CN108171725A - Object detection method under a kind of dynamic environment based on normal direction stream information - Google Patents
Object detection method under a kind of dynamic environment based on normal direction stream information Download PDFInfo
- Publication number
- CN108171725A CN108171725A CN201711421030.XA CN201711421030A CN108171725A CN 108171725 A CN108171725 A CN 108171725A CN 201711421030 A CN201711421030 A CN 201711421030A CN 108171725 A CN108171725 A CN 108171725A
- Authority
- CN
- China
- Prior art keywords
- frame image
- label
- field
- normal direction
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/215—Motion-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides object detection method under a kind of dynamic environment based on normal direction stream information, mainly includes three big steps, step 1：Ballot method estimates FOE；Step 2：Generate initial markers field；Step 3：Optimal Label Field is generated, so as to fulfill the detection of target.The sharpest edges of the present invention are that algorithm is relatively low to noise sensitivity, and the image texture and characteristics of image to photographed scene do not require, therefore are applicable in scene than wide, have broad prospects.
Description
Technical field
The present invention relates to a kind of method of target detection under dynamic environment based on normal direction stream information, to noise sensitivity compared with
Low, image texture and characteristics of image to photographed scene do not require, therefore are applicable in scene than wide, belong to computer vision
Field.
Background technology
Target detection refers to by analyzing the consecutive image sequence shot by motion cameras in different moments under dynamic environment
The movable information included in row detected the self-movement target in photographed scene from each frame image.Due to camera shooting
Machine is ceaselessly moving, therefore, comprising there are two types of movable informations in the consecutive image sequence of acquisition：The movable information of video camera and
The movable information of self-movement target in photographed scene.Both movable informations are mixed so that transporting in each frame image
The detection of moving-target becomes more complicated and difficulty.But in actual life, by the consecutive image sequence of mobile equipment acquisition
It is seen everywhere, and this technology is still in developing stage.Therefore, in the near future, this technology will have more extensive
Application field and good development prospect.At present, target detection technique is widely used in intelligent video monitoring, base under dynamic environment
In multiple fields such as the video frequency searchings, Visual Navigation of Mobile Robots, automatic Pilot of content.So have herein very important
Academic research meaning and actual application value.
So far, existing moving object detection algorithm is only applicable to the static situation of video camera, i.e. static field mostly
The detection of moving target in scape, only a small part detection algorithm can be suitable for the situation of camera motion, i.e. dynamic ring
The detection of target under border.These algorithms are broadly divided into：Frame differential method, the background subtraction of extension and the light of extension of extension
Stream method.The frame differential method of extension employs characteristic matching when estimating camera motion, thus its application range by
Great limitation, normally only can be suitably used for the scene that texture information enriches, has obvious characteristic.The background subtraction of extension is in structure
Image registration techniques are employed when building panorama background model image, since the technology is vulnerable to the influence of parallax, the target
Detection method is not used widely.The optical flow method of extension is complicated and easily affected by noise without obtaining extensively due to calculating
Application.
Invention content
The technology of the present invention solves the problems, such as：Overcome the deficiencies of the prior art and provide a kind of dynamic based on normal direction stream information
The method of target detection under environment, acquisition of this method based on normal direction stream information and be unfolded to implement, it is relatively low to noise sensitivity, it is right
The image texture and characteristics of image of photographed scene do not require, and are applicable in scene than wide.
The present invention is directed to the moving target under dynamic environment is detected.Its input quantity is calculated for adjacent two field pictures
The normal direction flow field arrived is exported to detect obtained moving target.The tool of target detection under dynamic environment is described below in the present invention
Body technique scheme.
Step 1：The 1st, 2 frame images in video are selected, the normal direction flow vector field of the 2nd frame image are calculated, while right
2nd frame image is split, and obtains the Color Segmentation result figure of the 2nd frame image；
Step 2：Using the normal direction flow vector field and Color Segmentation figure being calculated in step 1, using ballot method estimation
Obtain the pure translational movement information FOE of video camera；
Step 3：It, will with reference to the normal direction flow vector field of the 2nd frame image after the pure translational movement information FOE of video camera
Color Segmentation result figure carries out region division, is divided into background area and foreground area, foreground area, that is, motion target area, then right
Two different zones carry out the judgement of translation half-plane constraint, so as to generate the corresponding initial markers field of the 2nd frame image；
Step 4：The corresponding initial markers field of the 2nd frame image is optimized, the 2nd frame image of generation is corresponding most
Then excellent Label Field models the problem of optimal Label Field using markov random file theory, obtain target energy letter
Number；Algorithm being cut with figure again and solving the objective energy function, final realize detected moving target from the 2nd frame image；
Step 5：3rd frame image is split, obtains Color Segmentation result figure, it is corresponding optimal using the 2nd frame image
Label Field, with reference to the Color Segmentation result figure of the 3rd frame image, the corresponding initial markers field of the 3rd frame image of generation；
Step 6：The corresponding initial markers field of the 3rd frame image is optimized again, the 3rd frame image of generation corresponds to
Optimal Label Field, so that moving target detected from the 3rd frame image；
Step 7：It repeats Step 5: six, detects the target in other frame images, until will be in whole frame image
Moving object detection comes out.
The present invention having the beneficial effect that compared with present technology：Method is utilized during present invention estimation camera motion information
To the directional information of flow vector, FOE is obtained by ballot method.This method do not have to photographed scene any limitation and to noise not
Sensitivity can accurately be estimated to obtain the movable information of video camera.Therefore, method of the invention can be applied in wider field
Jing Zhong.
Description of the drawings
Fig. 1 is half-plane constraint principles figure；
Fig. 2 is the method for the present invention overhaul flow chart；
Fig. 3 is true picture sequential experimentation result figure, wherein：(a) it is original image；(b) it is based on Mean-Shift's
Color Segmentation result figure；(c) it is initial markers field；(d) it is optimal Label Field；(e) it is true value.
Specific embodiment
Technical solution for a better understanding of the present invention below further retouches the specific embodiment of the present invention
It states：
Institute's extracting method of the present invention is realized under based on the MATLAB 2013a language environments in 7 systems of Windows
's.This method utilizes the camera motion information and present frame that the normal direction flow field that is calculated by adjacent two field pictures estimates
Color Segmentation result figure, obtain an initial markers field, recycle the suitable target energy of markov random file the Theory Construction
Flow function cuts optimization algorithm with reference to figure and initial markers field is optimized, the optimal Label Field of present frame finally obtained, so as to examine
Measure self-movement target.
The present invention be a kind of innovation the dynamic environment based on normal direction stream information under target detection method, the stream of this method
Journey is as shown in Figure 2.This method includes the following steps：
First, ballot method estimation FOE
For the present invention under research trends environment during the method for target detection, consideration video camera makees the feelings of pure translational movement
Condition.The acquisition in normal direction flow field is described below and the method for camera translational motion parameter is estimated by normal direction flow field.
Normal direction flow vector field in image can be directly calculated by the gray matrix of adjacent two field pictures.Assuming that work as
Unit gradient direction vector on prior image frame at any one pixel p=(i, j) is n：
Wherein [Ex Ey]TRepresent the shade of gray direction at pixel p.By the definition of normal direction flow vector it is found that pixel p
The normal direction flow vector V at placenIt can be expressed as form：
Wherein EtRepresent the variation of gray value at corresponding pixel points p in adjacent two field pictures.Ex, Ey, EtCalculation expression
It is as follows：
Wherein I1, I2It is the gray matrix of adjacent two field pictures.Therefore, for true consecutive image sequence, each frame figure
As upper normal direction flow vector field can be calculated by formula (2) and formula (3).
After normal direction flow vector field on the 2nd frame image is calculated, normal direction stream when the present invention utilizes pure translational movement
The half-plane of vector constrains the translational motion information to estimate video camera, i.e. FOE.Utilize the base of this constraints estimation FOE
Present principles are as shown in Figure 1.
In Fig. 1, three pixel p1, p2And p3The normal direction flow vector at place is respectively Vn(p1), Vn(p2) and Vn(p3)；l
(p1), l (p2) and l (p3) it is respectively by pixel p1, p2And p3And perpendicular to the straight line of each pixel normal direction flow vector.For
Pixel p1For, by the half-plane constraint of normal direction flow vector during pure translational movement it is found that characterization camera pan information
FOE should be located at do not include normal direction flow vector Vn(p1) half-plane in；Similarly, for pixel p2For, FOE should position
In not comprising normal direction flow vector Vn(p2) half-plane in；For pixel p3For, FOE should be located at and not include normal direction stray arrow
Measure Vn(p3) half-plane in.Assuming that video camera makees pure translational movement, and scenery is static, then characterizes video camera pure translational movement letter
The positions of the FOE of breath in the picture should be able to make the normal direction flow vector in image at all pixels point meet half-plane about simultaneously
Beam.Therefore, in Fig. 1, in order to simultaneously meet three pixel p1, p2And p3The half-plane constraint at place, FOE should be located at by straight
Line l (p1), l (p2) and l (p3) in the delta-shaped region that surrounds.It can be seen that estimated by the normal direction flow vector of three pixels
The possibility distributed areas of obtained FOE are significantly less than the possibility of FOE estimated by the normal direction flow vector of a pixel
Distributed areas.Therefore, with for estimating that the pixel number of FOE gradually increases, utilizing normal direction flow vector during pure translational movement
The possibility distributed areas of FOE estimated of half-plane constraints be gradually reduced.Pixel until being used to estimate FOE
Number increases to after certain numerical value, and the precision of the FOE estimated can reach Pixel-level.
Based on above-mentioned principle, the present invention proposes the method using temporal voting strategy estimation camera pan information, that is, throws
Ticket method.This method estimation FOE is as follows：
(1) it establishes a two dimension to add up array, and it is zero to enable the initial values of all elements in array.The size of the array
It is identical with the spatial resolution of the consecutive image sequence of pretreatment.The range of its first and second dimension is FOE respectively in y-axis, x-axis
Possibility value range.
(2) for some pixel p in the 2nd frame image, the normal direction flow vector V at the pixel is utilizedt n's
Half-plane constraints can determine the possibility distributed areas of FOE on the image plane.
(3) it after the possibility distributed areas for obtaining FOE, finds the corresponding position in the region in the cumulative array of two dimension and makes
The element value of its corresponding position adds 1.
(4) step (2) and (3) is performed to other pixels in current frame image in addition to pixel p, until in image
All elements be disposed.
(5) find the maximum value of element in the cumulative array of two dimension, corresponding column and row call number be FOE in the picture
Coordinate value.
After above-mentioned five steps, you can accurately estimate the pure translational movement information of video camera, i.e. FOE.By
During Yu Yong ballot method estimations FOE, merely with the directional information of normal direction flow vector, therefore, during this method estimation FOE
It is relatively low to noise sensitivity.But in the pure translational movement of the half-plane constraints estimation video camera using normal direction flow vector
During information, need to exclude interference caused by self-movement target in photographed scene, the present invention excludes this dry using following methods
It disturbs.
First, the 2nd frame image is handled using the image partition method based on Mean-Shift algorithms, obtains its coloured silk
Color segmentation result figure.In the figure, the pixel with similar features spatial information flocks together, and forms several different companies
Logical region.Meanwhile these connected regions are endowed different RGB color values.Secondly, according to the RGB in Color Segmentation result figure
Color value assigns different label values to each region, so as to form an area label figure.Then, each label value is calculated
Corresponding region area size is to count the pixel number of each region.Finally, due to the larger region of area belongs to background
Therefore the possibility bigger of image, selects the normal direction flow vector at the corresponding pixel in these regions to estimate the pure flat of video camera
Shifting movement information, i.e. FOE.Meanwhile the corresponding label value in these regions is updated to " 0 ", wherein " 0 " represents background area.
Pass through above step, you can estimation obtains the pure translational movement information of video camera.So it is taken the photograph with the estimation of ballot method
During the pure translational movement information of camera, the normal direction flow vector at the corresponding pixel in the region is selected, and by the label in the region
Value is updated to " 0 ", that is, thinks that the region belongs to background.
2nd, the generation of initial markers field
After estimating to obtain the pure translational movement information i.e. FOE of video camera with ballot method, with reference to the normal direction stream of the 2nd frame image
Vector field, you can by the region division in its Color Segmentation result figure be two groups：Background area and foreground area (i.e. moving target
Region).In the 2nd frame image, the normal direction flow vector at all pixels for belonging to background area all meets translation half-plane about
Beam, and normal direction flow vector at the pixel of motion target area is belonged to for those, due to being influenced by displacement, make
It obtains the normal direction flow vector in the region at most of pixel and is unsatisfactory for translation half-plane constraint.The fact that utilization, you can right
Classify each region in its Color Segmentation result figure.
For any region R in its Color Segmentation result figure, first, the normal direction at each pixel in the region is examined
Flow vector whether meet translation half-plane constraint, and count wherein meet translation half-plane constraint pixel number Sup (R) and
Ungratified pixel number Vio (R).Secondly, the pixel number institute in this region for meeting translation half-plane constraint is calculated
The ratio Ratio (R) accounted for：
Finally, the attribute in the region is differentiated according to the Ratio (R) the being calculated and threshold epsilon ＞ 0 being previously set.At this
In invention, empirical value ε=0.8 is can use according to abundant experimental results.If Ratio (R) >=ε, the region belong to background, simultaneously will
The corresponding label value in the region is updated to " 0 "；Conversely, Ratio if (R) ＜ ε, which belongs to moving target, updates simultaneously
The corresponding label value in the region is " 1 ".
Above three step is repeated to each region in Color Segmentation result figure to draw whole image region
It is divided into two classes --- background and moving target, so as to generate initial markers field (the Initial Labeling of the 2nd frame image
Field)。
3rd, the generation of optimal Label Field
It the initial markers field that previous step obtains can not be accurately by the self-movement target in photographed scene from the 2nd frame figure
Divide as in and extract, therefore, the present invention also needs to first to this by some prioris included in consecutive image sequence
Beginning Label Field optimizes, to obtain the corresponding optimal Label Field (Optimal Labeling Filed) of the 2nd frame image, from
And the accurate Ground Split of self-movement target in photographed scene is extracted.In the optimal Label Field, " 0 " represents background area
Domain；" 1 " represents motion target area.
The it is proposed of markov random file (Markov Random Field, MRF) theory so that this optimization problem obtains
Obtain perfect solution.The priori included in image can well be expressed by Markov random field model.It utilizes
Markov random file theory is to be regarded as a Bayes to label (Bayesian Labeling) when being modeled to the problem
Problem acquires a maximum a-posteriori estimation (Maximum a of authentic signature field (True Labeling)
Posteriori, MAP)." 1 " in the authentic signature field represents the self-movement target in photographed scene, and " 0 " represents shooting field
Background area in scape.Since markov random file Joint Distribution and Gibbs Distribution Equivalence Theorem (i.e. Hammersley-
Clifford theorems) be suggested after, just so that solutions of the problem is converted to solution objective energy function minimum value and asks
Topic acquires one group of optimal label configuration and objective energy function is caused to obtain minimum value.And it is current relatively stream that figure, which cuts algorithm,
The method of capable solution tag combination optimization problem.Therefore, the present invention first with markov random file theory to the optimization
Problem is modeled, and obtains objective energy function；Then algorithm (Graph Cuts) is cut with figure solve the objective energy function again.
The process for being optimized under MAP-MRF frame structures to initial markers field and obtaining optimal Label Field is described more fully below.
For given tag set L={ 0,1 }, the observation field D={ d of the 2nd frame image1,d2,...,dNAnd the 2nd frame
The corresponding normal direction flow vector field of imageWherein diIt represents in the 2nd frame image at ith pixel point
RGB color value is a trivector；It represents the normal direction flow vector at ith pixel point, is a two-dimensional vector；N
Represent the pixel sum in the 2nd frame image.The target of the present invention seeks to obtain the corresponding optimal Label Field of the 2nd frame imageWherein fi *Expression is assigned to the optimal label value of ith pixel point, and value may be 0 or 1.
Under MAP-MRF frame structures, optimal Label Field f*It can be obtained by solving following formula：
Wherein U (f, V) represents objective energy function, and F represents the collection for being possible to Label Field f corresponding to the 2nd frame image
It closes.It follows that the process of the formula of solution (5) is exactly to find that objective energy function U (f, V) is made to obtain most in Label Field set F
One Label Field of small value, i.e. f*.Objective energy function U (f, V) is defined as follows：
U (f, V)=U1(f)+U2(f|V) (6)
Wherein U1(f) the priori group energy function of image is represented, the summation of all group energy in representative image；U2(f|
V it) represents likelihood energy function, represents the summation of required energy when assigning a label value to each pixel in image.
U1(f), U2Shown in (f | V) is defined as follows：
U in formula (7)1(fi,fj) represent the discontinuous cost of adjacent two pixels (i.e. i-th, j pixel) feature, i.e. base
Group's energy；CiRepresent the eight neighborhood pixel point set at ith pixel point.U in formula (8)2(fi|Vi n) it is represented to ith pixel point
Assign label value fiThe energy of Shi Suoxu.In formula, α and β is zero mean Gaussian white noise.U1(fi,fj),Determine
Justice is as follows：
Dist (d in formula (9)i,dj) in represent diAnd djEuclidean distance, τ is the distance threshold being previously set, 0 ＜ α1＜
α2；Cor in formula (10)iRepresent the coordinate of ith pixel point,Represent the normal direction flow vector at ith pixel point, μ is normal direction
The flow noise factor, δ (i) are defined as follows：
Wherein, num represents the pixel for meeting translation half-plane constraint in the eight neighborhood pixel point set of ith pixel point
Number.
In order to obtain the globally optimal solution of formula (5), the present invention solves the tag combination optimization problem using figure segmentation method.By
This defines a non-directed graph G=(V, E), and V represents the set on vertex in non-directed graph (comprising N number of node for representing image slices vegetarian refreshments
The terminal s for representing background with one and terminal a t), E for representing moving target represent that the set on side in non-directed graph (includes two
The side of type：A kind of is the side between connecting node, and weights are set as U1(fi,fj)；Another kind be connecting node and terminal it
Between side, weights are set asIt, can be by net by finding cutting for Least-cost in the non-directed graph
Node in trrellis diagram is divided into two set (S and T), wherein the node in set S is all connect with terminal s, the node in set T is all
It is connect with terminal t.
Complete above-mentioned steps, you can obtain the corresponding optimal Label Field f of the 2nd frame image*。
4th, in video self-movement target detection
Front has described several key steps in object detection method under the dynamic environment based on normal direction flow field in detail
Suddenly.The specific steps that self-movement target in algorithm extraction video is put forward using the present invention, flow chart such as Fig. 2 is explained below
It is shown.
Self-movement target in video is extracted using the method for the present invention to be as follows：
(1) the 1st, 2 frame images in video are selected, the normal direction stream of the 2nd frame image is calculated using formula (2) and formula (3)
Vector field；The 2nd frame image is split using the image partition method based on Mean-Shift algorithms simultaneously, obtains its colour
Segmentation result figure.
(2) ballot method is estimated to obtain the pure translational movement information of video camera, i.e. FOE.
(3) the corresponding initial markers field of the 2nd frame image of generation.
(4) the initial markers field is optimized, the corresponding optimal Label Field of the 2nd frame image of generation, so as to which mesh will be moved
Mark detected.
(5) using the corresponding optimal Label Field of the 2nd frame image, with reference to the Color Segmentation result figure of the 3rd frame image, generation the
The corresponding initial markers field of 3 frame images.The method for generating the initial markers field is as follows：First, according to the colour of the 3rd frame image point
The RGB color value in result figure is cut, different label values (its value is greater than 0 positive number) is assigned to each region, so as to raw
Into an area label figure.Secondly, by pixel corresponding with " 0 " in the optimal Label Field of the 2nd frame image in the area label figure
Label value at point becomes " 0 ", generates a pseudo- initial markers field.Again, each non-zeros label in pseudo- initial markers field is calculated
It is worth corresponding pixel number ratio Ratio (label shared in the same label area of former area label figurej), j=1,
2 ..., m, wherein labeljRepresent j-th of non-zeros label value, m represents the number of non-zeros label value.Finally, if Ratio
(labeljJ-th of non-zeros label value is then updated to " 1 " by) >=τ；Conversely, it is updated to " 0 ".By the above process, you can generation
The corresponding initial markers field of 3rd frame image.
(6) the initial markers field is optimized again, the corresponding optimal Label Field of the 3rd frame image of generation, so as to transport
Moving-target detected.
(7) (5) are repeated, (6) step, you can complete the detection of moving target in subsequent frame image.
Effectiveness of the invention and accuracy are carried out by the consecutive image sequence of emulation experiment and actual photographed
Verification, achieves good estimated result.Experimental result is shown in Fig. 3, Fig. 3 is the movement of the 57th frame image in true picture sequence
Object detection results, wherein (a) is original image, (b) is the Color Segmentation result figure based on Mean-Shift, and (c) is initial
Label Field, (d) are optimal Label Field, and (e) is true templates of moving objects, and (f) is DECOLOR algorithm testing results.This hair
Bright sharpest edges are that algorithm is relatively low to noise sensitivity, and the image texture and characteristics of image to photographed scene do not require,
Therefore scene is applicable in than wide.From the experimental results, the present invention can be effectively realized carries out target under dynamic environment
Detection.
Above example is provided just for the sake of the description purpose of the present invention, and is not intended to limit the scope of the present invention.This
The range of invention is defined by the following claims.It the various equivalent replacements that do not depart from spirit and principles of the present invention and make and repaiies
Change, should all cover within the scope of the present invention.
Claims (6)
1. object detection method under a kind of dynamic environment based on normal direction stream information, it is characterised in that include the following steps：
Step 1：The 1st, 2 frame images in video are selected, the normal direction flow vector field of the 2nd frame image are calculated, while to the 2nd
Frame image is split, and obtains the Color Segmentation result figure of the 2nd frame image；
Step 2：Using the normal direction flow vector field and Color Segmentation figure being calculated in step 1, estimate to obtain using ballot method
The pure translational movement information FOE of video camera；
Step 3：It, will be colored with reference to the normal direction flow vector field of the 2nd frame image after the pure translational movement information FOE of video camera
Segmentation result figure carries out region division, is divided into background area and foreground area, foreground area, that is, motion target area, then to two
Different zones carry out the judgement of translation half-plane constraint, so as to generate the corresponding initial markers field of the 2nd frame image；
Step 4：The corresponding initial markers field of the 2nd frame image is optimized, the corresponding optimal mark of the 2nd frame image of generation
Remember field, then the problem of optimal Label Field is modeled using markov random file theory, obtain objective energy function；Again
Algorithm is cut with figure and solves the objective energy function, and final realize detected moving target from the 2nd frame image；
Step 5：3rd frame image is split, obtains Color Segmentation result figure, utilizes the corresponding optimal label of the 2nd frame image
, with reference to the Color Segmentation result figure of the 3rd frame image, the corresponding initial markers field of the 3rd frame image of generation；
Step 6：The corresponding initial markers field of the 3rd frame image is optimized again, the 3rd frame image of generation is corresponding most
Excellent Label Field, so as to which moving target detected from the 3rd frame image；
Step 7：It repeats Step 5: six, detects the target in other frame images, until by the movement in whole frame image
Target detection comes out.
2. object detection method under the dynamic environment according to claim 1 based on normal direction stream information, it is characterised in that：Institute
It states in step 1, the Color Segmentation result figure obtained using the image partition method based on Mean-Shift algorithms.
3. object detection method under the dynamic environment according to claim 1 based on normal direction stream information, it is characterised in that：Institute
It states in step 3, the 2nd corresponding initial markers field of frame image of generation is implemented as follows：
(1) it for any region R in the Color Segmentation result figure of the 2nd frame image, examines and counts the region R satisfactions translation
The pixel number Sup (R) of half-plane constraint and ungratified pixel number Vio (R), and calculate satisfaction translation half-plane about
The pixel number of beam shared ratio Ratio (R) in this region：
(2) differentiate the attribute of the region R according to the Ratio (R) the being calculated and threshold epsilon ＞ 0 being previously set, it is of the invention according to
Empirical value ε=0.8 is taken according to abundant experimental results, if Ratio (R) >=ε, the region R belongs to background, while by the region pair
Label value is answered to be updated to " 0 "；Otherwise, the region R belongs to moving target, while updates the region corresponding label value for " 1 "；
It to be two classes by whole image region division that above-mentioned steps are repeated to each region in Color Segmentation result figure
Background area and foreground area, so as to generate the initial markers field of the 2nd frame image (Initial Labeling Field).
4. object detection method under the dynamic environment according to claim 3 based on normal direction stream information, it is characterised in that：Institute
State ε=0.8.
5. object detection method under the dynamic environment according to claim 1 based on normal direction stream information, it is characterised in that：Institute
It states in step 4, the corresponding optimal Label Field of the 2nd frame image of generation is specific as follows：
(1) for given tag set L={ 0,1 }, the observation field D={ d of the 2nd frame image1,d2,...,dNAnd the 2nd frame
The corresponding normal direction flow vector field of imageWherein diIt represents in the 2nd frame image at ith pixel point
RGB color value is a trivector；Vi nIt represents the normal direction flow vector at ith pixel point, is a two-dimensional vector；N
Represent the pixel sum in the 2nd frame image, target is to obtain the corresponding optimal Label Field of the 2nd frame image
Wherein fi *Expression is assigned to the optimal label value of ith pixel point, and value may be 0 or 1；
(2) under MAP-MRF frame structures, optimal Label Field f*It is obtained by solving following formula：
Wherein U (f, V) represents objective energy function, and F represents the set for being possible to Label Field f corresponding to the 2nd frame image；It asks
The process of solution formula (2) is that the Label Field that objective energy function U (f, V) is made to obtain minimum value is found in Label Field set F,
That is f*, objective energy function U (f, V) is defined as follows：
U (f, V)=U1(f)+U2(f|V) (3)
Wherein U1(f) the priori group energy function of image is represented, the summation of all group energy in representative image；U2(f | V) table
Show likelihood energy function, represent the summation of required energy when assigning a label value to each pixel in image；U1
(f), U2Shown in (f | V) is defined as follows：
U1(fi,fj) represent the discontinuous cost of adjacent two pixels (i.e. i-th, j pixel) feature, i.e. group energy；CiTable
Show the eight neighborhood pixel point set at ith pixel point, U2(fi|Vi n) it is represented to ith pixel point imparting label value fiShi Suoxu
Energy；α and β is zero mean Gaussian white noise；U1(fi,fj), U2(fi|Vi n) be defined as follows shown in：
dist(di,dj) in represent diAnd djEuclidean distance, τ is the distance threshold being previously set, 0 ＜ α1＜ α2；CoriRepresent the
The coordinate of i pixel, Vi nRepresent the normal direction flow vector at ith pixel point, δ (i) is defined as follows：
Wherein, num represents the pixel for meeting translation half-plane constraint in the eight neighborhood pixel point set of ith pixel point
Number；
(3) in order to obtain globally optimal solution, f is solved using figure segmentation method*, a non-directed graph G=(V, E) is defined, V represents undirected
The set on vertex in figure represents background terminal s comprising N number of node for representing image slices vegetarian refreshments and one and one represents movement mesh
Mark terminal t；E represents the set on side in non-directed graph, includes two kinds of side：A kind of is the side between connecting node, weights
It is set as U1(fi,fj)；Another kind is the side between connecting node and terminal, and the weights on side are set as U2(fi|Vi n)；By undirected
Cutting for Least-cost is found in figure, makes this minimum cutting that the node in grid chart can be divided into two set S and T,
Node in wherein set S is all connect with background terminal s, and the node in set T is all connect with target terminal t；Complete above-mentioned step
Suddenly, that is, the corresponding optimal Label Field of the 2nd frame image is obtained, so as to which moving target detected from the 2nd frame image.
6. object detection method under the dynamic environment according to claim 1 based on normal direction stream information, it is characterised in that：Institute
It states in step 5, the corresponding initial markers field of the 3rd frame image of generation is specific as follows：
(1) it according to the RGB color value in the Color Segmentation result figure of the 3rd frame image, is assigned to the region of different colours different
Label value, so as to generate an area label figure；
(2) by the area label figure in the optimal Label Field of the 2nd frame image " 0 ", i.e., most having the background picture in Label Field
Plain label is denoted as " 0 ", and object pixel label is denoted as " 1 ", and the label value at corresponding pixel becomes " 0 ", at the beginning of generating a puppet
Beginning Label Field；
(3) it calculates each non-zeros label in pseudo- initial markers field and is worth same mark of the corresponding pixel number in former area label figure
Sign ratio Ratio (label shared in regionj), j=1, wherein 2 ..., m, labeljRepresent j-th of non-zeros label value, m tables
Show the number of non-zeros label value；If Ratio (labelj) >=τ, τ are threshold value, then are updated to j-th of non-zeros label value " 1 "；Instead
It, is updated to " 0 ", that is, generates the corresponding initial markers field of the 3rd frame image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711421030.XA CN108171725A (en) | 2017-12-25 | 2017-12-25 | Object detection method under a kind of dynamic environment based on normal direction stream information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711421030.XA CN108171725A (en) | 2017-12-25 | 2017-12-25 | Object detection method under a kind of dynamic environment based on normal direction stream information |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108171725A true CN108171725A (en) | 2018-06-15 |
Family
ID=62520276
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711421030.XA Pending CN108171725A (en) | 2017-12-25 | 2017-12-25 | Object detection method under a kind of dynamic environment based on normal direction stream information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108171725A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104616298A (en) * | 2015-01-30 | 2015-05-13 | 浙江工业大学之江学院 | Method for detecting moving target of ink-jet printing fabric based on mixed-state Gauss MRF (Markov Random Field) model |
| US20160073062A1 (en) * | 2014-09-05 | 2016-03-10 | Toyota Jidosha Kabushiki Kaisha | Approaching object detection apparatus for a vehicle and approaching object detection method for the same |
| CN106780559A (en) * | 2016-12-28 | 2017-05-31 | 中国科学院长春光学精密机械与物理研究所 | A kind of moving target detecting method and device |
-
2017
- 2017-12-25 CN CN201711421030.XA patent/CN108171725A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160073062A1 (en) * | 2014-09-05 | 2016-03-10 | Toyota Jidosha Kabushiki Kaisha | Approaching object detection apparatus for a vehicle and approaching object detection method for the same |
| CN104616298A (en) * | 2015-01-30 | 2015-05-13 | 浙江工业大学之江学院 | Method for detecting moving target of ink-jet printing fabric based on mixed-state Gauss MRF (Markov Random Field) model |
| CN106780559A (en) * | 2016-12-28 | 2017-05-31 | 中国科学院长春光学精密机械与物理研究所 | A kind of moving target detecting method and device |
Non-Patent Citations (2)
| Title |
|---|
| DING YUAN, YALONG YU: "Moving Object Detection under Moving Camera by Using Normal Flows", 《PROCEEDINGS OF THE 2014 IEEE INTERNATIONAL CONFERENCE ON ROBOTICS AND BIOMIMETICS》 * |
| 丁亚龙: "基于法向流场的动态场景中运动目标检测", 《道客巴巴HTTP://WWW.DOC88.COM/P-6317753742744.HTML》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Herbst et al. | Toward object discovery and modeling via 3-d scene comparison | |
| CN108648233A (en) | A kind of target identification based on deep learning and crawl localization method | |
| CN107564022B (en) | Saliency detection method based on Bayesian Fusion | |
| CN111563442A (en) | Slam method and system for fusing point cloud and camera image data based on laser radar | |
| CN104156693B (en) | A kind of action identification method based on the fusion of multi-modal sequence | |
| CN109559320A (en) | Realize that vision SLAM semanteme builds the method and system of figure function based on empty convolution deep neural network | |
| CN109447082B (en) | Scene moving object segmentation method, system, storage medium and equipment | |
| CN109919059B (en) | Salient object detection method based on deep network layering and multi-task training | |
| CN106709938B (en) | Based on the multi-target tracking method for improving TLD | |
| CN106204594A (en) | A kind of direction detection method of dispersivity moving object based on video image | |
| CN104680546A (en) | Image salient object detection method | |
| CN106991686A (en) | A kind of level set contour tracing method based on super-pixel optical flow field | |
| CN112446882A (en) | Robust visual SLAM method based on deep learning in dynamic scene | |
| CN108804992A (en) | A kind of Demographics' method based on deep learning | |
| CN108846416A (en) | The extraction process method and system of specific image | |
| CN105335949A (en) | Video image rain removal method and system | |
| CN107967449B (en) | A kind of multispectral image unknown object recognition methods based on broad sense evidence theory | |
| Dou et al. | Moving object detection based on improved VIBE and graph cut optimization | |
| Matas | Colour-based object recognition | |
| CN104077603B (en) | Outdoor scene monocular vision space recognition method in terrestrial gravity field environment | |
| CN109658441A (en) | Foreground detection method and device based on depth information | |
| CN116883897A (en) | Low-resolution target identification method | |
| Paruchuri et al. | Spatially adaptive illumination modeling for background subtraction | |
| CN104077611A (en) | Indoor scene monocular vision space recognition method in terrestrial gravity field environment | |
| Gardziński et al. | Crowd density estimation based on voxel model in multi-view surveillance systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180615 |