enhancing are:

3.1) calculate the gradient data field of 3 d data field;

3.2) statistical gradient histogram data field, finds that data fields has double-hump characteristics or multimodal characteristic;

3.3) based on histogram of gradients, carry out the crest that thresholding is partitioned into intensity maximum, obtain target object;

3.4) other crest data are carried out to gaussian filtering Fuzzy Processing, obtain denoising effect.

Use the same shape of M under water cable sonar data, carry out step 1) described target detection and denoising strengthen and can obtain following result: as shown in Figure 6, if ignore denoising step, the 3 d data field real-time rendering image only obtaining by all the other flow processs, only shows water body part in figure.Add after denoise algorithm, most of noise is eliminated, and as shown in Figure 7, M type object noise spot around significantly reduces, and in order better to demonstrate target and noise, has adopted Transparent color while playing up water body in Fig. 7.

Described step 2) be: use GPU to optimize following steps:

4.1) the sonar data field based on after gradient data field and denoising is calculated weighted sum and is obtained new data field

C _trans＝αC _volumedata+βC _gradientdata

C wherein _volumedatarepresent source data field, C _gradientdatarepresent gradient data field, α and β are weighting coefficient, C _transthe synthetic result of representative;

4.2) the gradient field statistical computation based on new data field obtains the transport function that can change alternately;

4.3) take viewpoint as starting point, to RayCasting carrier cube face projection radiation, equidistantly sample and carry out color addition mixing;

c _o＝α _sc _s+(1-α _s)c _d

α wherein _srepresent voxel transparency, by transport function, provided.C _sfor voxel intensities, C _drepresent the intensity level that superposeed, C _othe intensity level that representative newly obtains, until the opacity of sampled intensity value surpasses 1, or sampling depth stops calculating over data fields.

Use the same shape of M under water cable sonar data carry out step 2), obtain following result: the gradient field statistic histogram that statistics obtains new data field as shown in Figure 8, the transport function generating according to this as shown in Figure 9, utilize this transport function to render the visual image of seabed data fields, as shown in figure 10.In this example in order to highlight target object, by step 4.1) in β be taken as 1, and α is taken as 0, calculate.In actual applications, suitable change can obtain better display effect.

The marchingCube method based on range detection and GPU accelerated method described in step 3) are:

5.1) cubical eight points are encoded to 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴, 2 ⁵, 2 ⁶, 2 ⁷, the pattern that cube is all is carried out to code index, and then all patterns is placed on to processing and the inquiry of carrying out GPU end in pattern texture texture_Pattern;

5.2) utilize separate characteristic between the parallel processing capability of GPU and cube, all cubes of data fields are classified, systematic searching formula is:

Wherein index is the index of current cube classification, vertex[i] be the value of the data fields of cube summit i, [isovalue1, isovalue2] be the possible threshold value dynamic range of target, utilize index in texture_Pattern, to search the number of vertices numofvertex. that cube can produce and then upgrade corresponding state variable:

cubeVertexNumber[indexofcube]＝numofvertex

$\mathrm{cubeOccupied}\left[\mathrm{indexofcube}\right]=\left\{\begin{array}{cc}0& \mathrm{numbofvertex}<=0\\ 1& \mathrm{numberofvertex}>0\end{array}\right.$

Wherein indexofcube is the coding of current cube on GPU;

5.3) utilize Prefix-Sum algorithm to process cubeOccupid array, in cubeOccupidScan array, preserve the current cube that comprises summit in all cubical sequences that comprise summit;

5.4) result of utilizing Prefix-Sum to process is compressed processing to cube, and the cube that does not contain summit is rejected away, and compression method is:

compactCube[cubeOccupidScan[i]]＝i if(cubeOccupid[i]＞0)

Wherein i is current cubical index; After processing finishes, what in compactCube array, preserve is all cubical index that comprise tri patch;

5.5) utilize cube 256 in the pattern cube that respectively each contained to tri patch process, produce final summit, normal direction and tri patch, computing formula is:

$\mathrm{vertex}=v_1+(\mathrm{isovalue}-\mathrm{value}1)\frac{v_2-v_1}{{\mathrm{value}}_2-{\mathrm{value}}_1}$

Wherein vertex is the apex coordinate of generation, v ₁, v ₂be the coordinate of two end points, value ₁, value ₂be the volume data value of two end points, the mean value of the threshold range that Isovalue is volume data:

isovalue＝(isovalue1+isovalue2)/2

Normal is the normal direction of tri patch:

normal＝(v2-v1)×(v3-v2)

5.6) utilize openGL to show final result, added Tessellation Shader, the grade of surface subdivision is set, obtain level and smooth final demonstration result;

5.7) result marching cube being obtained is preserved with obj form.

Because M shape cable does not have border intuitively, in the time of the marchingCube method based on range detection described in step 3) and GPU accelerated method, use submerged wreck to receive data, obtaining following result: Figure 11 is a drawing result of the present invention.The present invention draws out a smoother result in the situation that not improving three-dimensional data resolution.

Claims (5)

1. the marine scene modeling and the real-time, interactive method for drafting that based on GPU, accelerate, is characterized in that comprising the steps:

1) marine scene sonar data adopts morphological method and filtering to carry out pre-service to raw data for sonar data feature, is divided into the separation of water body stratum, the target detection based on gradient based on binaryzation and strengthens based on gradient statistics and the denoising of filtering;

2) utilize GPU to accelerate Ray Tracing Algorithm, the statistics based on data generates transport function, and adds real-time, interactive function, the marine scene three-dimensional data visual modeling that realization can be mutual;

3) use marchingCube method and the GPU accelerated method based on range detection, realize marine target sonar data surface extraction and real-time rendering.

2. a kind of marine scene modeling and real-time, interactive method for drafting accelerating based on GPU according to claim 1, is characterized in that step 1) described in water body stratum be separated into:

2.1) by LOG, convert and improve data area, widen active sonar data area;

2.2) adopt the quick denoising of gaussian filtering to weaken the impact of noise in water body, and retained the marginal original character in water body stratum;

2.3) by the opening operation in the morphological operation on gray scale territory, remove target and noise object, only retain water body and formation data;

2.4) that based on water body, from stratum, sonar signal is reacted is different, adopts robotization Threshold segmentation water body stratum, and threshold calculations formula is:

In formula, σ represents variance, and T is for band table data fields, σ _within(T) rear prospect variance and the summation of background variance weighted, σ are cut apart in representative data field _between(T) represent the variance of prospect and background, the difference of prospect and background is measured, and N representative data point number, as N _fgrnd(T) foreground data point number in the T of representative data field, N _fgrnd(T) number of background data point in the T of representative data field, N representative data field whole-sample point number.F[x, y] be positioned at the intensity level that point [x, y] is located in representative data field, three-dimensional widenable to f[x, y, z], μ represents all data mean value, μ _fgrndand μ _bgrndrespectively with the just mean value of prospect and background data point.

3. a kind of marine scene modeling and real-time, interactive method for drafting accelerating based on GPU according to claim 1, is characterized in that step 1) described target detection and denoising enhancing be:

3.1) calculate the gradient data field of 3 d data field;

3.2) statistical gradient histogram data field, finds that data fields has double-hump characteristics or multimodal characteristic;

3.3) based on histogram of gradients, carry out the crest that thresholding is partitioned into intensity maximum, obtain target object;

3.4) other crest data are carried out to gaussian filtering Fuzzy Processing, obtain denoising effect.

4. according to a kind of marine scene modeling and the real-time, interactive method for drafting accelerating based on GPU described in claim 1 or 3, it is characterized in that described step 2) be: use GPU to optimize following steps:

4.1) the sonar data field based on after gradient data field and denoising is calculated weighted sum and is obtained new data field

C _trans＝αC _volumedata+βC _gradientdata

C wherein _volumedatarepresent source data field, C _gradientdatarepresent gradient data field, α and β are weighting coefficient, C _transthe synthetic result of representative;

4.2) the gradient field statistical computation based on new data field obtains the transport function that can change alternately;

4.3) take viewpoint as starting point, to RayCasting carrier cube face projection radiation, equidistantly sample and carry out color addition mixing;

c _o＝α _sc _s+(1-α _s)c _d

α wherein _srepresent voxel transparency, by transport function, provided.C _sfor voxel intensities, C _drepresent the intensity level that superposeed, C _othe intensity level that representative newly obtains, until the opacity of sampled intensity value surpasses 1, or sampling depth stops calculating over data fields.

5. a kind of marine scene modeling and real-time, interactive method for drafting accelerating based on GPU according to claim 1, is characterized in that the marchingCube method based on range detection and the GPU accelerated method described in step 3) is:

5.1) cubical eight points are encoded to 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴, 2 ⁵, 2 ⁶, 2 ⁷, the pattern that cube is all is carried out to code index, and then all patterns is placed on to processing and the inquiry of carrying out GPU end in pattern texture texture_Pattern;

5.2) utilize separate characteristic between the parallel processing capability of GPU and cube, all cubes of data fields are classified, systematic searching formula is:

Wherein index is the index of current cube classification, vertex[i] be the value of the data fields of cube summit i, [isovalue1, isovalue2] be the possible threshold value dynamic range of target, utilize index in texture_Pattern, to search the number of vertices numofvertex. that cube can produce and then upgrade corresponding state variable:

cubeVertexNumber[indexofcube]＝numofvertex

Wherein indexofcube is the coding of current cube on GPU;

5.3) utilize Prefix-Sum algorithm to process cubeOccupid array, in cubeOccupidScan array, preserve the current cube that comprises summit in all cubical sequences that comprise summit;

5.4) result of utilizing Prefix-Sum to process is compressed processing to cube, and the cube that does not contain summit is rejected away, and compression method is:

compactCube[cubeOccupidScan[i]]＝i if(cubeOccupid[i]＞0)

Wherein i is current cubical index; After processing finishes, what in compactCube array, preserve is all cubical index that comprise tri patch;

5.5) utilize cube 256 in the pattern cube that respectively each contained to tri patch process, produce final summit, normal direction and tri patch, computing formula is:

Wherein vertex is the apex coordinate of generation, v ₁, v ₂be the coordinate of two end points, value ₁, value ₂be the volume data value of two end points, the mean value of the threshold range that Isovalue is volume data:

isovalue＝(isovalue1+isovalue2)/2

Normal is the normal direction of tri patch:

normal＝(v2-v1)×(v3-v2)

5.6) utilize openGL to show final result, added Tessellation Shader, the grade of surface subdivision is set, obtain level and smooth final demonstration result;

5.7) result marching cube being obtained is preserved with obj form.

Images