A method for maneuvering multi-target tracking under doppler blind zone

Abstract

The application belongs to the field of motorized multi-target tracking, and particularly relates to a multi-target tracking method under a Doppler blind area, comprising the following steps: constructing a Doppler radar detection probability model based on target motion state, clutter notch function and sensor state; constructing a measurement model for target state estimation and tracking based on the Doppler radar detection probability model, and performing a prediction process to obtain a multi-target prior density; in the algorithm updating process, additional pseudo-measurement components and enhanced measurement components are added, so that the target still maintains a high weight to prevent the track from being trimmed in the case that the target loses measurement in the Doppler blind area, and the track is redistributed after the target drives out of the blind area, thereby ensuring the normal operation of the tracker; based on the measurement set of the Doppler radar, an updating process is performed to calculate the multi-target posterior density of the IMM-GLMB algorithm; and the target motion state is extracted according to the IMM-GLMB model to obtain a current target motion state estimation result.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of motorized multi-target tracking, and particularly relates to a motorized multi-target tracking method under a Doppler blind zone based on an interacting multiple model-generalized labeled multi-Bernoulli (IMM-GLMB) filter. BACKGROUND

[0002] The biggest feature of Doppler radar is that the Doppler measurement can be obtained in addition to the position measurement, and the Doppler measurement can be introduced into the tracking algorithm to effectively improve the target tracking performance. However, due to the physical limitations of the Doppler radar sensor, there is inevitably a Doppler blind zone (DBZ). When the Doppler measurement of the target falls within the minimum detectable speed interval of the Doppler radar sensor, the Doppler radar cannot detect the moving target, resulting in problems such as intermittent, temporary disappearance, re-starting batch, and batch interruption. The existence of the Doppler blind zone greatly increases the complexity of the multi-target tracking problem of the Doppler radar. In the process of radar tracking, the motion state of the target and the matching degree of the motion model in the tracker directly affect the performance of the tracker. When the two are matched, the tracking performance can be improved, otherwise the tracking performance will be worsened, and even the filter will diverge. In the actual tracking scene, the tracked target usually does not maintain a single running state, such as turning, rolling, and diving during the flight of the aircraft. A single motion model cannot guarantee the tracking accuracy.

[0003] Therefore, how to reduce the influence of the Doppler blind zone on the tracker and track the motorized target under the blind zone is the research focus of scholars in the relevant field. SUMMARY

[0004] In order to overcome the problems in the prior art, the application provides a motorized multi-target tracking method under a Doppler blind zone.

[0005] The technical solution of the application to solve the above technical problems is as follows:

[0006] The application provides a motorized multi-target tracking method under a Doppler blind zone, comprising the following steps:

[0007] Based on the target motion state, the clutter notch function, and the sensor state, a Doppler radar detection probability model is constructed;

[0008] Based on the Doppler radar detection probability model, a measurement model for estimating and tracking the target state is constructed, and a prediction process is performed to obtain a multi-target prior density;

[0009] During the measurement model update process, additional pseudo-measurement components and enhanced measurement components are added; and the multi-target posterior density is calculated.

[0010] Extract the target motion state according to the measurement model to obtain the current target motion state estimation result.

[0011] Furthermore, based on the minimum detectable velocity, clutter notch function, and detection probability of the Doppler radar sensor, a radar detection probability model is constructed, specifically including:

[0012]

[0013] where p , (x) is the probability that the Doppler radar sensor detects the target; p D represents the detection probability when the Doppler velocity of the target is away from the Doppler blind zone; represents the normalization factor; R f = MDV 2 / (2ln2) represents the variance of the pseudo-measurement in the Doppler domain; represents the pseudo-measurement function; is the pseudo-measurement matrix; x represents the motion state of the target;

[0014] The clutter notch n<00​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​+，γ (·, l) denote the survival and birth target densities, respectively, for a given label set L; is a normalization parameter, τ(μ) denotes the model probability for model μ; c denotes the component, denotes the component set; the prediction set weight is:

[0019]

[0020] The single-target mixture density is:

[0021]

[0022] where, denotes the weight of the survival label denotes the weight of the birth label and p +，γ (·, l) denote the survival and birth target densities, respectively.

[0023] Further, the multi-target posterior density is:

[0024]

[0025] where τ(μ) denotes the model probability for model μ;

[0026] where the single-target posterior probability density under model μ r

[0027]

[0028] are the regular loss measurement mean, covariance, weight; are the pseudo measurement mean, covariance, weight; are the position measurement mean, covariance, weight; are the augmented measurement mean, covariance, weight;

[0029] The update weight is:

[0030]

[0031] The single-target normalization constant is:

[0032]

[0033] In the above formula, is the regular loss measurement weight; is the pseudo measurement weight; is the position measurement weight;​​​ represents the enhanced measurement weight; represents the track association mapping of the label l, when θ(l) = 0, it means that the measurement is not associated with the track, at this time the track is missed, when θ(l) > 0, it means that the track is updated with the measurement information; δ0(X) represents the generalized Kronecker delta function.

[0034] Compared with the prior art, the present application has the following technical effects:

[0035] The present application models the change rule of the detection probability of the Doppler radar sensor as a function related to the state of the moving target and the current azimuth of the Doppler radar sensor, and then substitutes the modeled detection probability model into the MM-GLMB model, and on the basis of the original conventional missing measurement component and position measurement component, the pseudo measurement component and the enhanced measurement component are added, when the observed target enters the Doppler blind area, although the measurement information is lost, the corresponding weights of the pseudo measurement component and the enhanced measurement component will also increase, so that the track can be avoided to be trimmed off, when the target drives out of the blind area, the track information can be immediately distributed, and the influence of the blind area on the tracker is reduced. At the same time, the interactive multiple model algorithm is used to estimate the motion state of the target, which can adapt to the maneuvering change of the target and improve the tracking precision. The method has strong robustness and adaptability, and can realize effective tracking of the maneuvering multi-target in a complex radar environment. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art and the advantages thereof, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows, obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without any creative labor.

[0037] Figure 1 IMM-GLMB filtering diagram of the present application;

[0038] Figure 2 IMM-GLMB-MDV filtering diagram of the present application;

[0039] Figure 3 OSPA (Optimal sub-pattem assignment) distance comparison diagram of the present application;

[0040] Figure 4 Flowchart of the present application. DETAILED DESCRIPTION

[0041] In order to further illustrate the technical means taken by the present application and the effects achieved in order to achieve the predetermined inventive objectives, the specific implementation, structure, features and effects of the technical solutions according to the present application are described in detail below in combination with the drawings and preferred embodiments. The specific features, structures or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as understood by those skilled in the art to which the present application belongs.

[0042] The present application aims at the situation that the existing GLMB (Generalized Labeled Multi-Bernoulli, generalized labeled multi-Bernoulli) algorithm faces the Doppler blind area in the process of tracking the maneuvering target, the tracking performance drops sharply or even cannot continue tracking, and proposes a maneuvering multi-target tracking method suitable for the GLMB under the Doppler blind area.

[0043] The present application simultaneously uses MDV (Minimum Detectable Velocity, minimum detectable velocity) information to model the detection probability, so that the target can still have a high weight corresponding to the track after entering the blind area without measurement information, so as to avoid being cut, and the track information can be immediately redistributed when the target drives out of the blind area, thereby reducing the influence of the Doppler blind area on the tracker.

[0044] Reference Figure 4 A maneuvering multi-target tracking method under the Doppler blind area, comprising the following steps:

[0045] Step 1. Based on the target motion state, the clutter notch function and the Doppler radar sensor state, a Doppler radar detection probability model is constructed.

[0046] Suppose that the target motion state at k time is Where (x k ,y k ) represents the target position component, represents the target velocity component; correspondingly, the Doppler radar sensor state is Where, represents the Doppler radar sensor position component, represents the Doppler radar sensor velocity component.

[0047] The probability p D,k (x) that the Doppler radar sensor detects the target is directly affected by the clutter notch function n c , n c represents the difference between the Doppler velocity of the target relative to the sensor and the Doppler velocity of the nearby clutter.

[0048]

[0049] where, denotes the Doppler measurement of the target:

[0050]

[0051] denotes the background clutter Doppler measurement:

[0052]

[0053] In the above equation, x k denotes the target state matrix, x, vx, y, vy, respectively, denote the x-axis coordinate, x-direction velocity, y-axis coordinate, y-direction velocity, x s is the Doppler radar state matrix,

[0054] The main role of the clutter notch is to suppress the clutter, but at the same time it will also affect the detection probability of the target. Here the detection probability is modeled as a function p D = p D,k (x) related to the target state and the current Doppler radar sensor azimuth.

[0055] When the target enters the Doppler blind zone, i.e. n c <MDV (minimum detectable velocity), the Doppler radar sensor is difficult to obtain the measurement value of the target, at this time p Dk (x) tends to 0; on the contrary, when the target is far away from the Doppler blind zone, i.e. n c > MDV, at this time, p D,k (x) will tend to p D , p D denotes the detection probability when the Doppler velocity of the target is far away from the Doppler blind zone, so the Doppler blind zone can be described as DBZ = {x | n c (x k ) < MDV}, the probability p D,k (x) that the Doppler radar sensor detects the target is:

[0056] p D,k (x) ≈ p D [1-exp(-(n c (x k ) / MDV) 2 log2)]

[0057] Since the detection probability in the above equation is in exponential form, it needs to be converted into Gaussian form before subsequent calls. For the nonlinear clutter notch function n c , a first-order Taylor expansion can be performed near the predicted value , which can be obtained:

[0058]

[0059] Among them, pseudo-measurement function for:

[0060]

[0061] The pseudo-measurement matrix is:

[0062]

[0063] In the above formula:

[0064]

[0065] These are the x and y coordinates of the predicted value, respectively. These represent the predicted velocities along the x and y axes, respectively.

[0066] The above approximation of n c Substituting (x) into the equation, we get:

[0067]

[0068] In the above formula, p D R represents the detection probability when the target's Doppler velocity moves away from DBZ; f =MDV 2 / (2ln2) represents the variance of the pseudo-measurement in the Doppler domain; This represents the normalization factor. H represents the Gaussian density of the mean m and variance P. k Let R be the observation matrix, and R be the position measurement noise covariance.

[0069] Step 2. Based on the Doppler radar detection probability model, construct a measurement model for target state estimation and tracking, and perform a prediction process to obtain the multi-target prior density.

[0070] Let the multi-objective state at time k-1 be X = {x1, x2, ... x}. |X|}, where x = [x k-1 ,μ k-1 ,l k-1 ]∈X; x, μ, l represent the motion state, motion model, and label state of the target, respectively.

[0071] Define the transition probability matrix of the Markov model as follows:

[0072]

[0073] Where, χ(μ) i|μ j ) represents the probability that the motion model jumps from μ j to μ i .

[0074] The state jumps from x + to x + at time k, and the Markov state transition density is given by:

[0075] φ(x + ,μ + ,l + |x,μ,l)=φ(x + |x,μ + ,l)χ(μ + |μ)IMM algorithm (interacting multiple model algorithm) needs to use model probability to mix the target motion state before prediction, and the multi-target prior density of IMM-GLMB after model state mixing:

[0076]

[0077] where the prediction set weight is:

[0078]

[0079] The single-target mixed density is:

[0080]

[0081] In the above formula, represents the weight of the surviving label , and represents the weight of the new label . and respectively represent the surviving target density and the new target density for a given label set L. x + represents the target prediction state, μ + and l respectively represent the corresponding model and label.

[0082] The single-target densities of the surviving target and the new target are respectively:

[0083]

[0084] where, is the surviving component weight, and are the mean and covariance of the surviving component respectively, is the new target weight, and are the mean and covariance of the new target respectively.

[0085] For the normalization parameter:

[0086]

[0087] In the above equation, τ(μ) denotes the model probability of the Markov model μ.

[0088] The parameters required for the density propagation process are given by:

[0089] The weight of the surviving target prediction density is:

[0090]

[0091] In the above equation, p denotes the scaling factor, p S denotes the survival probability of the target at the next time instant.

[0092] The surviving target requires an interaction input of the input state before the prediction step:

[0093]

[0094] The mean and covariance of the state mixture are used for the time prediction.

[0095]

[0096] In the above equation, F(μ and Q(μ denote the mean and covariance of the surviving component, respectively, F(μ + ) denotes the state transition matrix, and Q(μ + ) denotes the noise covariance.

[0097] The weight of the newborn target prediction density is:

[0098]

[0099] In the above equation, p denotes the existence probability of the newborn target.

[0100] Since the newborn target does not have a prior model probability, the model mixture step is not required, and the mean and covariance of the newborn target can be directly calculated.

[0101]

[0102]

[0103] Step 3. In the measurement model updating process, additional pseudo-measurement components and enhanced measurement components are added to keep the target with high weight to prevent track pruning in the Doppler blind area and to reallocate the track after the target leaves the blind area, ensuring normal operation of the tracker; and the multi-target posterior density is calculated.

[0104] Let the measurement set of the Doppler radar sensor at time k be Z = {z1, z2, …, z |Z|}, z = [y c , y d ], y c denote the position measurement information, and y d denote the Doppler measurement information.

[0105] At this time, the multi-target posterior density is:

[0106]

[0107] where the single-target posterior probability density under the model μ r is:

[0108]

[0109] The updated weight is:

[0110]

[0111] The single-target normalization constant is:

[0112]

[0113] In the above formula, θ(l) denotes the track association mapping with label l, when θ(l) = 0, it means that the measurement is not associated with the track, at this time the track is missed, and when θ(l) > 0, it means that the track is updated with the measurement information. δ0(X) denotes the generalized Kronecker delta function, which is defined as:

[0114] The mean, covariance and weight in the conventional lost measurement are given by the following formula:

[0115]

[0116] where and are the mean, covariance and weight obtained by the prediction process.

[0117] The mean, covariance and weight in the pseudo-measurement are obtained by the following formula:

[0118]

[0119] where the pseudo-measurement gain K i,f is given by

[0120]

[0121] The pseudo-measurement innovation covariance is

[0122]

[0123] The weight of the pseudo-measurement component is

[0124]

[0125] In the above equation, c f is a normalization factor given in step one in modeling the detection probability

[0126] The likelihood probability of the component is

[0127]

[0128] The mean, covariance and weight in the position measurement are obtained by processing the position measurement and Doppler measurement. First, the position measurement information y c is used to update the parameters:

[0129]

[0130]

[0131] where H c is the target observation matrix, the gain and innovation covariance of the position measurement are

[0132]

[0133] Then the Doppler information y d is used to sequentially update:

[0134]

[0135] The weight of the position component is

[0136]

[0137] where are the clutter intensities for the position component and Doppler component, respectively.

[0138] The Doppler measurement gain and covariance are

[0139] ​

[0140] where R c and σ d are the position measurement and Doppler measurement noise standard deviations, respectively.

[0141] Augmented measurement The mean, covariance and weights in the augmented measurement are given by:

[0142]

[0143]

[0144] where the gain and covariance of the augmented measurement are:

[0145]

[0146] Augmented measurement matrix By substituting into the following equation.

[0147]

[0148] In the equation:

[0149]

[0150] The weights of each component are given by:

[0151]

[0152] Step 4. Extract the target motion state according to the measurement model to obtain the current target motion state estimation result.

[0153] Unlike the multi-model, the IMM algorithm needs to perform model state mixing after all components are calculated to extract the target state. The model mixing process is as follows:

[0154]

[0155] The model probability updating process is:

[0156]

[0157] where c is a normalization parameter

[0158] It is worth noting that the symbolic expressions of each component in this embodiment are consistent in meaning, and the subscript is only used to distinguish different components.

[0159] The experimental condition in the application is: the observation scene is a two-dimensional plane of [-1000 1000] (m) x [-1000 1000] (m), the scanning period of the Doppler radar sensor is 1s, the observation time is 100s, and the target motion model set is composed of a uniform linear motion, a right turn model (2° / s turning rate cooperative turning model), and a left turn (-2° / s turning rate cooperative turning model). In this experimental scene, there are three observation objects, which appear at different positions at different times, and the new positions are (-400, 380), (-400, 410), and (-700, 400). The motion state of the target is shown in Table 1, and the state transition equation is: x + =F(μ)x+v, wherein the state transition matrix corresponding to the three motion models is obtained by bringing ω=0 (when μ=1), ω=2π / 180 (when μ=2), and ω=-2π / 1 (when μ=3) into F(μ), and is F(μ1), F(μ2), and F(μ3) respectively.

[0160] Table 1 Motion state of target

[0161]

[0162]

[0163]

[0164] wherein T represents the scanning period of the Doppler radar sensor, T=1s, v represents process noise with a covariance Q and a mean value of 0. v =10m / s, the survival probability of each target is p s =0.99, the saturation value of the Doppler radar sensor detection probability is p D =0.98, the prior velocity standard deviation of the target is σ s =17m / s, the position measurement noise standard deviation is σ c =10m / s, the Doppler observation noise standard deviation is σ d =0.5m / s, the number of clutter points in each period obeys a Poisson distribution with a mean value of 25, and the position of each clutter point is uniformly distributed in the measurement range. The model probability of the new target is τ0=[1 / 3 1 / 3 1 / 3], in order to facilitate comparison, the pruning parameter of all filters is T=10 -5 , the merging threshold is U=4, and the maximum number of Gaussian components is J max =100, and the multi-target extraction threshold is 0.5.

[0165] The initial motion state of the target is:

[0166] x1=[-400 7 380 -7]

[0167] x2 = [-400 -6410 -10]

[0168] x3 = [-700 3400 -14]

[0169] The Markov matrix for switching between target motion models is set as:

[0170]

[0171] Figure 1 and Figure 2 The IMM-GLMB filtering result and the IMM-GLMB-MDV filtering result respectively, it can be seen that when the target drives into the Doppler blind area, the measurement will be immediately lost, and even after driving out of the blind area, it still cannot be normally tracked, and the filter combined with MDV information will lose part of the measurement in the blind area, but can be immediately restored when the target drives out of the blind area, effectively reducing the influence of the Doppler blind area on the performance of the tracker.

[0172] In summary, the direct influence of the Doppler blind area on the radar is that it will cause continuous loss of measurements of the moving target, when the target enters the Doppler blind area, once the continuously lost measurement points meet the track deletion threshold, the target will be missed, and after the target drives out of the Doppler blind area, because the distance between the new measurement information and the inherent threshold has a large error, the target will be misdetected. The present application models the detection probability of the Doppler radar sensor according to the reason for the formation of the Doppler blind area, and applies the IMM algorithm to the GLMB tracking algorithm according to the characteristics of the observed target high maneuverability, and realizes a maneuvering multi-target tracking under the sheltering of the Doppler blind area.

[0173] The above examples are only used to illustrate the technical solutions of the present application, but not limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for maneuvering multi-target tracking under Doppler blind zone, characterized in that, The method comprises the following steps: Based on the target motion state, clutter notch function and sensor state, a Doppler radar detection probability model is constructed; Based on the Doppler radar detection probability model, a measurement model for target state estimation and tracking is constructed, and a prediction process is performed to obtain a multi-target prior density; In the measurement model updating process, additional pseudo-measurement components and enhanced measurement components are added; And calculate the multi-target posterior density; According to the measurement model, the target motion state is extracted to obtain the current target motion state estimation result.

2. The method of claim 1, wherein, Based on the minimum detectable speed, the clutter notch function and the detection probability of the Doppler radar sensor, a radar detection probability model is constructed, specifically including: wherein is the probability that the Doppler radar sensor detects the target; represents the detection probability when the Doppler velocity of the target is far from the Doppler blind zone; represents a normalization factor; represents the variance of the measurement in the Doppler domain; represents the pseudo-measurement function; is the pseudo-measurement matrix; x represents the motion state of the target; represents the Gaussian density of the mean , variance . The clutter notched function is expressed as a difference in Doppler velocity of the target relative to the nearby clutter Doppler velocity; When the target enters the Doppler blind zone, i.e. where MDV is the minimum detectable velocity, the Doppler radar sensor is difficult to obtain the measurement value of the target, at this time tends to 0; on the contrary, when the target is far away from the Doppler blind zone, i.e. , at this time will tend to .

3. A method of maneuvering multi-target tracking under Doppler blind zone according to claim 2, characterized in that, The multi-target prior density is: wherein, denotes a label mutual distinctiveness indicator, defined as i.e. when the labels of the elements in the set X are mutually distinct, ; otherwise, ; denotes the Dirac delta function, denotes the state probability density of the single objective, denotes the weight of the corresponding component, is a normalization parameter; denotes the label is the weight denotes the single target density for a given set of labels ; is a normalization parameter, , denotes the model probability when the model is ; representing components, representing a set of components; the prediction set weights are: The single-target mixed density is: wherein denotes the weight of the survival label , denotes the weight of the newborn label ; and denote the survival target density and the newborn target density, respectively.

4. The method of claim 3, wherein, The multi-target posterior density is: wherein the model probability when the model is the model probability when the model is where the single-target posterior probability density under the model is given by = mean, cov, weight of regular loss measurement = mean, cov, weight of pseudo measurement = mean, cov, weight of position measurement = mean, cov, weight of augmentation measurement The update weight is: The single-target normalization constant is: In the above formula, conventional loss-of-measurement weight; representative pseudo-measurement weight; representative position-measurement weight; representative augmented-measurement weight; represents track association mapping for a label l when when there is no association between the measurement and the track, in which case the track is missed, when the track is updated with the measurement information; represents the generalized Kronecker delta function.

Images