CN112946625B - B-spline shape-based multi-extended target track tracking and classifying method - Google Patents
B-spline shape-based multi-extended target track tracking and classifying method Download PDFInfo
- Publication number
- CN112946625B CN112946625B CN202110166401.4A CN202110166401A CN112946625B CN 112946625 B CN112946625 B CN 112946625B CN 202110166401 A CN202110166401 A CN 202110166401A CN 112946625 B CN112946625 B CN 112946625B
- Authority
- CN
- China
- Prior art keywords
- shape
- target
- spline
- information
- measurement
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 80
- 238000005259 measurement Methods 0.000 claims abstract description 99
- 238000000605 extraction Methods 0.000 claims abstract description 7
- 230000008569 process Effects 0.000 claims description 21
- 238000001914 filtration Methods 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 10
- 238000010586 diagram Methods 0.000 claims description 10
- 239000013598 vector Substances 0.000 claims description 10
- 238000001514 detection method Methods 0.000 claims description 9
- 230000004927 fusion Effects 0.000 claims description 7
- 238000000638 solvent extraction Methods 0.000 claims description 7
- 230000011218 segmentation Effects 0.000 claims description 5
- 238000007499 fusion processing Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 238000013138 pruning Methods 0.000 claims description 4
- 230000009471 action Effects 0.000 claims description 2
- 238000003672 processing method Methods 0.000 claims description 2
- 230000007704 transition Effects 0.000 claims description 2
- 230000010365 information processing Effects 0.000 abstract description 2
- 230000001788 irregular Effects 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 description 23
- 230000006870 function Effects 0.000 description 19
- 101001086191 Borrelia burgdorferi Outer surface protein A Proteins 0.000 description 9
- 230000000694 effects Effects 0.000 description 6
- 238000012544 monitoring process Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 3
- 238000011160 research Methods 0.000 description 2
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000009707 neogenesis Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000004451 qualitative analysis Methods 0.000 description 1
- 238000004445 quantitative analysis Methods 0.000 description 1
- 229910052704 radon Inorganic materials 0.000 description 1
- SYUHGPGVQRZVTB-UHFFFAOYSA-N radon atom Chemical compound [Rn] SYUHGPGVQRZVTB-UHFFFAOYSA-N 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 230000004083 survival effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/66—Radar-tracking systems; Analogous systems
- G01S13/72—Radar-tracking systems; Analogous systems for two-dimensional tracking, e.g. combination of angle and range tracking, track-while-scan radar
- G01S13/723—Radar-tracking systems; Analogous systems for two-dimensional tracking, e.g. combination of angle and range tracking, track-while-scan radar by using numerical data
- G01S13/726—Multiple target tracking
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/02—Systems using reflection of radio waves, e.g. primary radar systems; Analogous systems
- G01S13/06—Systems determining position data of a target
- G01S13/42—Simultaneous measurement of distance and other co-ordinates
Landscapes
- Engineering & Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- Remote Sensing (AREA)
- Computer Networks & Wireless Communication (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a B-spline shape-based multi-extended target track tracking and classifying method, which relates to the technical field of information processing and comprises the following steps: firstly, the irregular shape estimated by the B spline is used in a KDE-SSP method, the adjacent target measurement set is divided for the second time, and the center position of the candidate shape is searched by using a kernel density estimation method, so that the efficiency of the algorithm is improved; and then, target shape category information is extracted to assist the updating and extraction of the target state, so that the problems of missing tracking, wrong tracking and the like of updating of an adjacent target measurement set are solved, and meanwhile, the track management and the target classification of the extended target are effectively realized according to the extracted target motion state, the target shape information and the target track.
Description
Technical Field
The invention relates to the technical field of information processing, in particular to a B-spline-shape-based multi-extension target track tracking and classifying method.
Background
The research on the multi-target tracking (MTT) technology has been a key point in the target tracking problem, and has also been a difficult point, and has been widely applied to the fields of civilian use, military use, and the like, such as tracking systems for video monitoring and the like, monitoring systems for early warning of enemies, missile-borne systems, sea, land, air and multi-dimensional integrated cooperative systems, and the like.
The MTT mainly utilizes observation information obtained by a sensor to estimate the state of a target, but because many uncertain interference factors exist in a complex real scene, the observation information is inaccurate, and a great challenge is brought to tracking. Early multi-target tracking mostly adopts the idea of data association, such as Multiple Hypothetic Tracking (MHT), Joint data association (JPDA), etc., but these methods have too high time complexity when the number is increased, and it is difficult to effectively process multi-target tracking with a changed number. Aiming at the problem, in Mahler and 2003, a Random Finite Set (RFS) multi-target tracking theory is proposed in a document Multi target Bayes filtering sight-order multi target movements, so that the uncertain and variable multi-target tracking can be effectively processed, and a new thought is provided for the multi-target tracking.
With the increasing resolution of modern radar and other novel detection devices, the echo signals of a target may be distributed in different range resolution units, the detection field of the target is no longer equivalent to one point, the target may produce multiple measurements, and a target with such characteristics is called an Extended target. In the conventional point target tracking problem, the assumption that one target corresponds to one measurement is no longer true, but the problem that a plurality of measurements correspond to the same target is solved. In 2009, Mahler and the like further popularize the RFS theory to the multi-extension target tracking field, provide a multi-extension target measurement updating frame, provide a new idea for multi-extension target tracking, become the leading edge and hot spot problems of the research of the extension target tracking field, and particularly have great challenges for irregular-shape multi-extension target tracking under the uncertain observation complex environment. Random matrix technology is introduced into document A PHD Filter for Tracking Multiple Extended Targets Using Random matrix, and a Gaussian inverse Weishate (GIW-PHD) filtering method is proposed to realize the shape estimation of the Extended target, but only the shape of the target can be estimated to be an ellipse; subsequently, a Random Hypersurface technology is introduced in the document Extended Object and Group Tracking with orthogonal Random Hypersurface Models to further realize star-shaped multi-Extended target Tracking. However, the algorithms do not fully consider the shape information of the extended targets, track tracking and target classification of the extended targets are realized, especially when multiple extended targets are adjacent, the tracking precision is seriously reduced, and aiming at the problem, the invention provides a method for driving the track tracking and the classification of the multiple extended targets based on the B-spline shape.
Disclosure of Invention
Aiming at the problems and the technical requirements, the invention provides a method for tracking and classifying multi-extended target tracks based on B-spline shape drive, and the technical scheme of the invention is as follows:
a B-spline-shape-based multi-extension target track tracking and classifying method comprises the following steps:
establishing a B-Spline-GM-PHD filter, initializing a target state at the moment k:
wherein x iskRepresents a state of motion, BkRepresenting B-spline shape information, pk,i、θk,iRespectively represents the polar diameter and polar angle information at the ith angle at the moment k, LkA track label representing the gaussian component at time k,and a track label representing the jth Gaussian component at time k.
When k is larger than or equal to 1, performing primary division on the measurement set by a distance division method;
if the divided measurement set has an adjacent target, performing secondary division on the divided measurement set, and if the divided measurement set does not have an adjacent target, performing multi-hypothesis filtering on the target state by adopting a B-Spline-GM-PHD filter;
performing multi-hypothesis filtering on the target state by adopting a B-Spline-GM-PHD filter to update the target state and obtain a shape likelihood function after weight updating;
fusing target components meeting a distance fusion threshold, pruning Gaussian components with undersized weights after fusion, taking the Gaussian components with high weights as extraction targets in the fusion process, and extracting the targets comprising target motion states, target shape information and target tracks;
the likelihood is obtained by the shape likelihood function for the extracted target shape information and the preset target shape with the category information, and the target shape category information is obtained;
and if the observation information arrives at the next moment, re-executing the step of dividing the measurement set once by using the distance division method, otherwise, ending the target track tracking process.
The beneficial technical effects of the invention are as follows:
the application provides a method for tracking and classifying multi-extension target tracks based on B-spline shape driving, aiming at the phenomenon of tracking missing and wrong tracking when the shape information of an adjacent target is not fully considered in the multi-extension target tracking based on ET-GM-PHD. The method adopts a B spline fitting method to realize the estimation of any irregular shape of the extended target, fully excavates the shape information of the extended target, adopts a kernel density shape partitioning (KDE-SSP) method to carry out secondary partitioning on a measurement set of the adjacent extended target, extracts target shape category information to assist the update and extraction of the target state, effectively improves the tracking performance of the adjacent extended targets, and effectively realizes the track management and the target classification of the extended target according to the extracted target motion state, the target shape and the target track.
Drawings
FIG. 1 is a flow chart of multi-extended target tracking based on B-Spline-GM-PHD provided by the present application.
FIG. 2 is a diagram of a metrology set W in which two extended targets are in close proximity.
FIG. 3 is a three-dimensional nuclear density estimate plot of the metrology in the immediate vicinity of the extended target metrology set W.
FIG. 4 is a schematic diagram of the measure point determination in the KDE-SSP algorithm.
Fig. 5 is a polar angle and radius map after dimension expansion in the shape matching algorithm.
Fig. 6 is a polar angle and diameter map after processing by a shape matching algorithm.
FIG. 7 is a graph of the effect of the KDE-SSP algorithm in partitioning the set of proximate target metrics.
FIG. 8 is a diagram of the effect of the B-Spline-GM-PHD algorithm without shape information update.
FIG. 9 is a diagram illustrating the effect of the updated B-Spline-GM-PHD algorithm provided by the present application.
FIG. 10 is a graph of track following results without shape information update.
FIG. 11 is a graph of the results of track following two targets.
FIG. 12 is a graph illustrating the average OSPA distance and the weights and sums of two targets.
Fig. 13 is a diagram of the result of shape classification of two objects.
FIG. 14 is a graph of two target average shape classification accuracy.
FIG. 15 is a graph showing the classification results and the average classification accuracy for three targets.
FIG. 16 is a graph of the results of track following three targets.
FIG. 17 is a graph illustrating three target average OSPA distances.
FIG. 18 is a graph of three target average weight sums.
Detailed Description
The following description of the embodiments of the present invention will be made with reference to the accompanying drawings.
The application discloses a B-spline-shape-based multi-extension target track tracking and classifying method, a flow chart of which is shown in figure 1, and the method comprises the following steps:
step 1: establishing a B-Spline-GM-PHD filter, and initializing a target state at the moment k:
wherein x iskRepresents a state of motion, BkRepresenting B-spline shape information, pk,i、θk,iRespectively represents the polar diameter and polar angle information at the ith angle at the moment k, LkA tag representing the gaussian component track at time k,and (3) associating the same labels at different moments by using the track label representing the jth Gaussian component at the moment k, so as to realize association between tracks.
Step 2: and when k is more than or equal to 1, dividing the measurement set once by a distance division method.
And step 3: and (4) judging whether a plurality of measurement neighbors exist in the measurement set after the primary division, if so, entering the step 4, and if not, entering the step 5.
And 4, step 4: if the divided measurement set has an adjacent target, performing secondary division on the divided measurement set, including:
step 41: initializing a candidate shape parameter set C when k is 1:
wherein,candidate shape parameters for fitting a B-spline curve,respectively representing polar diameter and polar angle information at the ith angle; initializing polar anglesIs composed ofn1Is a predetermined dimension, and n1< 360, and the polar diameter information corresponding to the initialized polar angle isAlternatively, suppose n1When 20, the polar angle is initialized to
When k is more than or equal to 2, selecting m targets with weight omega more than 0.5 in the posterior target state, and putting the shape information of the m posterior target states into the candidate shape parameter setIn the method, the targets with the weight omega more than 0.5 are selected to play a role in removing clutter.
Step 42: determining the central point position of each candidate shape in the candidate shape parameter set C by adopting a Kernel Density Estimation (KDE), and establishing a candidate central position set L:
wherein the coordinates corresponding to the density peak points of the candidate shape to which the B-spline curve is fitted are represented by (ω)q (x),ωq (y)) And n represents the number of density peak points.
Assume that the metrology set W contains two closely adjacent targets, representing metrology sets of multiple targets of the same metrology set. Since the candidate shape in extended target tracking is the result of fitting with a B-spline curve, all measured density peak points within the shape may be approximated as the candidate shape center point.
Specifically, the shape in the whole filtering process is fitted by using a B-spline, and the method for fitting the shape by using a B-spline curve comprises the following steps:
b-spline curve is composed of B-spline basis function Ni,k(u) and control vertex PiCollectively determined, the k-th order B-spline function is represented as:
wherein, Ni,k(u) is a k-1 degree B-spline basis function whose recurrence formula is:
wherein U is { U ═1,u2,…,uk,…,un+kIs a set of node vectors for the B-spline.
When k is 3, the B-spline basis function is expressed as:
the process of obtaining the control vertex set in the B-spline function by using the position information of the measurement points, in which the measurement set of the extended target includes a plurality of measurement points, is as follows:
suppose thatIn order to expand the measurement set of the target, the mean position of the measurement set is calculated, the mean position is used as the origin of coordinates to establish a coordinate system, the coordinate system with the angle of 2 pi is divided into n equal parts to obtain n control vertex positions, and the polar coordinates are used for representing the control vertexes Lambdai={ρi,θi},ρi、θiRespectively, the pole diameter and the pole angle, and the pole diameter is expressed as:
|Zi|={zk|d(zk,βi)<λ1,k(zk,αi)=1}
βi=[-tan(θi),1]
wherein, | - | represents the number of elements, d (z)k,βi) Indicates the measurement point zkTo betaiDistance of line, λ1Distance threshold for constraint width, κ (z)k,αi) As a constraint, βiIs polar angle thetaiThe calculation formula of the determined vector, d (-) is as follows:
αi=[1,tan(θi)]
wherein, | | · | | represents an absolute value, αiRepresentation and vector betaiThe vector perpendicular to the straight line.
And through the calculation, solving a control vertex set Lambda under polar coordinates, converting the control vertex set Lambda into a corresponding set P under a rectangular coordinate system, and generating a smooth curve related to the shape of a measurement set under the combined action of the control vertex set Lambda and a B spline basis function.
FIG. 2 shows two metrology sets W of two closely spaced targets, which include all of the measurement points of two extended targets, one of which is a "Y-shaped" target and the other of which is a "cross-shaped" target. By using a Gaussian kernel density function and selecting a proper bandwidth, a three-dimensional density estimation graph of two-dimensional discrete point coordinates in the measurement set W shown in FIG. 3 can be obtained. In the density map, the coordinates corresponding to each density peak point can be easily found
Step 43: the candidate shape parameter set C and the candidate center position set L are combined intoThe method comprises m × n combinations, and the Shape Selection Partitioning (SSP) is adopted to partition and combine the m × n combinations to obtain an in-Shape measurement subset and an out-Shape measurement subset, wherein the in-Shape measurement subset and the out-Shape measurement subset are respectively:
wherein,a subset of the measurements within the shape is represented,representing a subset of extratopographic measurements, zkRepresents the measurement points in the measurement set W,indicates the measurement point zkTo the central point position omegaqThe distance of (c). As shown in FIG. 4, phi denotesThe point of the corresponding point is displayed on the display,representing the angle between the connecting line of the central point position and the h-th point on the shape and the x-axis,an angle between a connecting line representing the position of the center point and the measuring point and the x-axis is set, and when the absolute value of the difference between the two is minimum, a point on the shape at the angle is selected as psi,representing psi to center point position omegaqOf the distance of (c).
wherein, bΨVector, p, representing the location of the center point to Ψk|k-1Coordinate information representing control vertices generated from the intra-shape metrology subset, Ni,3(ψ) represents a B-spline basis function.
Each combination may be in the region of the metrology set W, exactly ωqAs the position of the center point,for the fitted shape of the parameterAnd (4) counting. In addition, in the judgment of the measuring points, due to the influence of clutter or sensor instability, the measuring points closer to the edge are likely to be misjudged as out-of-shape points, and the shape parameters for segmentation are expanded by 1.3 times in the calculation process, wherein the times are empirical times obtained through multiple experiments. In fig. 4, the inner circle curve represents the unexpanded parting line, the outer circle curve represents the parting line expanded by 1.3 times, and when the shape parameter is expanded by 1.3 times, the shape edge measuring point can be well divided into the shape without affecting the division of the adjacent target measuring point.
And step 44: respectively matching the shape internal measurement subset and the shape external measurement subset to obtain the likelihood, and obtaining the condition of the maximum weight product of the shape internal measurement subset and the shape external measurement subset, wherein the condition comprises the following steps:
for intra-shape measurement subsetsThe likelihood is calculated using the shape used when the intra-shape metrology subset fits the shape and the segmentation combination, and the process of calculating the likelihood is a process of updating the shape likelihood function with a B-spline-GM-PHD filter.
For out-of-shape measurement subsetsSolving the weights by using the same updated shape likelihood function process, traversing all candidate shapes to respectively obtain likelihood with a measuring set fitting shape, and selecting the maximum value of the weights as an external measuring subsetAnd finding the condition that the likelihood product of the shape internal measuring subset and the shape external measuring subset is maximum as the final segmentation result.
And 5: and performing multi-hypothesis filtering on the target state by adopting a B-Spline-GM-PHD filter, updating the target state, and acquiring the shape likelihood function after the weight is updated.
Step 51: sequentially carrying out target prediction and target update of a B-Spline-GM-PHD filter on the target state, and specifically comprising the following steps:
the prediction step of the B-Spline-GM-PHD filter is represented as:
Lk|k-1=Lk-1∪Lβ
wherein,representing the probability hypothesis density of the predicted target, Jk|k-1Represents the number of predicted gaussian components at the moment k,respectively representing the weight, mean and covariance matrix of the jth Gaussian component predicted at the time k, N (-) is Gaussian distribution, Lk|k-1And the label set representing the prediction at the time k comprises the label sets of the original target and the new target at the time k-1, and the predicted polar diameter and polar angle information are consistent with the state of the posterior target at the time k-1.
wherein, Fk|k-1Represents a state transition matrix, IdRepresenting a d-dimensional identity matrix, QkRepresenting the process noise covariance matrix.
The update steps of the B-Spline-GM-PHD filter are represented as:
wherein, the probability hypothesis density of the missed detection target is expressed as:
Lk|k=Lk|k-1
wherein, γjA desired value representing the number of target production measurements,representing a target detection probability;
the probability hypothesis density for target detection at time k is given by:
where σ is expressed as a preset forgetting factor, ρW,iFor the polar diameter information at the ith polar angle of the shape to which the current measurement set is fitted,the sum of the measurement set numbers in all the division methods is represented, and each innovation parameter is represented as:
step 52: the weight update formula considering the target shape information is:
wherein,the position likelihood of the current measurement set is shown, and it should be noted that in the above formulaIn shorthand form, the position likelihood is solved using a gaussian distribution,show the shape likeThe calculation process of the shape matching likelihood of the function, namely the B spline, is as follows:
wherein HW、RWDenotes an innovation parameter, r'k|k-1,iRepresents the post-treatment predicted pole diameter, r'W,iDenotes the measured collector diameter after processing, and λ denotes the shape similarity index.
There are two cases for the shape likelihood function, when r'k|k-1,i|=|r′W,iIf not, the shape likelihood is calculated by the second formula.
The processing method of the pole diameter comprises the following steps:
known predicted path informationPolar diameter information corresponding to the current measurement setTwo polar diameter information are divided into n1The dimension expansion is 360 dimensions, resulting in:
as shown in FIG. 5, r can be obtained in a two-dimensional rectangular coordinate systemk|k-1And rWPolar angle and polar diameter diagram (1). FIG. 6 shows r 'after polar angle and polar diameter processing after dimensional expansion'k|k-1,iAnd r'W,iCorresponding poleAngle pole diameter diagram. From predicted polar diameter information r after dimension expansionk|k-1And measuring collector radius information r after dimension expansionWThe peak values of the wave crest and the wave trough obtained from the polar angle polar diameter diagram are respectively Pk|k-1And PWThe corresponding peak numbers are respectivelyAnd
by comparing the peak numbers of the two, the predicted polar diameter information after dimension expansion and the measured polar diameter information after dimension expansion are reordered respectively, and then the polar angle information is correspondingly changed, specifically comprising:
Pk|k-1(1,1:2)≈PW(1,u0:v)
v=mod(u0+1,Num),or v=Num
r′k|k-1=[rk|k-1(lk|k-1(1)),…,rk|k-1(end),rk|k-1(1),…,rk|k-1(lk|k-1(1)-1)]
r′W=[rW(lW(u0)),…,rW(end),rW(1),…,rW(lW(u0)-1)]
λ=1
Pk|k-1(1)≈PW(u0)
r′k|k-1=[rk|k-1(lk|k-1(1)),…,rk|k-1(end),rk|k-1(1),…,rk|k-1(lk|k-1(1)-1)]
r′W=[rW(lW(u0)),…,rW(end),rW(1),…,rW(lW(u0)-1)]
Pk|k-1(1)≠PW(u0)
r′k|k-1=rk|k-1
r′W=rW
r′k|k-1=rk|k-1
r′w=rw
wherein "≈" indicates that the peak size of the predicted shape is equal to the peak size of the metrology set shape within the threshold range.
Therefore, the shape likelihood function of the measurement set is updated, and the updated shape likelihood function formula is used for updating the likelihood of a target state, target shape classification, track tracking and the like, so that the filtering result is more accurate.
Step 6: setting a distance fusion threshold U, fusing target components meeting the distance fusion threshold U, pruning Gaussian components with undersized weights after fusion, taking the Gaussian components with high weights as extraction targets in the fusion process, and extracting the targets comprising target motion states, target shape information and target tracks.
If the number t of the fused Gaussian components is larger than the preset maximum threshold J of the Gaussian componentsmaxTaking J out of the fused Gaussian componentsmaxAnd (4) completing pruning operation by the Gaussian component with the maximum weight value.
And in the fusion process of corresponding Gaussian components of the track label, selecting the track label with the maximum weight value of the Gaussian components as the fused track label.
And 7: the likelihood of the extracted target shape information and a preset target shape with category information is solved through a shape likelihood function, and the target shape category information is obtained, and the method comprises the following steps:
to achieve classification of the target shapes, a set of preset target shapes with class information is set asThe set including a predetermined "Y-shape'and' cross shape"class information, i.eWherein
Using all the extracted object shape informationSet Bk|kEach shape of (2) is respectively integrated with a preset target shapeThe likelihood is obtained for each shape, and the formula is obtained by using the shape likelihood function of the B-Spline-GM-PHD filterSolving for each extracted shapeAll have the shapes with the maximum likelihood of matchingWhen the likelihood exceeds a category threshold, the extracted s-th shape is classified into the c-th class, and when the likelihood is smaller than the category threshold, the extracted s-th shape is classified into other classes, namely the s-th shape cannot be accurately classified into a preset target shape at the first moments of filtering, so that the s-th shape is classified into other classes, and the s-th shape can be matched with the corresponding preset target shape after the likelihood exceeds the category threshold.
And 8: and if the observation information arrives at the next moment, re-executing the step of dividing the measurement set once by using the distance division method, otherwise, ending the target track tracking process.
In order to verify the effect of the B-Spline-GM-PHD filtering algorithm and the KDE-SSP method adopted by the application, the design experiment is as follows:
1. implementation conditions and parameters
In a two-dimensional plane monitoring area [ -1000,1000] × [ -1000,1000], the number of targets in the monitoring area is unknown and varies with time. Setting two targets to have adjacent cross motion at the same time, and carrying out experimental analysis on the measurement set of filtering at the current time. In addition, the OSPA distance of the target, the weight of the target, and the shape classification accuracy were recorded in 100 experiments.
2. Software, hardware and related parameter setting in experimental process
The method is carried out on a machine with a processor of AMD Ryzen 74700U with Radon Graphics, 4.1GHz and 8 cores, a memory of 16GB and a display card of AMD Radon (TM) Graphics, and is written by Matlab R2020a software.
The target states set in the experiment were:
ζk={xk,Bk,Lk}
the parameters in the motion and metrology models are as follows:
the experiment makes uniform linear motion under the sampling period of T-1 s, wherein sigma isa=2m/s2、σg3 m. Probability of detection pD0.99, survival probability psEach time step, the number of generated target measurements is poisson's ratio γ 10.
The new target intensity is expressed as:
Db=0.1N(ζk,mb,Pb)
3. qualitative analysis of the results
The specific experiment mainly aims at the division of measurement sets at the time of the close proximity of multiple extended targets, the influence of a likelihood updating method on position, flight path and shape classification is evaluated, the influence on the accuracy of an experiment result when a target is newly generated is evaluated, and the experiment result is as follows:
experiment one: proximity extension target measurement set partitioning
In the process of movement, when two targets are close to each other, the distance between the measurement sets of the two extended targets detected by the sensor is short, and the measurement sets are easily wrongly divided into one measurement set by using a traditional measurement set dividing method. Therefore, the KDE-SSP algorithm is adopted for measurement set division of the adjacent targets.
In the uniform linear motion at the time 100, the targets move closely at the three times 47, 48 and 49. As shown in fig. 7, are the results of partitioning the measurement set using the KDE-SSP algorithm for two closely adjacent targets at time 48. It can be seen that the present invention can better divide two closely adjacent target metrology sets.
Experiment two: two target position tracking
As shown in fig. 8, at time 48, in the measurement update of the target close to the ET-GM-PHD, the target update is performed by using the position information of the measurement set that is subject to the gaussian distribution, and the phenomenon of missing and missing the target due to the erroneous update of the measurement set occurs, and at time 48 shown in the figure, only the "cross-shaped" target is tracked, which results in the "Y-shaped" missing, and at time 49, the "Y-shaped" target is missed. Therefore, at the time of the immediate vicinity of the target, the likelihood update is prone to errors using only the position information.
As shown in fig. 9, at time 48 as well, shape information is added to the likelihood of the update process, so that the problems of wrong tracking and missed tracking when the targets are close to each other are solved well.
Meanwhile, the figure also shows a comparison experiment of the B-spline method and the shape estimation method of the GIW-PHD, and the accuracy of the B-spline estimation shape can be seen.
Experiment three: two target track tracking
As shown in fig. 10, the weight is updated only by using the position information in the target tracking, and fig. 10- (1) shows the track tracking result at 100 times, and fig. 10- (2) shows that the track is updated only by using the position information, which may cause the phenomena of tracking error and tracking missing.
In the application, the track tracking is to add a track label in a Gaussian component. And performing track association by using the same label at different time. In the track association of the adjacent targets, the maximum weight is obtained by the likelihood obtaining method of the application, and the track association is carried out. As shown in fig. 11, a graph of the track following results using the filter likelihood update of the present invention. Fig. 11- (1) is a track tracking result at 100 time instants, and fig. 11- (2) is a detailed diagram of the time instants immediately adjacent to each other, and it can be seen from fig. 11- (2) that correct track tracking can still be performed at the time instants immediately adjacent to each other by using the filtering algorithm of the present invention.
To evaluate the quasi-certainty and stability of the invention, 100 experiments under the same conditions were performed, and the average OSPA distance and weight sum of 100 experiments was calculated. As shown in FIG. 12, FIG. 12- (1) shows the average OSPA distance over 100 time instants, and it can be seen that the OSPA value rises slightly in the immediate time instant, but the overall OSPA distance can always be controlled below 2. Fig. 12- (2) shows the average weighted sum over 100 time instants, and it can be seen that the weighted sum is rounded to 2, i.e. the number per time instant is estimated to be 2.
Experiment four: two object shape classifications
In the filtering updating method, the likelihood degree of the two shapes can be extracted better by utilizing the shape likelihood information. The method has great help to the tracking and classification of the target, and after useful target information is extracted by a filtering algorithm, the likelihood judgment is carried out on the target shape information and the preset shape information.
As shown in fig. 13, the classification result is a graph of shape information. Fig. 13- (1) shows the classification result at 100 times, and fig. 13- (2) shows that the shape fitting of the target at the new time is not enough, and the target is determined as another shape class before the likelihood with the shape information in the class information is low. In the shape classification after the filtering state extraction, the classification accuracy is evaluated, fig. 14 shows the average shape classification accuracy of 100 experiments, and it can be seen from the figure that the shape classification accuracy of the target at the time immediately adjacent is slightly reduced, but the accuracy is still maintained above 0.9.
Experiment five: target neogenesis
On the basis of two moving targets, a new target is set at the 60 th moment, falls into the monitoring areas of the two moving targets, and moves to the 100 th moment, and the new target keeps constant-speed linear motion.
As shown in fig. 15, fig. 15- (1) shows the classification results of three targets, and fig. 15- (2) shows the average classification accuracy obtained in 100 experiments. It can be seen that the new object moves at the beginning of the 60 th moment, and the shape is not well fitted, so that the category information of the new object cannot be accurately identified, and the accuracy rate is reduced. In the subsequent shape classification, the shape classification can be well divided. Fig. 16 shows the results of three target track traces. Fig. 17 and 18 show the average OSPA distance and the average weight sum of 100 experiments in the case of a targeted newborn by the filtering algorithm, respectively.
4. Quantitative analysis of the results of the experiment
In experiment one, verification of KDE-SSP method is carried out. In the method, only a plurality of determined peak points need to be found in the selection of candidate positions. Therefore, the method and the device can well divide the measurement set close to the target, and can control the time complexity in a lower range, so that the operation efficiency is higher.
In experiments two, three and four, the weights are updated by using likelihood calculation in the filtering updating algorithm respectively. When the target moves closely, the phenomenon of wrong tracking and missed tracking caused by inaccurate measurement set updating is easily caused by only using position updating. The shape information is added in the likelihood updating, the corresponding target state updating can be carried out on the measurement set, further the track information in the target motion can be effectively found out, and the target category can be accurately judged.
In experiment five, a nascent target was added to detect the effect on the method of the present application. As can be seen from the experimental result graph, the OSPA distance, the weight sum and the shape classification accuracy can be well adapted to the new situation with the target, and the algorithm has stronger stability.
What has been described above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiment. It is to be understood that other modifications and variations directly derivable or suggested by those skilled in the art without departing from the spirit and concept of the present invention are to be considered as included within the scope of the present invention.
Claims (10)
1. The method for tracking and classifying the multi-extension target track based on the B spline shape drive is characterized by comprising the following steps:
establishing a B-Spline-GM-PHD filter, initializing a target state at the moment k:
wherein x iskRepresenting a state of motion, BkRepresenting B-spline shape information, pk,i、θk,iRespectively represents the polar diameter and polar angle information at the ith angle at the moment k, LkA tag representing the gaussian component track at time k,a track label representing the jth Gaussian component at the time k;
when k is larger than or equal to 1, performing primary division on the measurement set by a distance division method;
if the divided measurement set has an adjacent target, performing secondary division on the divided measurement set, and if the divided measurement set does not have an adjacent target, performing multi-hypothesis filtering on the target state by using the B-Spline-GM-PHD filter;
performing multi-hypothesis filtering on the target state by adopting the B-Spline-GM-PHD filter to update the target state and obtain a shape likelihood function after weight updating;
fusing target components meeting a distance fusion threshold, pruning Gaussian components with undersized weights after fusion, and taking the Gaussian components with high weights as extraction targets in the fusion process, wherein the extraction targets comprise target motion states, target shape information and target tracks;
the likelihood is obtained by the shape likelihood function for the extracted target shape information and a preset target shape with category information, and target shape category information is obtained;
and if the observation information arrives at the next moment, re-executing the step of carrying out primary division on the measurement set by the distance division method, otherwise, ending the target track tracking process.
2. The method for tracking and classifying the multi-extension target track based on the B-spline shape driving according to claim 1, wherein the obtaining of the target shape category information by the likelihood of the extracted target shape information and the preset target shape with the category information by the shape likelihood function comprises:
set the preset target shape with the category information asThe set includes a preset' Y-shapeAnd a cross shapeClass information, wherein
Using all the extracted object shape information sets Bk|kEach shape of (a) is respectively associated with the preset target shape setUsing the shape likelihood function of the B-Spline-GM-PHD filter to obtain a formulaSolving for each extracted shapeAll have the most likely shape to matchAnd when the likelihood exceeds a category threshold, classifying the s-th extracted shape into a c-th category, and when the likelihood is smaller than the category threshold, classifying the s-th extracted shape into other categories.
3. The method for tracking and classifying the target track based on the B-Spline shape driving multiple extensions according to claim 1 or 2, wherein the performing multiple hypothesis filtering on the target state by using the B-Spline-GM-PHD filter to update the target state and obtain the shape likelihood function after the weight update comprises:
and sequentially carrying out target prediction and target update of the B-Spline-GM-PHD filter on the target state, wherein a weight update formula considering the target is as follows:
wherein,representing a position likelihood under a current metrology set, the position likelihood solved using a Gaussian distribution,representing the shape likelihood function, the calculation process is as follows:
wherein HW、RWDenotes an innovation parameter, r'k|k-1,iIndicates the predicted pole diameter, r 'after treatment'W,iRepresenting the measured collector diameter after processing, and lambda represents a shape similarity mark;
there are two cases for the shape likelihood function, when r'k|k-1,i|=|r′W,iIf not, the shape likelihood is calculated by the second formula.
4. The B-spline shape-driven multi-extended target track tracking and classifying method according to claim 3, wherein the processing method of the polar path comprises the following steps:
known predicted polar path informationPolar diameter information corresponding to the current measurement setTwo polar diameter information are divided into n1The dimension expansion is 360 dimensions, resulting in:
from the predicted polar diameter information r after dimension expansionk|k-1And measuring collector radius information r after dimension expansionWThe peak values of the wave crest and the wave trough obtained from the polar angle polar diameter diagram are respectively Pk|k-1And PWThe corresponding peak numbers are respectivelyAnd
and by comparing the number of the peak values of the two, the predicted polar diameter information after dimension expansion and the measured collective polar diameter information after dimension expansion are respectively reordered, and the polar angle information is correspondingly changed.
5. The method for tracking and classifying the flight path of the multi-extended target based on the B-spline shape drive of claim 4, wherein the step of reordering the predicted polar path information after the dimension expansion and the measured collective polar path information after the dimension expansion by comparing the number of peaks of the two comprises the steps of:
case 1, match to two peak points:
Pk|k-1(1,1:2)≈PW(1,u0:v)
v=mod(u0+1,Num),or v=Num
r′k|k-1=[rk|k-1(lk|k-1(1)),…,rk|k-1(end),rk|k-1(1),…,rk|k-1(lk|k-1(1)-1)]
r′W=[rW(lW(u0)),…,rW(end),rW(1),…,rW(lW(u0)-1)]
λ=1
case 2, matching to a peak point:
Pk|k-1(1)≈PW(u0)
r′k|k-1=[rk|k-1(lk|k-1(1)),…,rk|k-1(end),rk|k-1(1),…,rk|k-1(lk|k-1(1)-1)]
r′W=[rW(lW(u0)),…,rW(end),rW(1),…,rW(lW(u0)-1)]
case 3, no peak point matched:
Pk|k-1(1)≠PW(u0)
r′k|k-1=rk|k-1
r′W=rW
r′k|k-1=rk|k-1
r′w=rw
wherein "≈" indicates that the peak size of the predicted shape and the peak size of the metrology set shape are equal within a threshold range.
6. The method for tracking and classifying a multi-extended target track based on B-spline shape driving according to claim 1, wherein the secondarily partitioning the partitioned measurement sets comprises:
initializing a candidate shape parameter set C when k is 1:
wherein,candidate shape parameters for fitting a B-spline curve,respectively representing polar diameter and polar angle information at the ith angle; initializing polar anglesIs composed ofn1Is a predetermined dimension, and n1< 360, and the polar diameter information corresponding to the initialized polar angle is
When k is larger than or equal to 2, selecting m targets in the posterior target states, and putting shape information of the m posterior target states into the candidate shape parameter set C;
determining the central point position of each candidate shape in the candidate shape parameter set C by adopting a kernel density estimation method, and establishing a candidate central position set L:
wherein the coordinates corresponding to the density peak points of the candidate shape to which the B-spline curve is fitted are represented by (ω)q (x),ωq (y)) N represents the number of density peak points;
the candidate shape parameter set C and the candidate center position set L are combined intoAnd adopting a shape selection division method to divide the combination to obtain an inner shape measurement subset and an outer shape measurement subset which are respectively as follows:
wherein,a subset of the intra-shape measurements is represented,representing a subset of extratopographic measurements, zkRepresenting the measurement points in the measurement set W,represents the measurement point zkTo the central point position omegaqIs indicated byThe point corresponding to the point(s) of the image,the included angle between the connecting line of the central point position and the h point on the shape and the x axis is shown,an angle between a connecting line representing the position of the center point and the measuring point and the x-axis is set, and when the absolute value of the difference between the two is minimum, a point on the shape at the angle is selected as psi,representing psi to center point position omegaqThe distance of (a);
wherein, bΨVector, p, representing the location of the center point to Ψk| k-1Coordinate information representing control vertices generated from the subset of intra-shape measurements, Ni,3(ψ) represents a B-spline basis function;
and respectively matching the shape internal measurement subset and the shape external measurement subset to obtain the likelihood, and obtaining the condition of the maximum weight product of the shape internal measurement subset and the shape external measurement subset.
7. The method for tracking and classifying tracks of multiple extended targets based on B-spline shape driving according to claim 6, wherein the matching the in-shape measurement subset and the out-shape measurement subset respectively for the likelihood to obtain the case of the maximum weight product of the in-shape measurement subset and the out-shape measurement subset comprises:
for intra-shape measurement subsetsCalculating likelihood by using the shape used when the shape is fitted and combined with the segmentation by using the intra-shape measurement subset, wherein the process of calculating the likelihood is a process of updating a shape likelihood function by using the B-S pline-GM-PHD filter;
for the out-of-shape measurement subsetSolving the weights by using the same process of updating the shape likelihood function, traversing all candidate shapes to respectively fit the shapes with the measurement set to obtain likelihood, and selecting the maximum value of the weights as the external measurement subset of the shapesAnd finding the condition that the likelihood product of the shape internal measuring subset and the shape external measuring subset is maximum as the final segmentation result.
8. The method for tracking and classifying the multi-extended target track based on the B-spline shape driving according to claim 6, wherein the method for fitting the shape by adopting the B-spline curve comprises the following steps:
the B-spline curve is composed of a B-spline basis function Ni,k(u) and control vertex PiCollectively determined, the k-th order B-spline function is represented as:
wherein, Ni,k(u) is a k-1 degree B-spline basis function whose recurrence formula is:
wherein, U is { U ═1,u2,…,uk,…,un+kIs a set of node vectors for the B-spline;
when k is 3, the B-spline basis function is expressed as:
the process of obtaining the control vertex set in the B spline function by using the position information of the measuring points, wherein the measuring set of the extended target comprises a plurality of measuring points, comprises the following steps:
suppose thatCalculating the mean position of the measurement set for the measurement set of the extended target, establishing a coordinate system by taking the mean position as the origin of coordinates, dividing the coordinate system with the angle of 2 pi into n equal parts to obtain n control vertex positions, and expressing the control vertexes Lambda by polar coordinatesi={ρi,θi},ρi、θiRespectively, the pole diameter and the pole angle, and the pole diameter is expressed as:
|Zi|={zk|d(zk,βi)<λ1,k(zk,αi)=1}
βi=[-tan(θi),1]
wherein, | · | represents the number of elements, d (z)k,βi) Represents the measurement point zkTo betaiDistance of straight line, λ1Distance threshold for constraint width, κ (z)k,αi) As a constraint, βiIs polar angle thetaiThe calculation formula of the determined vector, d (-) is as follows:
αi=[1,tan(θi)]
wherein, | | · | | represents an absolute value, αiRepresentation and vector betaiA vector perpendicular to the straight line;
and through the calculation, solving a control vertex set Λ under a polar coordinate, converting the control vertex set Λ into a corresponding set P under a rectangular coordinate system, and generating a smooth curve related to the shape of the measurement set under the combined action of the control vertex set Λ and the B spline basis function.
9. The B-Spline shape-driven multi-expansion target track tracking and classifying method according to claim 3, wherein the prediction step of the B-Spline-GM-PHD filter is represented as:
Lk|k-1=Lk-1∪Lβ
wherein,representing the probability hypothesis density of the predicted target, Jk|k-1Represents the number of predicted gaussian components at the moment k,respectively representing the weight, mean and covariance matrix of the jth Gaussian component predicted at the time k, N (-) is Gaussian distribution, Lk|k-1The label set representing the prediction at the time k comprises the label sets of the original target and the new target at the time k-1, and the predicted polar diameter and polar angle information are consistent with the posterior target state at the time k-1;
wherein, FK|k-1Representing a state transition matrix, IdRepresenting a d-dimensional identity matrix, QkRepresenting the process noise covariance matrix.
10. The method for tracking and classifying the track of the target based on the B-Spline shape driving multi-extension target of claim 3, wherein the updating step of the B-Spline-GM-PHD filter is expressed as:
wherein, the probability hypothesis density of the missed detection target is expressed as:
Lk|k=Lk|k-1
wherein, γjA desired value representing the number of measurements made by the target,representing the target detection probability:
the probability hypothesis density for target detection at time k is given by:
where σ is expressed as a preset forgetting factor, ρW,iFor the polar diameter information at the ith polar angle of the shape to which the current measurement set is fitted,the sum of the measurement set numbers in all the division methods is represented, and each innovation parameter is represented as:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110166401.4A CN112946625B (en) | 2021-02-04 | 2021-02-04 | B-spline shape-based multi-extended target track tracking and classifying method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110166401.4A CN112946625B (en) | 2021-02-04 | 2021-02-04 | B-spline shape-based multi-extended target track tracking and classifying method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112946625A CN112946625A (en) | 2021-06-11 |
CN112946625B true CN112946625B (en) | 2022-07-15 |
Family
ID=76243010
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110166401.4A Active CN112946625B (en) | 2021-02-04 | 2021-02-04 | B-spline shape-based multi-extended target track tracking and classifying method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112946625B (en) |
Families Citing this family (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114002667A (en) * | 2021-10-29 | 2022-02-01 | 西安交通大学 | Multi-neighbor extended target tracking algorithm based on random matrix method |
WO2024005063A1 (en) * | 2022-07-01 | 2024-01-04 | 株式会社デンソー | Object tracking device |
CN115222300B (en) * | 2022-09-20 | 2023-02-28 | 珠海翔翼航空技术有限公司 | Flight simulator retraining duration distribution system and method based on intelligent evaluation algorithm |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2471276A (en) * | 2009-06-22 | 2010-12-29 | Bae Systems Plc | Terrain sensing apparatus for an autonomous vehicle |
CN102879766B (en) * | 2011-07-11 | 2015-05-06 | 哈尔滨工业大学 | Method and apparatus for detecting and tracking faint target of high frequency ground wave radar |
US20180372857A1 (en) * | 2016-01-06 | 2018-12-27 | Mitsubishi Electric Corporation | Target tracking apparatus |
CN108072865A (en) * | 2016-11-10 | 2018-05-25 | 北京机电工程研究所 | Tracking before a kind of surface vessel target detection |
CN110320512A (en) * | 2019-07-09 | 2019-10-11 | 大连海事大学 | A kind of GM-PHD smothing filtering multi-object tracking method based on tape label |
-
2021
- 2021-02-04 CN CN202110166401.4A patent/CN112946625B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN112946625A (en) | 2021-06-11 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN112946625B (en) | B-spline shape-based multi-extended target track tracking and classifying method | |
Zhang et al. | Deep neural networks for wireless localization in indoor and outdoor environments | |
Shen et al. | Fast online tracking with detection refinement | |
Groß et al. | Alignnet-3d: Fast point cloud registration of partially observed objects | |
CN106780557B (en) | Moving object tracking method based on optical flow method and key point features | |
CN104200495B (en) | A kind of multi-object tracking method in video monitoring | |
CN110197502B (en) | Multi-target tracking method and system based on identity re-identification | |
CN110298865B (en) | Space-based starry sky background weak small point target tracking method based on threshold separation clustering device | |
CN111553409A (en) | Point cloud identification method based on voxel shape descriptor | |
Goli et al. | nerf2nerf: Pairwise registration of neural radiance fields | |
Cao et al. | Correlation-based tracking of multiple targets with hierarchical layered structure | |
CN112652003B (en) | Three-dimensional point cloud registration method based on RANSAC measure optimization | |
Soleimanitaleb et al. | Single object tracking: A survey of methods, datasets, and evaluation metrics | |
CN108537805A (en) | A kind of target identification method of feature based geometry income | |
Zeng et al. | Robust multivehicle tracking with wasserstein association metric in surveillance videos | |
CN111830501A (en) | HRRP (high resolution representation protocol) historical characteristic assisted signal fuzzy data association method and system | |
Li et al. | Adaptive and compressive target tracking based on feature point matching | |
Gong et al. | A two-level framework for place recognition with 3D LiDAR based on spatial relation graph | |
Wei et al. | RSSI-based location fingerprint method for RFID indoor positioning: a review | |
Zhu et al. | Deepego: Deep instantaneous ego-motion estimation using automotive radar | |
Chen et al. | Multi-neighborhood guided Kendall rank correlation coefficient for feature matching | |
Yu et al. | Cattrack: Single-stage category-level 6d object pose tracking via convolution and vision transformer | |
CN113532438A (en) | Improved ICCP terrain matching method under large initial positioning error | |
Zhou et al. | Place recognition and navigation of outdoor mobile robots based on random Forest learning with a 3D LiDAR | |
Usuelli et al. | Radarlcd: Learnable radar-based loop closure detection pipeline |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |