A multi-target tracking method and processing device with adaptive new target initiation
Through the adaptive new target starting method of Gaussian mixed probability assumption density filter, the problem of new target detection in multi-objective tracking is solved, the target number and state estimation in the case of unknown new target intensity is achieved, and the steady-state accuracy and Computational efficiency.
Patent Information
- Application Number
- CN201910663755.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2019-07-19
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2039-07-19
AI Technical Summary
The prior art has random errors, clutter interference and detection uncertainty in multi-object tracking, especially when the intensity of the new target is unknown, it is difficult to effectively detect the new target, resulting in a biased target number estimate.
Adaptive nascent target starting method based on Gaussian mixed probability assumption density filter, by predicting and updating the Gaussian elements and intensity of the surviving target, and pruning and combining the measurements, assigning labels to distinguish nascent targets and reducing unnecessary Calculation quantity.
In the case where the intensity of the new target is unknown, the number and state of the target can be estimated simultaneously, the steady-state estimation accuracy can be improved, the calculation amount will be reduced, and the new target and clutter can be effectively distinguished.
Smart Images

Figure CN112241006B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of target tracking, and in particular to a multi-target tracking method that is adaptive to the initiation of new targets. Background Art
[0002] In the prior art, traditional multi-target tracking solutions often estimate the motion states and numbers of multiple targets based on the measurement sequences detected by sensors and generated by multiple moving targets and environmental noise. However, in the actual multi-target tracking problem, the sensor's measurement has random errors and there is clutter interference. In addition, targets appear and disappear randomly in the observation area, and the number of targets is often unknown. Secondly, there is detection uncertainty and association uncertainty in the tracking process. All of these factors make multi-target tracking challenging.
[0003] The multi-target tracking method based on random finite sets (RFS) avoids explicit data association and can be applied when the target is time-varying and unknown. It has received widespread attention in recent years. The probability hypothesis density (PHD) filter proposed by Mahler makes the multi-target tracking method based on RFS practical. At present, the main methods for implementing PHD filters are the sequential Monte Carlo (SMC-PHD) method and the Gaussian mixture (GM-PHD) method. Compared with the SMC-PHD method, the GM-PHD filter peak extraction method is simple, and the amount of calculation can be effectively controlled by merging and pruning Gaussian elements.
[0004] The standard PHD recursive process assumes that the target's new intensity is known a priori. However, when a new target appears in a place that is not covered by the new intensity, the PHD filter cannot detect it. Generating new targets in the entire state space is a common solution. However, covering the entire state space requires a large amount of computation. Therefore, a measurement-driven new target generation method has emerged to overcome this problem. This method can adaptively use measurements to generate new target intensities without any prior information. However, directly applying the measurement-driven new target intensity generation method to PHD will produce biased estimates of the number of targets. A PHD filter with adaptive new target initiation is proposed to solve this problem. This method can distinguish between existing targets and new targets, and recurse separately in the prediction and update steps. After searching the existing technical literature, it was found that Ouyang Cheng et al. (Ouyang Cheng, Hua Yun, Gao Shangwei. Improved adaptive new target intensity PHD filter [J]. Systems Engineering and Electronics, 2013, 35(12): 2452-2458.) proposed an improved adaptive new target initiation PHD filter, but this method did not utilize all targets to generate measurements. When the number of targets is large, a lot of computing time is spent on processing the surviving target measurements generated from the clutter. Summary of the invention
[0005] In view of the above-mentioned defects of the prior art multi-target tracking method with adaptive new target initiation, the present invention provides a multi-target tracking method and processing device with adaptive new target initiation based on Gaussian mixture probability hypothesis density filter. The method can be used in the fields of radar, infrared, visible light target tracking, detection and tracking of multi-rotor drones, and tracking and positioning of mobile robots.
[0006] According to one aspect of the present invention, a multi-target tracking method with adaptive new target initiation is provided, comprising the following steps:
[0007] Step S1: Given multi-target parameters before the current moment, the multi-target parameters include Gaussian elements and intensities of surviving targets, Gaussian elements and intensities of newborn targets, a measurement set not associated with the target, and a corresponding measurement number set;
[0008] Step S2: performing a one-step prediction on the Gaussian element and intensity of the surviving target at the current moment to obtain the predicted Gaussian element set and intensity of the surviving target;
[0009] Step S3: If there is a new sensor observation value at the current moment, the Gaussian element and intensity of the surviving target in step S2 are updated to obtain the predicted Gaussian element set and intensity of the surviving target, and the measurement and its corresponding Gaussian element are assigned corresponding labels;
[0010] Step S4: Prune and merge the updated Gaussian elements to eliminate redundant Gaussian elements;
[0011] Step S5: extracting multi-target states from the pruned and merged Gaussian elements to obtain an estimated set of multi-target states and a corresponding measurement set, and also obtaining a measurement set that is not associated with the target;
[0012] Step S6: obtaining the predicted strength of the new target according to the measurement set that is not associated with the target;
[0013] Step S7: Update the predicted strength of the new target in step S6.
[0014] In a specific embodiment, the step S1 includes:
[0015] Given the multi-target parameters before the current moment: Gaussian element of the surviving target at (k-1) moment and intensity D k-1 (y,0), the superscript j represents the Gaussian element index, and the letter p represents the survival target. represents the weight of the jth surviving target Gaussian element at time (k-1), represents the state of the jth surviving target Gaussian element at time (k-1), represents the state error variance matrix of the jth surviving target Gaussian element at time (k-1), represents the number of Gaussian elements of surviving targets at time (k-1); the set of Gaussian elements of newly born targets at time (k-1) With intensity D k-1 (y,1), the letter b indicates a new target, represents the weight of the jth new target Gaussian element at time (k-1), represents the state of the jth new target Gaussian element at time (k-1), represents the state error variance matrix of the jth new target Gaussian element at time (k-1), represents the number of new target Gaussian elements at time (k-1); until time (k-1), the unassociated measurement set And the corresponding number set
[0016] In a specific embodiment, step S2 includes:
[0017] The survival target Gaussian element used for prediction can be obtained by the following formula:
[0018]
[0019] here make is the predicted Gaussian element of the surviving target, then the predicted strength of the surviving target D k|k-1 (y,0) can be calculated using the following formula:
[0020]
[0021] Where: To predict Gaussian element weights; It means that the Gaussian random variable y follows a mean of The variance is The subscript k|k-1 indicates the prediction from time (k-1) to time k; the letter p indicates the survival target; the superscript j indicates the parameter of the jth Gaussian element;
[0022] The Gaussian mean and variance of the surviving targets at time k are predicted by the following formula:
[0023]
[0024]
[0025]
[0026] Where: is the predicted value of the state at time k; is the prediction state error variance matrix at time k, f(·) is the state transfer equation, F k-1 is the Jacobian matrix of f(·), Q k-1 is the process noise variance matrix; p S is the target survival probability.
[0027] In a specific embodiment, step S3 includes:
[0028] When there are new sensor observations, the mean and variance of the surviving target Gaussian element are updated, and the calculation formula is:
[0029]
[0030]
[0031]
[0032]
[0033]
[0034] Where: represents the predicted value of the observation at time k; represents the predicted value of the state at time k; h(·) represents the measurement equation, H krepresents the Jacobian matrix of h(·); z represents the observation; represents the target state after being updated by measuring z at time k; I represents the identity matrix; Represents the prediction state error variance matrix at time k; represents the state error variance matrix at time k; represents the filter gain at time k; represents the measurement prediction error variance matrix at time k; R k represents the measurement noise variance matrix at time k; the letter p represents the survival target; the superscript j represents the parameter of the jth Gaussian element;
[0035] After obtaining the mean and variance of the survival target Gaussian element, the updated survival target strength can be obtained, denoted as D k (y,0), the calculation formula is:
[0036]
[0037] in:
[0038]
[0039]
[0040] Where: is the Gaussian element weight corresponding to the measurement z at time k; p D represents the detection probability; Z k represents the set of measurements at time k; represents the prediction weight of Gaussian element; J k-1 Indicates the number of surviving target Gaussian elements; represents the likelihood of measuring z at time k, represents the predicted measure of the survival target at time k, represents the predicted state of the surviving target at time k, represents the measurement prediction error variance matrix at time k; L(z) represents the normalization coefficient; κ(z) represents the clutter density; is the a priori expected number of new targets, and the letter b represents the new target;
[0041] At time k, there is a set of measurements Here m k is the number of measurements, for each measurement Assign a unique label to obtain the measurement label set:
[0042]
[0043] Where L k is the measurement label set; is the label corresponding to the undetected part; is the measurement label corresponding to the corresponding measurement,
[0044] When updating, for each measured z k,i The updated predicted Gaussians are assigned corresponding measurement labels Then the updated Gaussian element becomes here: represents the updated Gaussian element weight at time k, represents the updated Gaussian mean at time k; represents the Gaussian error variance matrix after the update at time k; Represents the updated Gaussian meta-measurement label at time k.
[0045] In a specific embodiment, step S5 includes:
[0046] After pruning and merging, the multi-target state can be obtained by extracting the weight greater than a certain threshold ω Th The Gaussian method yields:
[0047]
[0048] Where: represents the mean value of the i-th Gaussian element at time k, ω Th represents the Gaussian element pruning threshold, J k represents the number of Gaussian elements after the update at time k, represents the Gaussian element weight at time k,
[0049] After state extraction, the state set of the target is Then the corresponding label set can be obtained as According to the measurement label The set of measurements associated with the state estimation can be obtained This completes the association between measurement and estimation, so the set of measurements that are not associated with estimates is Then the new target exists in the unrelated measurement set middle,
[0050] Once we have the k-time uncorrelated set of measurements Then the previous unrelated measurement set Compare and obtain a possible set of new targets.
[0051] In a specific embodiment, step S6 includes:
[0052] make is the predicted Gaussian element of the new target, then the predicted intensity of the new target D k|k-1(y,1) can be calculated by the following formula
[0053]
[0054] in:
[0055]
[0056]
[0057] Where: represents the predicted Gaussian element weight; represents the predicted Gaussian mean; represents the Gaussian element error variance matrix of the prediction; represents the jth unassociated measurement at time k; h(·) represents the measurement equation, H k represents the Jacobian matrix of h(·); h -1 (·) is the inverse function of the measurement equation; R k is the measurement noise variance; is the a priori expected number of newborn targets; represents the predicted number of new targets; the letter b represents a new target; the superscript j represents the parameter of the jth Gaussian element.
[0058] In a specific embodiment, the step S7 includes:
[0059] The strength of the new target is updated to
[0060]
[0061] in:
[0062]
[0063]
[0064]
[0065]
[0066] Where: represents the updated Gaussian element weight; represents the updated Gaussian mean; represents the updated Gaussian element error variance matrix; Indicates an unassociated measurement, represents a set of uncorrelated measurements; p D represents the probability of detection; represents the likelihood of measuring z at time k, represents the predicted measure of the survival target at time k, represents the predicted state of the target at time k; represents the normalization coefficient; Indicates clutter density; represents the number of new targets after update; the letter b represents the new target; the superscript j represents the parameter of the jth Gaussian element.
[0067] According to another aspect of the present invention, there is provided a multi-target tracking processing device for adaptively starting a new target, comprising:
[0068] A measurement input module for receiving measurements containing clutter;
[0069] A Gaussian mixture filter module, connected to the measurement input module, is used to filter the received measurements and new targets, obtain the Gaussian element set and strength of the predicted surviving targets, and output measurement labels and Gaussian element labels;
[0070] A track management module, connected to the Gaussian mixture filter module, for processing the Gaussian meta-tags to finally obtain a target track;
[0071] a new target initiation module, connected to the Gaussian mixture filter module, for obtaining the new target at the current moment according to the measurement tag, and outputting the new target to the Gaussian mixture filter module; and
[0072] The track output module is connected to the track management module, and is used to receive the target track and send the multi-target tracking results corresponding to the target track to the user.
[0073] In a specific embodiment, when the Gaussian mixture filter module processes the received measurements and new targets, targets not associated with the measurements are marked by the measurement tags in the update step and are used as new targets in the next time step.
[0074] In a specific embodiment, the Gaussian element tag is recursively transferred over time to determine whether the Gaussian elements at different times are from the same target; the measurement tag is used to mark the Gaussian element at the current moment to distinguish whether the target is from missed detection or measurement.
[0075] The method for multi-target tracking using adaptive new target initiation of the present invention firstly gives the multi-target parameters before the current moment, which include the Gaussian elements and intensities of surviving targets, the Gaussian elements and intensities of new targets, the measurement set not associated with the targets, and the corresponding measurement number set; at the current moment, the Gaussian elements and intensities of the surviving targets are predicted in one step to obtain the predicted Gaussian element set and intensities of the surviving targets; if there are new sensor observations at the current moment, the Gaussian elements and intensities of the surviving targets are updated to obtain the predicted Gaussian element set and intensities of the surviving targets, and the measurements and their corresponding Gaussian elements are assigned corresponding labels; the updated Gaussian elements are pruned and merged to eliminate redundant Gaussian elements; the pruned and merged Gaussian elements are subjected to multi-target state extraction to obtain the estimated set of multi-target states and the corresponding measurement set, and at the same time, the measurement set not associated with the targets is obtained; the predicted intensity of the new targets is obtained according to the measurement set not associated with the targets; the predicted intensity of the new targets in the previous step is updated.
[0076] Compared with the prior art, the multi-target tracking method of the present invention can estimate the number and state of targets at the same time when the strength of the new target is unknown. The target with no associated measurement is marked by the measurement tag in the update step and is used as a new target in the next time step. The unassociated measurement of the current step and the unassociated measurement in the previous time step are used to search for possible new targets through the measurement gate. After the search process, the number of measurements used to generate new targets is greatly reduced, thereby reducing unnecessary calculations and improving the accuracy of steady-state estimation. BRIEF DESCRIPTION OF THE DRAWINGS
[0077] After reading the specific embodiments of the present invention with reference to the accompanying drawings, readers will understand the various aspects of the present invention more clearly.
[0078] Figure 1 A flowchart of a multi-target tracking method for adaptive new target initiation according to one aspect of the present invention is shown.
[0079] Figure 2 A schematic structural diagram of a multi-target tracking processing device for adaptively initiating new targets according to another aspect of the present invention is shown.
[0080] Figure 3 Shows the use of Figure 1 Schematic diagram of the tracking scenario of the multi-target tracking method.
[0081] Figure 4 A schematic diagram showing the comparison between the data estimated values and the true values after 100 Monte Carlo simulations using the MEA-GM-EPHDF method and the GM-EPHDF method.
[0082] Figure 5 Show Figure 4 Schematic diagram of the comparison of OSPA distances of the MEA-GM-EPHDF method and the GM-EPHDF method after 100 Monte Carlo simulations.
[0083] Figure 6 Show Figure 4 Schematic diagram of the comparison of ICE between the MEA-GM-EPHDF method and the GM-EPHDF method under different clutter densities. DETAILED DESCRIPTION
[0084] In order to make the technical content disclosed in this application more detailed and complete, reference may be made to the accompanying drawings and the following various specific embodiments of the present invention, in which the same reference numerals represent the same or similar components. However, it should be understood by those skilled in the art that the embodiments provided below are not intended to limit the scope of the present invention. In addition, the accompanying drawings are only used for schematic illustration and are not drawn according to their original size.
[0085] The specific implementation modes of various aspects of the present invention are further described in detail below with reference to the accompanying drawings.
[0086] Figure 1 A flowchart of a multi-target tracking method for adaptive new target initiation according to one aspect of the present invention is shown.
[0087] Reference Figure 1 In this embodiment, the multi-target tracking method for adaptively starting a new target of the present application is implemented through steps S101 to S113. In detail, in step S101, the multi-target parameters before the current moment are given. Here, the multi-target parameters include the Gaussian element and intensity of the surviving target, the Gaussian element and intensity of the new target, the measurement set that is not associated with the target, and the corresponding measurement number set.
[0088] Then, in steps S103 and S105, the Gaussian element and intensity of the surviving target are predicted at the current moment to obtain the predicted Gaussian element set and intensity of the surviving target. If there is a new sensor observation value at the current moment, the Gaussian element and intensity of the surviving target in step S103 are updated to obtain the predicted Gaussian element set and intensity of the surviving target, and the measurement and its corresponding Gaussian element are assigned corresponding labels, that is, the measurement label and Gaussian element label are obtained.
[0089] Next, in steps S107 and S109, the updated Gaussian elements are pruned and merged to eliminate redundant Gaussian elements, and multi-target state extraction is performed on the pruned and merged Gaussian elements to obtain an estimated set of multi-target states and a corresponding measurement set, and at the same time, a measurement set not associated with a target is obtained.
[0090] Finally, in steps S111 and S113, the prediction strength of the new target is obtained according to the measurement set that is not associated with the target, and the prediction strength of the new target in the above step S111 is updated.
[0091] Furthermore, in step S101, the multi-target parameters before the current moment are given as: the Gaussian element of the surviving target at the (k-1) moment and intensity D k-1 (y,0), the superscript j represents the Gaussian element index, and the letter p represents the survival target. represents the weight of the jth surviving target Gaussian element at time (k-1), represents the state of the jth surviving target Gaussian element at time (k-1), represents the state error variance matrix of the jth surviving target Gaussian element at time (k-1), represents the number of Gaussian elements of surviving targets at time (k-1); the set of Gaussian elements of newly born targets at time (k-1) With intensity D k-1 (y,1), the letter b indicates a new target, represents the weight of the jth new target Gaussian element at time (k-1), represents the state of the jth new target Gaussian element at time (k-1), represents the state error variance matrix of the jth new target Gaussian element at time (k-1), represents the number of new target Gaussian elements at time (k-1); until time (k-1), the unassociated measurement set And the corresponding number set
[0092] Furthermore, in step S103, the Gaussian element of the survival target used for prediction can be obtained by the following formula:
[0093]
[0094] here make is the predicted Gaussian element of the surviving target, then the predicted strength of the surviving target D k|k-1 (y,0) can be calculated using the following formula:
[0095]
[0096] Where: To predict Gaussian element weights; It means that the Gaussian distribution random variable y follows a mean of The variance is The subscript k|k-1 indicates the prediction from time (k-1) to time k; the letter p indicates the survival target; the superscript j indicates the parameter of the jth Gaussian element;
[0097] The Gaussian mean and variance of the surviving targets at time k are predicted by the following formula:
[0098]
[0099]
[0100]
[0101] Where: is the predicted value of the state at time k; is the prediction state error variance matrix at time k, f(·) is the state transfer equation, F k-1 is the Jacobian matrix of f(·), Q k-1 is the process noise variance matrix; p S is the target survival probability.
[0102] According to a specific embodiment, in step S105, when there is a new sensor observation value, the mean and variance of the surviving target Gaussian element are updated, and the calculation formula is:
[0103]
[0104]
[0105]
[0106]
[0107]
[0108] Where: represents the predicted value of the observation at time k; represents the predicted value of the state at time k; h(·) represents the measurement equation, H k represents the Jacobian matrix of h(·); z represents the observation; represents the target state after being updated by measuring z at time k; I represents the identity matrix; Represents the prediction state error variance matrix at time k; represents the state error variance matrix at time k; represents the filter gain at time k; represents the measurement prediction error variance matrix at time k; R krepresents the measurement noise variance matrix at time k; the letter p represents the survival target; the superscript j represents the parameter of the jth Gaussian element;
[0109] After obtaining the mean and variance of the survival target Gaussian element, the updated survival target strength can be obtained, denoted as D k (y,0), the calculation formula is:
[0110]
[0111] in:
[0112]
[0113]
[0114] Where: is the Gaussian element weight corresponding to the measurement z at time k; p D represents the detection probability; Z k represents the set of measurements at time k; represents the prediction weight of Gaussian element; J k-1 Indicates the number of surviving target Gaussian elements; represents the likelihood of measuring z at time k, represents the predicted measure of the survival target at time k, represents the predicted state of the surviving target at time k, represents the measurement prediction error variance matrix at time k; L(z) represents the normalization coefficient; κ(z) represents the clutter density; is the a priori expected number of new targets, and the letter b represents the new target;
[0115] At time k, there is a set of measurements Here m k is the number of measurements, for each measurement Assign a unique label to obtain the measurement label set:
[0116]
[0117] Where L k is the measurement label set; is the label corresponding to the undetected part; is the measurement label corresponding to the corresponding measurement,
[0118] When updating, for each measured z k,i The updated predicted Gaussians are assigned corresponding measurement labels Then the updated Gaussian element becomes here: represents the updated Gaussian element weight at time k, represents the updated Gaussian mean at time k; represents the Gaussian error variance matrix after the update at time k; Represents the updated Gaussian meta-measurement label at time k.
[0119] According to a specific embodiment, in step S109, after pruning and merging, the multi-target state can be obtained by extracting the weight greater than a certain threshold ω Th The Gaussian method yields:
[0120]
[0121] Where: represents the mean value of the i-th Gaussian element at time k, ω Th represents the Gaussian element pruning threshold, J k represents the number of Gaussian elements after the update at time k, represents the Gaussian element weight at time k,
[0122] After state extraction, the state set of the target is Then the corresponding label set can be obtained as According to the measurement label The set of measurements associated with the state estimation can be obtained This completes the association between measurement and estimation, so the set of measurements that are not associated with estimates is Then the new target exists in the unrelated measurement set middle,
[0123] Once we have the k-time uncorrelated set of measurements Then the previous unrelated measurement set Compare and obtain a possible set of new targets.
[0124] According to a specific embodiment, in step S111,
[0125] make is the predicted Gaussian element of the new target, then the predicted intensity of the new target D k|k-1 (y,1) can be calculated by the following formula
[0126]
[0127] in:
[0128]
[0129]
[0130] Where: represents the predicted Gaussian element weight; represents the predicted Gaussian mean; represents the Gaussian element error variance matrix of the prediction; represents the jth unassociated measurement at time k; h(·) represents the measurement equation, H k represents the Jacobian matrix of h(·); h -1 (·) is the inverse function of the measurement equation; R k is the measurement noise variance; is the a priori expected number of newborn targets; represents the predicted number of new targets; the letter b represents a new target; the superscript j represents the parameter of the jth Gaussian element.
[0131] According to a specific embodiment, in step S113, the strength of the new target is updated to
[0132]
[0133] in:
[0134]
[0135]
[0136]
[0137]
[0138] Where: represents the updated Gaussian element weight; represents the updated Gaussian mean; represents the updated Gaussian element error variance matrix; Indicates an unassociated measurement, represents a set of uncorrelated measurements; p D represents the probability of detection; represents the likelihood of measuring z at time k, represents the predicted measure of the survival target at time k, represents the predicted state of the target at time k; represents the normalization coefficient; Indicates clutter density; represents the number of new targets after update; the letter b represents the new target; the superscript j represents the parameter of the jth Gaussian element.
[0139] From the above, it can be seen that the multi-target tracking method of the present invention can estimate the number and state of targets at the same time when the strength of the new target is unknown. The target with no associated measurement is marked by the measurement tag in the update step and is used as a new target in the next time step. The unassociated measurement of the current step and the unassociated measurement in the previous time step are used to search for possible new targets through the measurement gate. After the search process, the number of measurements used to generate new targets is greatly reduced, thereby reducing unnecessary calculations and improving the accuracy of steady-state estimation.
[0140] Figure 2 A schematic structural diagram of a multi-target tracking processing device for adaptively initiating new targets according to another aspect of the present invention is shown.
[0141] Reference Figure 2 In this embodiment, the multi-target tracking processing device for adaptive new target initiation of the present application includes a measurement input module, a Gaussian mixture filter module, a track management module, a track output module and a new target initiation module.
[0142] Specifically, the measurement input module is used to receive measurements containing clutter. The Gaussian mixture filter module is connected to the measurement input module, and is used to filter the received measurements and new targets to obtain the Gaussian element set and intensity of the predicted surviving targets, and output the measurement label and Gaussian element label. The track management module is connected to the Gaussian mixture filter module, and is used to process the Gaussian element label to finally obtain the target track. The new target starting module is connected to the Gaussian mixture filter module, and is used to obtain the new target at the current moment according to the measurement label, and output the new target to the Gaussian mixture filter module. The track output module is connected to the track management module, and is used to receive the target track, and deliver the multi-target tracking results corresponding to the target track to the user.
[0143] As can be seen from the above, the multi-target tracking processing device of the present application is provided with a Gaussian mixture filter module, which uses Gaussian element labels and measurement labels to mark Gaussian elements at the same time in the update step when updating the Gaussian element and intensity of the surviving target, thereby improving the calculation efficiency and accuracy. Since the Gaussian element label is recursively and transmitted over time, it is determined whether the Gaussian elements at different times come from the same target (track); and because the measurement label is not recursively and transmitted over time, it only plays a marking role for the Gaussian element at the current moment, thereby distinguishing whether the target comes from missed detection or measurement. In this way, the method can estimate the number and state of targets at the same time when the intensity of the new target is unknown. The target that is not associated with the measurement is marked with the measurement label in the update step and is used as a new target in the next time step. The measurement in the current step and the previous time step is used to search for possible new targets through the measurement gate.
[0144] Figure 3 Shows the use of Figure 1 Schematic diagram of the tracking scenario of the multi-target tracking method. Figure 4 A schematic diagram showing the comparison between the data estimated values and the true values after 100 Monte Carlo simulations using the MEA-GM-EPHDF method and the GM-EPHDF method. Figure 5 Show Figure 4 Schematic diagram of the comparison of OSPA distances of the MEA-GM-EPHDF method and the GM-EPHDF method after 100 Monte Carlo simulations. Figure 6 Show Figure 4 Schematic diagram of the comparison of ICE between the MEA-GM-EPHDF method and the GM-EPHDF method under different clutter densities.
[0145] like Figure 3 As shown, consider a two-dimensional plane tracking problem, assuming that the state of the target is This vector contains the position of the target (p x,k ,p y,k ),speed and acceleration The target motion adopts the constant velocity model (CV). The sampling time is T = 1s, and the target survival probability p S =0.98. The target's movement is Figure 3 As shown in the figure (“o” indicates the starting point of the target), there are 8 targets in total. Assume that the strength of the first target is known, and the strengths of the other targets are unknown.
[0146] Assuming that the radar observations are slant range r and azimuth θ, the observation equation at time k is:
[0147]
[0148] In the formula, ε k is the observation noise and satisfies ε k N(0,R), where R = diag[σ r ,σ θ ],σ r =100m,σ θ =0.1°. Sensor position (x s ,y s ) at the coordinate origin.
[0149] The random finite set of clutter in the observation follows a Poisson distribution with probability density κ k (z) = λ c Vu(z). Where λ c =2.1×10 -4 (radm) -1is the number of clutter in a unit volume (that is, there are 20 clutters in the entire tracking area); the monitoring area is [-π / 2,π / 2]rad×[0,30000]m, V is the area of the observation area; u(·) is the probability density function of the clutter points in the detection area that obey a uniform distribution. A priori expectation number of new targets and here The measurement gate γ=20.
[0150] This example uses Matlab language to test the proposed method and compares two methods: the traditional adaptive new target starting multi-target probability hypothesis density filter (GM-EPHDF) and the efficient adaptive new target starting multi-target probability hypothesis density filter (MEA-GM-EPHDF) proposed in this application. The performance of the above two algorithms will be evaluated in the following aspects: (1) potential estimation, true value and estimated value; (2) comparison of OPSA distance; (3) comparison of computational efficiency.
[0151] Figure 4 The true value of the potential estimate and the estimated mean value after 100 Monte Carlo simulations are given. It can be seen that both methods can estimate the target number. In addition, Figure 4 It can also be seen that the estimation accuracy of MEA-GM-EPHDF in steady state is higher than that of GM-EPHDF, while GM-EPHDF responds faster when new targets are generated. This is because MEA-GM-EPHDF uses delayed logic and uses the measurements of the current step and the previous step to make judgments, while only using possible new targets for recursion will improve the accuracy. In addition, it can be seen that the overall estimation accuracy and steady-state estimation accuracy of MEA-GM-EPHDF are better than those of GM-EPHDF. Therefore, the estimation effect of the MEA-GM-EPHDF algorithm on the number of targets is better than that of GM-EPHDF.
[0152] Figure 5 The OSPA distance is given after 100 Monte Carlo simulations. The OSPA distance can simultaneously evaluate the number and element differences of two random sets. Figure 5It can be seen that when a new target is generated, the OSPA distances of both algorithms have a spike. This is because neither method can respond to the new target in real time. MEA-GM-EPHDF has to delay one or two steps. When a new target is generated, the OSPA distance of MEA-GM-EPHDF is greater than that of GM-EPHDF, while the opposite is true in steady state, that is, the OSPA distance of MEA-GM-EPHDF is smaller than that of GM-EPHDF. In addition, it can be seen that the overall OSPA distance of MEA-GM-EPHDF is greater than that of GM-EPHDF, while the steady-state OSPA distance is better than that of GM-EPHDF, because GM-EPHDF always responds faster to new targets than MEA-GM-EPHDF. However, the estimation accuracy of MEA-GM-EPHDF is better than that of GM-EPHDF.
[0153] In order to compare the computational efficiency of the algorithms, the following computational efficiency improvement is defined:
[0154]
[0155] in is the average computation time of the GM-EPHDF algorithm, is the average computation time of the MEA-GM-EPHDF algorithm.
[0156] Figure 6 The ICE (100 Monte Carlo simulations) of the MEA-GM-EPHDF algorithm relative to the GM-EPHDF under different clutter densities (average number of clutter per scanning cycle) is given. It can be seen from the figure that when the clutter density is 5, the ICE is 50%. As the clutter density increases, the ICE also increases. When the clutter density reaches more than 35, the ICE can reach 90%. It can be seen that under different clutter density conditions, the computational efficiency of the MEA-GM-EPHDF algorithm is higher than that of the GM-EPHDF method, and the efficiency is improved as the clutter density increases. This is because after the MEA-GM-EPHDF algorithm adopts measurement-estimation association and measurement wave gate screening, the measurement used to generate new targets is greatly reduced, so the computational efficiency is improved.
[0157] The method for multi-target tracking using adaptive new target initiation of the present invention firstly gives the multi-target parameters before the current moment, which include the Gaussian elements and intensities of surviving targets, the Gaussian elements and intensities of new targets, the measurement set not associated with the targets, and the corresponding measurement number set; at the current moment, the Gaussian elements and intensities of the surviving targets are predicted in one step to obtain the predicted Gaussian element set and intensities of the surviving targets; if there are new sensor observations at the current moment, the Gaussian elements and intensities of the surviving targets are updated to obtain the predicted Gaussian element set and intensities of the surviving targets, and the measurements and their corresponding Gaussian elements are assigned corresponding labels; the updated Gaussian elements are pruned and merged to eliminate redundant Gaussian elements; the pruned and merged Gaussian elements are subjected to multi-target state extraction to obtain the estimated set of multi-target states and the corresponding measurement set, and at the same time, the measurement set not associated with the targets is obtained; the predicted intensity of the new targets is obtained according to the measurement set not associated with the targets; the predicted intensity of the new targets in the previous step is updated.
[0158] Compared with the prior art, the multi-target tracking method of the present invention can estimate the number and state of targets at the same time when the strength of the new target is unknown. The target with no associated measurement is marked by the measurement tag in the update step and is used as a new target in the next time step. The unassociated measurement of the current step and the unassociated measurement in the previous time step are used to search for possible new targets through the measurement gate. After the search process, the number of measurements used to generate new targets is greatly reduced, thereby reducing unnecessary calculations and improving the accuracy of steady-state estimation.
[0159] In the above, the specific embodiments of the present invention are described with reference to the accompanying drawings. However, those skilled in the art will appreciate that various changes and substitutions may be made to the specific embodiments of the present invention without departing from the spirit and scope of the present invention. These changes and substitutions are all within the scope defined by the claims of the present invention.
Claims
1. A multi-target tracking method for adaptive new target initiation, characterized in that: The multi-target tracking method comprises the following steps: Step S1: Given multi-target parameters before the current moment, the multi-target parameters include Gaussian elements and intensities of surviving targets, Gaussian elements and intensities of newborn targets, a measurement set not associated with the target, and a corresponding measurement number set; Step S2: performing a one-step prediction on the Gaussian element and intensity of the surviving target at the current moment to obtain the predicted Gaussian element set and intensity of the surviving target; Step S3: If there is a new sensor observation value at the current moment, the Gaussian element and intensity of the surviving target in step S2 are updated to obtain the predicted Gaussian element set and intensity of the surviving target, and the measurement and its corresponding Gaussian element are assigned corresponding labels; Step S4: Prune and merge the updated Gaussian elements to eliminate redundant Gaussian elements; Step S5: extracting multi-target states from the pruned and merged Gaussian elements to obtain an estimated set of multi-target states and a corresponding measurement set, and also obtaining a measurement set that is not associated with the target; Step S6: obtaining the predicted strength of the new target according to the measurement set that is not associated with the target; Step S7: updating the predicted strength of the new target in step S6; Wherein the step S1 further comprises: Given the multi-target parameters before the current moment: Gaussian element of the surviving target at (k-1) moment and intensity D k-1 (y,0), the superscript j represents the Gaussian element index, and the letter p represents the survival target. represents the weight of the jth surviving target Gaussian element at time (k-1), represents the state of the jth surviving target Gaussian element at time (k-1), represents the state error variance matrix of the jth surviving target Gaussian element at time (k-1), represents the number of Gaussian elements of surviving targets at time (k-1); the set of Gaussian elements of newly born targets at time (k-1) With intensity D k-1 (y,1), the letter b indicates a new target, represents the weight of the jth new target Gaussian element at time (k-1), represents the state of the jth new target Gaussian element at time (k-1), represents the state error variance matrix of the jth new target Gaussian element at time (k-1), represents the number of new target Gaussian elements at time (k-1); until time (k-1), the unassociated measurement set And the corresponding number set Wherein the step S2 further comprises: The survival target Gaussian element used for prediction is obtained by the following formula: here make is the predicted Gaussian element of the surviving target, then the predicted strength of the surviving target D k|k-1 (y,0) is calculated using the following formula: Where: To predict Gaussian element weights; It means that the Gaussian random variable y follows a mean of The variance is The subscript k|k-1 indicates the prediction from time (k-1) to time k; the letter p indicates the survival target; the superscript j indicates the parameter of the jth Gaussian element; The Gaussian mean and variance of the surviving targets at time k are predicted by the following formula: Where: is the predicted value of the state at time k; is the prediction state error variance matrix at time k, f(·) is the state transfer equation, F k-1 is the Jacobian matrix of f(·), Q k-1 is the process noise variance matrix; p S is the target survival probability; Wherein the step S3 further comprises: When there are new sensor observations, the mean and variance of the surviving target Gaussian element are updated, and the calculation formula is: Where: represents the predicted value of observation at time k; represents the predicted value of the state at time k; h(·) represents the measurement equation, H k represents the Jacobian matrix of h(·); z represents the observation; represents the target state after being updated by measuring z at time k; I represents the identity matrix; Represents the prediction state error variance matrix at time k; represents the state error variance matrix at time k; represents the filter gain at time k; represents the measurement prediction error variance matrix at time k; R k represents the measurement noise variance matrix at time k; the letter p represents the survival target; the superscript j represents the parameter of the jth Gaussian element; After obtaining the mean and variance of the survival target Gaussian element, the updated survival target strength can be obtained, denoted as D k (y,0), the calculation formula is: in: Where: is the Gaussian element weight corresponding to the measurement z at time k; p D represents the detection probability; Z k represents the set of measurements at time k; represents the prediction weight of Gaussian element; J k-1 Indicates the number of surviving target Gaussian elements; represents the likelihood of measuring z at time k, represents the predicted measure of the survival target at time k, represents the predicted state of the surviving target at time k, represents the measurement prediction error variance matrix at time k; L(z) represents the normalization coefficient; κ(z) represents the clutter density; is the a priori expected number of new targets, and the letter b represents the new target; At time k, there is a set of measurements Here m k is the number of measurements, for each measurement Assign a unique label to obtain the measurement label set: Where L k is the measurement label set; is the label corresponding to the undetected part; is the measurement label corresponding to the corresponding measurement, When updating, for each measured z k,i The updated predicted Gaussians are assigned corresponding measurement labels Then the updated Gaussian element becomes here: represents the updated Gaussian element weight at time k, represents the updated Gaussian mean at time k; represents the Gaussian error variance matrix after the update at time k; represents the updated Gaussian element measurement label at time k; Wherein the step S5 further comprises: After pruning and merging, the multi-target state is obtained by extracting the weight greater than a certain threshold ω Th The Gaussian method yields: Where: represents the mean value of the i-th Gaussian element at time k, ω Th represents the Gaussian element pruning threshold, J k represents the number of Gaussian elements after the update at time k, represents the Gaussian element weight at time k, After state extraction, the state set of the target is Then the corresponding label set can be obtained as According to the measurement label Get the set of measurements associated with the state estimate This completes the association between measurement and estimation, so the set of measurements that are not associated with estimates is Then the new target exists in the unrelated measurement set middle, Once we have the k-time uncorrelated set of measurements Then the previous unrelated measurement set Compare and obtain a possible new target set; Wherein the step S6 further comprises: make is the predicted Gaussian element of the new target, then the predicted intensity of the new target D k|k-1 (y,1) is calculated by the following formula in: Where: represents the predicted Gaussian element weight; represents the predicted Gaussian mean; represents the Gaussian element error variance matrix of the prediction; represents the jth unassociated measurement at time k; h(·) represents the measurement equation, H k represents the Jacobian matrix of h(·); h -1 (·) is the inverse function of the measurement equation; R k is the measurement noise variance; is the a priori expected number of newborn targets; represents the predicted number of new targets; the letter b represents a new target; the superscript j represents the parameter of the jth Gaussian element; Wherein the step S7 further comprises: The strength of the new target is updated to in: Where: represents the updated Gaussian element weight; represents the updated Gaussian mean; represents the updated Gaussian element error variance matrix; Indicates an unassociated measurement, represents a set of uncorrelated measurements; p D represents the probability of detection; represents the likelihood of measuring z at time k, represents the predicted measure of the survival target at time k, represents the predicted state of the target at time k; represents the normalization coefficient; Indicates clutter density; represents the number of new targets after update; the letter b represents the new target; the superscript j represents the parameter of the jth Gaussian element.
2. A multi-target tracking processing device using the multi-target tracking method of adaptive new target initiation as claimed in claim 1, characterized in that: The multi-target tracking processing device comprises: A measurement input module for receiving measurements containing clutter; A Gaussian mixture filter module, connected to the measurement input module, is used to filter the received measurements and new targets, obtain the Gaussian element set and strength of the predicted surviving targets, and output measurement labels and Gaussian element labels; A track management module, connected to the Gaussian mixture filter module, for processing the Gaussian meta-tags to finally obtain a target track; a new target initiation module, connected to the Gaussian mixture filter module, for obtaining the new target at the current moment according to the measurement tag, and outputting the new target to the Gaussian mixture filter module; and The track output module is connected to the track management module, and is used to receive the target track and send the multi-target tracking results corresponding to the target track to the user.
3. The multi-target tracking processing device according to claim 2, characterized in that: When the Gaussian mixture filter module processes the received measurements and new targets, targets not associated with the measurements are marked by the measurement tags in the update step and are used as new targets in the next time step.
4. The multi-target tracking processing device according to claim 2, characterized in that: The Gaussian element label is recursively transferred over time to determine whether the Gaussian elements at different times come from the same target; the measurement label is used to mark the Gaussian element at the current moment to distinguish whether the target comes from missed detection or measurement.
