Partial interaction multi-model tracking method and device based on Gaussian mixture initialization

Abstract

The application provides a partial interaction multi-model tracking method and device based on Gaussian mixture initialization, each maneuvering mode is described by multiple maneuvering models with the same structure and allowing slight changes of maneuvering parameters, the multiple maneuvering models with the same structure have one-to-one corresponding maneuvering parameter partitions, the maneuvering parameters of each model are initialized by Gaussian distribution of the corresponding maneuvering parameter partition, after the initialization of the maneuvering parameters, the models with the same structure are independently operated, and the models with different structures can interact to cope with possible mode jumps; the scheme can eliminate the mismatching problem of the maneuvering models and the deviation of the maneuvering parameters; in addition, the mechanism of partial model interaction instead of interaction of all models protects the best initialized filter from the interference of other error initialized filters, and can improve the estimation accuracy of the maneuvering mode dwell section. Therefore, the application can effectively improve the tracking accuracy of the maneuvering target.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of maneuvering target tracking, and in particular to a partially interacting multi-model tracking method and device based on Gaussian mixture initialization. Background Art

[0002] Maneuvering target tracking is the process of estimating the state of a maneuvering target based on sensor information. Due to the complex and ever-changing motion patterns of maneuvering targets, single-model approaches often struggle to meet the required tracking accuracy. In contrast, multi-model approaches demonstrate superior tracking performance. Multi-model approaches simultaneously estimate the state of a maneuvering target using multiple models that potentially match the target's true motion pattern.

[0003] To achieve high tracking accuracy, multi-model methods must ensure that a model in the model set matches the target's true motion pattern. This matching requires that the model types and maneuvering parameters are consistent. However, existing multi-model methods cannot effectively address both the uncertainty of model types and the uncertainty of maneuvering parameters, which affects the tracking accuracy of maneuvering targets.

[0004] Therefore, there is an urgent need to provide a multi-model maneuvering target tracking method that can simultaneously solve the uncertainty of model type and maneuvering parameters. Summary of the Invention

[0005] The embodiments of the present invention provide a partially interacting multi-model tracking method and device based on Gaussian mixture initialization, which can simultaneously solve the uncertainty of model types and the uncertainty of maneuvering parameters, and improve the tracking accuracy of maneuvering targets.

[0006] In a first aspect, an embodiment of the present invention provides a multi-model maneuvering target tracking method, comprising:

[0007] Step 100: Based on possible motion forms of the maneuvering target, motion models of corresponding motion forms are constructed to obtain a motion model set; wherein the motion forms include non-maneuvering motion modes and multiple maneuvering modes. The same maneuvering mode is described by multiple maneuvering models with the same structure, and the multiple maneuvering models with the same structure have one-to-one corresponding maneuvering parameter partitions; the models used to describe the same motion form constitute a model structure;

[0008] Step 102: Initialize the filtering for each model and each model structure in the motion model set, and after the filtering is initialized, start recursive filtering for each model;

[0009] Step 104: At each filtering moment, for each model, the following steps are performed: using a Gaussian mixture distribution to initialize the maneuver parameter state components of other model structures having different structures from the model, mixing the overall state estimate of the other model structures with the state estimate of the model at the previous moment to obtain a mixed state estimate; obtaining the current moment measurement, and using the current moment measurement and the mixed state estimate as input to the model, and calculating the conditional state estimate and likelihood function of the model at the current moment after the input;

[0010] Step 106: Calculate the conditional state estimate of each model structure based on the conditional state estimate and likelihood function of each model at the current moment, and fuse the conditional state estimate of each model structure at the current moment to obtain a fused state estimate; the fused state estimate is the state estimate result of the maneuvering target at the current moment;

[0011] Step 108: Take the current moment as the previous moment and continue to execute step 104 until the maneuvering target tracking is completed.

[0012] In a second aspect, an embodiment of the present invention further provides a partially interacting multi-model tracking device based on Gaussian mixture initialization, comprising:

[0013] a construction unit configured to construct motion models of corresponding motion forms according to possible motion forms of a maneuvering target, thereby obtaining a motion model set; wherein the motion forms include non-maneuvering motion modes and multiple maneuvering modes, wherein the same maneuvering mode is described by multiple maneuvering models having the same structure, and the multiple maneuvering models having the same structure have one-to-one corresponding maneuvering parameter partitions; and wherein the models describing the same motion form constitute a model structure;

[0014] an initialization unit, configured to perform filtering initialization on each model and each model structure in the motion model set, and after filtering initialization, each model starts recursive filtering;

[0015] The first computing unit is configured to, at each filtering moment, execute, for each model, the following: initializing, using a Gaussian mixture distribution, a maneuver parameter state component of another model structure having a different structure from the model, mixing an overall state estimate of the other model structure with a state estimate of the model at a previous moment to obtain a mixed state estimate; obtaining a current moment measurement, using the current moment measurement and the mixed state estimate as inputs to the model, and calculating a conditional state estimate and a likelihood function of the model at the current moment after the input;

[0016] The second computing unit is used to calculate the conditional state estimate of each model structure based on the conditional state estimate and likelihood function of each model at the current moment, and to fuse the conditional state estimate of each model structure at the current moment to obtain a fused state estimate; the fused state estimate is the state estimation result of the maneuvering target at the current moment; the current moment is taken as the previous moment, and the first computing unit is triggered to continue to perform the corresponding operation until the maneuvering target tracking is completed.

[0017] In a third aspect, an embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method described in any embodiment of this specification is implemented.

[0018] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, enables the computer to execute the method described in any embodiment of this specification.

[0019] An embodiment of the present invention provides a partially interacting multi-model tracking method and apparatus based on Gaussian mixture initialization. Each maneuvering mode is jointly described by multiple maneuvering models with the same structure that allow for slight changes in maneuvering parameters. The multiple maneuvering models with the same structure have one-to-one corresponding maneuvering parameter partitions. The maneuvering parameters of each model are initialized using the Gaussian distribution of the corresponding maneuvering parameter partition. After the maneuvering parameters are initialized, the models with the same structure operate independently, while models with different structures can interact to address possible mode jumps. This solution can eliminate both the problem of maneuvering model mismatch and the deviation of maneuvering parameters. In addition, the mechanism of partial model interaction rather than the interaction of all models protects the optimally initialized filter from interference from other incorrectly initialized filters, thereby improving the estimation accuracy of the maneuvering mode's lingering segments. Therefore, the present invention can effectively improve the tracking accuracy of maneuvering targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a flow chart of a partially interacting multi-model tracking method based on Gaussian mixture initialization provided by one embodiment of the present invention;

[0022] Figure 21 is a schematic diagram of the average Euclidean error of position provided by an embodiment of the present invention;

[0023] Figure 3 1 is a schematic diagram of the average Euclidean error of velocity provided by an embodiment of the present invention;

[0024] Figure 4 1 is a schematic diagram of the average Euclidean error of acceleration provided by an embodiment of the present invention;

[0025] Figure 5 This is a hardware architecture diagram of an electronic device provided by one embodiment of the present invention;

[0026] Figure 6 This is a structural diagram of a partially interacting multi-model tracking device based on Gaussian mixture initialization provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] As mentioned above, the existing multi-model method cannot effectively solve the uncertainty of model type and maneuvering parameters at the same time, thereby affecting the tracking accuracy of maneuvering targets. The inventive concept of the present invention is:

[0029] First, considering that actual maneuvering parameters typically do not remain constant but rather undergo slight variations, a model that considers maneuvering parameters as state components is adopted to mitigate model type mismatch. Second, for states with different maneuvering parameters, a strategy is proposed to initialize maneuvering parameters using a Gaussian mixture distribution. Third, multiple maneuvering models with the same structure are employed to jointly describe the corresponding maneuvering pattern, with each model initializing its maneuvering parameters using a corresponding Gaussian distribution. Each model is responsible for initializing a specific maneuvering parameter partition to mitigate model parameter bias. Given that the optimal parameter initialization after a maneuvering pattern takes effect is unique and time-invariant, models with the same structure should operate independently. Therefore, a partial model interaction mechanism is proposed. Models with the same structure operate independently after parameter initialization without interfering with each other, while models with different structures can interact to address potential model jumps. This partial interaction mechanism, rather than full model interaction, protects the optimally initialized filter from interference from other incorrect initializations.

[0030] The specific implementation of the above concept is described below.

[0031] Please refer to Figure 1 , an embodiment of the present invention provides a partial interaction multi-model tracking method based on Gaussian mixture initialization, the method comprising:

[0032] Step 100: Based on possible motion forms of the maneuvering target, motion models of corresponding motion forms are constructed to obtain a motion model set; wherein the motion forms include non-maneuvering motion modes and multiple maneuvering modes. The same maneuvering mode is described by multiple maneuvering models with the same structure, and the multiple maneuvering models with the same structure have one-to-one corresponding maneuvering parameter partitions; the models used to describe the same motion form constitute a model structure;

[0033] Step 102: Initialize the filtering for each model and each model structure in the motion model set, and after the filtering is initialized, start recursive filtering for each model;

[0034] Step 104: At each filtering moment, for each model, the following steps are performed: using a Gaussian mixture distribution to initialize the maneuver parameter state components of other model structures having different structures from the model, mixing the overall state estimate of the other model structures with the state estimate of the model at the previous moment to obtain a mixed state estimate; obtaining the current moment measurement, and using the current moment measurement and the mixed state estimate as input to the model, and calculating the conditional state estimate and likelihood function of the model at the current moment after the input;

[0035] Step 106: Calculate the conditional state estimate of each model structure based on the conditional state estimate and likelihood function of each model at the current moment, and fuse the conditional state estimate of each model structure at the current moment to obtain a fused state estimate; the fused state estimate is the state estimate result of the maneuvering target at the current moment;

[0036] Step 108: Take the current moment as the previous moment and continue to execute step 104 until the maneuvering target tracking is completed.

[0037] In an embodiment of the present invention, each maneuvering mode is described by multiple maneuvering models with the same structure that allow for slight variations in maneuvering parameters. These multiple maneuvering models with the same structure have one-to-one corresponding maneuvering parameter partitions. The maneuvering parameters of each model are initialized using the Gaussian distribution of the corresponding maneuvering parameter partition. After the maneuvering parameters are initialized, the models with the same structure operate independently, while models with different structures can interact to address possible mode jumps. This solution can eliminate both the problem of maneuvering model mismatch and the deviation of maneuvering parameters. In addition, the mechanism of partial model interaction rather than full model interaction protects the optimally initialized filter from interference from other incorrectly initialized filters, thereby improving the estimation accuracy of the maneuvering mode's lingering segments. Therefore, the present invention can effectively improve the tracking accuracy of maneuvering targets.

[0038] Described below Figure 1 How to perform the steps shown.

[0039] First, with respect to step 100 , according to possible motion forms of the maneuvering target, motion models of corresponding motion forms are constructed to obtain a motion model set.

[0040] In the embodiment of the present invention, different maneuvering targets may have different possible motion forms. Therefore, the possible motion forms of the maneuvering targets may be analyzed to construct motion models of corresponding motion forms, and the tracking of the maneuvering targets may be achieved through multiple motion models.

[0041] The motion forms can include non-maneuverable motion modes and multiple maneuverable motion modes. Non-maneuverable motion modes can be near-constant motion, while maneuverable motion modes can include acceleration, turning, deceleration, diving, and turning while decelerating. Considering that a maneuverable target can have more than one motion form and can be any combination of the above motion forms, the constructed motion model set can include multiple motion models.

[0042] In the embodiment of the present invention, when constructing a non-maneuvering model for a non-maneuvering motion mode, since the non-maneuvering motion mode is a nearly uniform motion, its state vector does not involve maneuvering parameters. Therefore, if the possible motion forms of the maneuvering target include the non-maneuvering motion mode, a non-maneuvering model is constructed for the non-maneuvering motion mode.

[0043] When constructing a maneuver model for a maneuver mode, since there is uncertainty in the size of its maneuver parameters when the maneuver mode becomes effective, if the possible motion forms of the maneuvering target include at least one maneuver mode, the following steps are performed for each maneuver mode: construct multiple maneuver models with the same structure for describing the maneuver mode, determine the maneuver parameter range based on the maneuver parameters corresponding to the maneuver mode, divide the maneuver parameter range equally into multiple maneuver parameter partitions, and the number of partitions is equal to the number of models of the multiple maneuver models constructed; and assign the multiple maneuver parameter partitions to the multiple maneuver models in a one-to-one correspondence.

[0044] This embodiment of the present invention considers the evolution of maneuvering parameters in actual tracking applications and uses multiple maneuvering models with identical structures that allow for small variations in maneuvering parameters to jointly describe a maneuvering pattern, effectively eliminating model mismatch. Furthermore, multiple Gaussian distributions with quantized means are used to initialize the maneuvering parameters. Each Gaussian distribution is responsible for initializing a parameter partition, and the variance of the corresponding partition is used as the initialization variance to ensure consistency in parameter initialization.

[0045] Models describing the same motion form form a single model structure. That is, if a single non-maneuvering model is constructed for a non-maneuvering motion mode, then that single non-maneuvering model forms a single model structure. Similarly, if 10 near-uniform acceleration maneuvering models are constructed for an acceleration maneuvering mode, then those 10 near-uniform acceleration maneuvering models form a single model structure.

[0046] It should be noted that when determining the number of models required to describe a maneuver pattern, it is understood that a greater number of models results in higher maneuvering target tracking accuracy, but also increases computational complexity. Therefore, the appropriate number of models can be determined based on acceptable accuracy. Furthermore, the width of the maneuver parameter range corresponding to the maneuver pattern can also be considered when determining the number of models.

[0047] The following describes this step by taking the movement forms of the maneuvering target, including the nearly uniform speed movement form (non-maneuvering mode) and the uniform acceleration movement form (maneuvering mode), as an example.

[0048] A two-dimensional near-uniform velocity model (NCV) can be constructed for near-uniform velocity motion, and multiple two-dimensional near-uniform acceleration models (NCA) with the same structure can be constructed for uniform acceleration motion.

[0049] The two-dimensional nearly uniform motion model is denoted as m 1 , its dynamic equation is:

[0050]

[0051] in, Indicates that the position x in the x direction at time kk , y-direction position y k , x-direction velocity y-direction speed The state vector of the NCV model is composed of is the state transition matrix of the NCV model, is the process noise gain matrix of the NCV model, To obey zero mean, the covariance is Gaussian white noise. State transfer matrix and the process noise gain matrix Given as follows:

[0052]

[0053] Where T is the data sampling interval.

[0054] The two-dimensional nearly uniform acceleration motion model is denoted as m 2 , its dynamic equation is:

[0055]

[0056] in, Indicates that the position x in the x direction at time k k , y-direction position y k , x-direction velocity y-direction speed x-direction acceleration y-direction acceleration The state vector of the NCA model is composed of is the state transition matrix of the NCA model, is the process noise gain matrix of the NCA model, To obey zero mean, the covariance is Gaussian white noise. State transfer matrix and the process noise gain matrix Given as follows:

[0057]

[0058] Assuming that the number of NCA models required to describe the uniform acceleration of a maneuvering target is 9, it is necessary to assign an acceleration partition to each NCA model. Specifically, the lower bound of the acceleration range in the x direction is denoted as a x , the upper bound of the x direction is denoted as b x , the lower bound of the y direction is denoted as a y , the upper bound of the y direction is denoted as b y , the βth NCA model is recorded as m 2,β(β∈{1,...,9}). Assuming that the acceleration in the x-direction and the y-direction can be decoupled, the two-dimensional acceleration area is divided into three equal parts in the x-direction and the y-direction, respectively, to obtain nine acceleration partitions, each of which corresponds to an NCA model. The mean value of the uniform distribution within the acceleration partition is:

[0059]

[0060]

[0061] in, and are the components of the mean value uniformly distributed in the βth acceleration partition in the x-direction and y-direction, n x and n y are the number of partitions in the x direction and the y direction respectively. In this embodiment, n x =n y = 3. The variances of the uniform distribution in the acceleration partition in the x and y directions are:

[0062]

[0063]

[0064] Among them, the mean and variance within the acceleration partition are used to initialize the acceleration state component of the corresponding NCA model.

[0065] It should be noted that the maneuver parameters to be divided are related to the corresponding maneuver model constructed. In this embodiment of the present invention, the NCA model is used as an example, so the division is based on the acceleration parameter. If other maneuver models are included in the motion model set, the corresponding maneuver parameters are selected for division. For example, the maneuver parameter to be divided in the coordinated turning model is the turning rate.

[0066] Then, for step 102, filtering initialization is performed on each model and each model structure in the motion model set, and after filtering initialization, recursive filtering is started for each model.

[0067] In the embodiment of the present invention, each model and each model structure can be initialized in the following manner:

[0068] During state initialization, the radar's state measurements z1 and z2 at the first two moments are used to initialize the common state components and covariance at the second moment. During initialization, the maneuvering parameters are initialized using the Gaussian mixture distribution corresponding to each high-dimensional model. This Gaussian mixture distribution can more accurately describe the prior information about the maneuvering parameters. Initializing the maneuvering parameters using multiple Gaussian distributions with quantized means improves the match between the actual and initial maneuvering parameters, alleviating the problem of maneuvering parameter mismatch.

[0069] Continuing with the example of a motion model set including one NCV model and n NCA models, then:

[0070] The state estimation and corresponding covariance of the NCV model when k=2 are:

[0071]

[0072]

[0073] in, is the state estimation of the NCV model at time k=2, z k =[x k y k ] T represents the state measurement result received by the radar at time k, z k (1) represents the target position observation in the x direction at time k, z k (2) represents the target position observation in the y direction at time k, and T represents the radar data sampling interval. is the state estimation covariance of the NCV model at time k = 2, and R is the observation noise covariance.

[0074] The state estimate and corresponding covariance of the βth NCA model when k = 2 are:

[0075]

[0076]

[0077] in, is the state estimate of the βth NCA model at time k = 2, is the state estimation covariance of the βth NCA model at time k = 2, 0 m×n Represents an m-row, n-column matrix consisting of 0s.

[0078] The overall state estimate and corresponding covariance of the NCA model structure when k=2 are:

[0079]

[0080]

[0081] It should be noted that, in addition to the two-point initialization method described above for filter initialization, other methods of filter initialization may also be used, such as a single-point initialization method and a three-point initialization method.

[0082] Next, for step 104: at each filtering moment, for each model, the following are performed: using a Gaussian mixture distribution to initialize the maneuver parameter state components of other model structures that have different structures from the model, mixing the overall state estimate of the other model structures with the state estimate of the model at the previous moment to obtain a mixed state estimate; obtaining the current moment measurement, and using the current moment measurement and the mixed state estimate as input to the model, and calculating the conditional state estimate and likelihood function of the model at the current moment after the input.

[0083] In this embodiment of the present invention, each model in the motion model set begins recursive filtering after initialization. At each filtering moment, it is necessary to determine the state estimate of each model at the previous moment and the overall state estimate of each model structure. The overall state estimate of each model and each model structure at the initial filtering moment can be achieved using the formula in step 102 above.

[0084] Considering that the optimal parameter initialization after the maneuver mode takes effect is unique and time-invariant, models with the same structure should operate independently. Therefore, this embodiment uses a partial model interaction mechanism. Models with the same structure operate independently after filter initialization without interfering with each other, while models with different structures can interact to handle possible model jumps. Therefore, in this embodiment of the present invention, for each model, it is necessary to: blend the state estimate of the model at the previous moment with the overall state estimate of each other model structure to obtain a blended state estimate input to the model.

[0085] For example, assuming that the motion model set includes three model structures A, B and C, then for model A1 in the first model structure A, it is necessary to mix the state estimate of model A1 at the previous moment, the overall state estimate of model structure B at the previous moment, and the overall state estimate of model structure C at the previous moment.

[0086] In one embodiment of the present invention, the state estimate of the model at the previous moment and the overall state estimate of each other model structure may be mixed to obtain a mixed state estimate in at least the following manner (1041-1042):

[0087] 1041: Calculate the mixed probability of each model at the current moment based on the preset model transition probability and the probability of the model at the previous moment;

[0088] In the embodiment of the present invention, the model transition probability needs to be pre-set first.

[0089] During motion, a maneuvering target undergoes both mode retention and mode transitions. When the mode retention period is maintained, the target remains in the dwell phase, while when the mode transition occurs, a mode jump occurs. Since the maneuvering parameters that take effect at the mode jump moment belong to only one parameter partition, the optimally initialized maneuvering model is unique and time-invariant during the maneuver dwell phase. This means that after maneuvering parameter initialization, models with the same structure operate independently, with no mode jumps between them. Therefore, the model transition probability between models with the same structure is set to zero. However, mode jumps between models with different structures can occur. Assuming the model set includes three model structures, a model transition probability must be set between any two model structures.

[0090] Continuing with the example of the above motion model set including one NCV model and n NCA models, the probability π that the NCV model can be maintained unchanged 1|1 The probability π that the βth NCA model remains unchanged 2,β|2 Set to P max , the probability π that the NCV model switches to the βth NCA model 2,β|1 Set to (1-P max ) / n, where n is the number of NCA models. In this embodiment, n=9, the probability that the NCA model structure switches to the NCV model is π 1|2 Set to 1-P max .

[0091] Then, it is necessary to pre-set the initial probability of each model structure and the initial probability of each model in the motion model set.

[0092] In the embodiment of the present invention, the initial probability of each model structure at the second moment is the same and can be set to Here, r is the number of model structures. In this embodiment, a motion model set including one NCV model and n NCA models is taken as an example, and r=2.

[0093] In the embodiment of the present invention, for each model structure in the motion model set, the initial probability of each model in the model structure can be set to Where r is the number of model structures, and n is the number of models in the model structure. In this embodiment, the motion model set includes one NCV model and n NCA models as an example. The non-motorized motion model has only one NCV model, so the initial probability of the NCV model is the same as the initial probability of the corresponding model structure, which is The maneuver model includes n NCA models, so the initial probability of each NCA model at the second moment can be set as:

[0094] After setting the model transition probabilities and the initial probabilities for each model and each model structure, the mixing probability needs to be calculated. Mixing refers to the weighted combination of the state estimates from each model at the previous moment and inputs them into the sub-filter corresponding to the model at the next moment. The mixing probability is the weighted proportion of the state at the previous moment. The mixing probability of the input needs to be calculated for each model in the motion model set.

[0095] Specifically, for the current moment k≥3, when the target mode changes, the known model m j,β Under the condition that the model structure m is effective at time k i The probability that (i≠j) takes effect at time k-1 is:

[0096]

[0097] Among them, π j,β|i is the model structure m i (i-th model structure) switch to model m j,β The model transition probability of (the βth model in the jth model structure), is the model structure m i The probability of the model structure at time k-1. When the target mode remains unchanged, if the model effective at time k is m j,β According to the static operation mechanism between models with the same structure, it can be inferred that the model that takes effect at time k-1 is also m j,β , the mixing probability at this time is:

[0098]

[0099] Among them, π j,β|j,β For model m j,β The probability of maintaining the same For model m j,β The probability of the model at time k-1, normalization constant For model m j,β The predicted probability of the model at time k is:

[0100]

[0101] 1042: Using the model mixing probability at the current moment, weight the overall state estimate from each other model structure and the state estimate of the model at the previous moment to obtain a mixed state estimate.

[0102] In the embodiment of the present invention, the mixed state estimation is obtained by:

[0103] For the βth model m in the jth model structure j,β , the calculated mixing probability Weighted k-1 moment model mj,β State estimation and the overall state estimate from other model structures get:

[0104]

[0105]

[0106] in, is the mixed state estimate, Estimate the covariance for the mixture state.

[0107] In an embodiment of the present invention, since the structures of the models in this model and other model structures are different, if the input dimension of the model is different from the output dimension of the other model structures, then when the state estimate of the model is mixed with the overall state estimate of the other model structures, dimension alignment processing is required. That is, the overall state estimate of the mixed model structure is obtained by dimensionally aligning the overall state estimate actually output by the model structure with the input dimension of the model. The alignment method may include:

[0108] If the input dimension of the model is lower than the output dimension of the other model structure, the first several dimensions of the overall state estimation from the other model structure at the previous moment are truncated so that the dimension of the overall state estimation after truncation is the same as the input dimension of the model;

[0109] If the input dimension of the model is higher than the output dimension of other model structures, the overall state estimate from the other model structure at the previous moment is expanded so that the dimension of the overall state estimate after expansion is the same as the input dimension of the model.

[0110] In this embodiment of the present invention, this dimensionality expansion approach is achieved by initializing the maneuver parameters using a Gaussian distribution for the maneuver parameter partition corresponding to the model. Initializing the maneuver parameters using a mixture of multiple Gaussian distributions with quantized means can mitigate deviations in the maneuver parameters. This initialization approach is equivalent to using a Gaussian mixture distribution to approximate a prior uniform distribution. Compared to a single Gaussian distribution, a Gaussian mixture can better describe the prior information of the maneuver parameters and provide better parameter initialization.

[0111] Continuing with the example of a motion model set including one NCV model and n NCA models, since the dimension of the NCV model is lower than the dimension of the NCA model structure, the following is true:

[0112] When calculating the mixed state estimate input to the NCV model at time k, the overall state estimate from the NCA model structure at time k-1 is directly intercepted. State Estimation Covariance The first four dimensions in the truncated state estimate are obtained and the truncated state estimate covariance

[0113] When calculating the mixed state estimate input to the βth NCA model at time k, the Gaussian mixture distribution corresponding to the βth NCA model is used to initialize the maneuver parameters, which is formally expressed as the state estimate from the NCV model at time k-1. and the corresponding state covariance Expand the dimension and obtain the state estimation after the expansion and the corresponding covariance

[0114]

[0115]

[0116] In the embodiment of the present invention, after obtaining the mixed state estimation and the mixed state estimation covariance, it is also necessary to obtain the current moment measurement.

[0117] In the embodiment of the present invention, a linear measurement model is considered:

[0118]

[0119] Among them, z k represents the state measurement result received by the radar at time k, is the model structure m j The observation matrix of (j∈{1,2}), w k is the observation noise with zero mean and covariance R.

[0120] The observation matrix of the NCV model is:

[0121]

[0122] The observation matrix of the NCA model is:

[0123]

[0124] After the measurement is obtained, the mixed state is estimated Estimating covariance with the mixed state And the current state measurement result z k As the k-time model m j,β Corresponding to the filter input.

[0125] Furthermore, when the above data is input into the model m j,β After that, we need to calculate the model m at the current moment j,β The conditional state estimate and the model m j,β The likelihood function of :

[0126] First, calculate the model m j,β State prediction value Covariance with the corresponding state prediction value

[0127]

[0128]

[0129] Then, according to the model m j,β State prediction value Calculate the model m j,β The residual And according to the model m j,β The state prediction value covariance calculates the residual covariance

[0130]

[0131]

[0132] Next, according to the model m j,β The covariance of the state prediction value and the residual covariance Calculate the model m j,β Kalman filter gain

[0133]

[0134] Furthermore, according to the state prediction value of the model residual and Kalman filter gain Calculate the model m j,β Conditional state estimation Estimated covariance with conditional state

[0135]

[0136]

[0137] Finally, according to the model m j,β The residual and residual covariance Calculate the model m j,β Likelihood function

[0138]

[0139] Further with respect to step 106: based on the conditional state estimate and likelihood function of each model at the current moment, the conditional state estimate of each model structure is calculated, and the conditional state estimate of each model structure at the current moment is fused to obtain a fused state estimate; the fused state estimate is the state estimation result of the maneuvering target at the current moment.

[0140] In the embodiment of the present invention, after obtaining the state estimate and likelihood function of each model at the current moment in step 104, it is necessary to fuse the state estimates from models with the same structure to obtain the conditional state estimate of each model structure.

[0141] In the embodiment of the present invention, the conditional state estimation of the current model structure is calculated as follows:

[0142] For the current model structure m j Each model m j,β Both execute:

[0143] 1061: Calculate the model m at the current k moment j,β The conditional predicted probability

[0144] Among them, if the current model structure m at the current k moment is given j If it takes effect, then the model m at time k j,β The conditional prediction probability is

[0145] 1062: Using the current k moment model m j,β The conditional predicted probability And the model m j,β Likelihood function Use the Bayesian formula to update the current model structure m j The model m is effective j,β The conditional probability of:

[0146]

[0147] in, The current model structure m at the current k moment j The likelihood function of :

[0148]

[0149] 1063: According to the current k moment, the model m j,β The conditional probability of And the model m j,β Conditional state estimation Calculate the current model structure m j Conditional state estimation And according to the model mj,β The state estimation covariance of Calculate the current model structure m j The state estimation covariance of

[0150]

[0151]

[0152] In the embodiment of the present invention, it is necessary to update the posterior probability of each model structure at the current moment, and fuse the conditional state estimates and conditional state estimate covariances of the model structures to output as the filtering result at the current moment.

[0153] Specifically, when fusing the conditional state estimates of each model structure at the current moment, the following steps may be included:

[0154] S1: Calculate the predicted probability of each model structure at the current moment;

[0155] Among them, the current model structure m at time k j The predicted probability is the predicted probability of the same structural model The sum of

[0156] S2: Update the posterior probability of each model structure based on the predicted probability of each model structure and the likelihood function of each model structure at the current moment;

[0157] Among them, the current model structure m at time k j The predicted probability And the model structure m j Likelihood function Update model structure m j The posterior probability for:

[0158]

[0159] Where c is the normalization constant

[0160] S3: Use zero-padding to fill the low-dimensional state estimate, and fuse the conditional state estimates of each model structure based on the updated posterior probability of each model structure to obtain a fused state estimate.

[0161] Using zero-padded low-dimensional state estimates and the corresponding state estimate covariance, the posterior probability of the updated model structure is Conditional state estimation by integrating various model structures and conditional state estimated covariance Get the fusion state estimate and the corresponding fusion state estimated covariance P k|k , as the filtering result output at time k:

[0162]

[0163]

[0164] At this point, the state estimation result of the maneuvering target at the current moment is obtained. Further, the current k moment can be regarded as the previous moment, and the process jumps to step 104 until the maneuvering target tracking is completed.

[0165] The effects of the embodiments of the present invention are described below through simulation:

[0166] The maneuvering target is x0 = [1000m, 1000m, 55m / s, 0m / s] T The initial state is uniform motion for 50s, then with acceleration [a, a] T Perform accelerated motion for 50 seconds, then switch to uniform motion for 50 seconds, and then continue with acceleration [-a, -a] T Perform 50s of acceleration. The total simulation time is 200s, and the radar data sampling interval is T = 1s. The process noise covariance is The measurement is the x and y position of the target in the Cartesian coordinate system reported by the radar at the coordinate origin. The measurement noise covariance is R = 400Im 2 2000 Monte Carlo simulations were performed, and each Monte Carlo simulation was performed from the acceleration range [-40,40]m / s 2 The acceleration a is randomly selected, and the model remains unchanged with probability P max =0.95.

[0167] The algorithm used in this embodiment is the GM-PIMM algorithm. The position average Euclidean error of the GM-PIMM algorithm and the IMM algorithm using a single Gaussian distribution to fill the low-dimensional state under 2000 Monte Carlo simulations is as follows: Figure 2 As shown in the figure, the average Euclidean error of the GM-PIMM and the IMM algorithm using a single Gaussian distribution to fill the low-dimensional state under 2000 Monte Carlo simulations is as follows: Figure 3 As shown in the figure, the average Euclidean error of the acceleration of GM-PIMM and IMM algorithm using a single Gaussian distribution to fill the low-dimensional state under 2000 Monte Carlo simulations is as follows: Figure 4 As shown. Figures 2-4 It can be seen from the figure that this embodiment achieves significantly better performance in the target acceleration segment and improves the tracking accuracy of the maneuvering target.

[0168] like Figure 5、 Figure 6 As shown, the embodiment of the present invention provides a partially interactive multi-model tracking device based on Gaussian mixture initialization. The device embodiment can be implemented by software, hardware, or a combination of software and hardware. From the hardware level, Figure 5 As shown in FIG. 1 , a hardware architecture diagram of an electronic device where a partially interactive multi-model tracking device based on Gaussian mixture initialization is provided in an embodiment of the present invention is located, except for Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, the electronic device in the embodiment may also include other hardware, such as a forwarding chip responsible for processing messages, etc. Taking software implementation as an example, Figure 6 As shown, as a device in a logical sense, the CPU of the electronic device in which it is located reads the corresponding computer program in the non-volatile memory into the internal memory and runs it. This embodiment provides a partially interactive multi-model tracking device based on Gaussian mixture initialization, including:

[0169] A construction unit 601 is configured to construct motion models corresponding to possible motion forms of a maneuvering target, thereby obtaining a motion model set. The motion forms include non-maneuvering motion modes and multiple maneuvering modes. The same maneuvering mode is described by multiple maneuvering models having the same structure, and the multiple maneuvering models having the same structure have one-to-one corresponding maneuvering parameter partitions. Models describing the same motion form constitute a single model structure.

[0170] Initialization unit 602, configured to perform filtering initialization on each model and each model structure in the motion model set, and after filtering initialization, each model starts recursive filtering;

[0171] The first computing unit 603 is configured to, at each filtering moment, for each model, perform the following: initializing the maneuver parameter state components of other model structures having different structures from the model using a Gaussian mixture distribution, mixing the overall state estimate of the other model structure with the state estimate of the model at the previous moment to obtain a mixed state estimate; obtaining a current moment measurement, using the current moment measurement and the mixed state estimate as inputs to the model, and calculating a conditional state estimate and a likelihood function of the model at the current moment after the input;

[0172] The second computing unit 604 is used to calculate the conditional state estimate of each model structure based on the conditional state estimate and likelihood function of each model at the current moment, and fuse the conditional state estimate of each model structure at the current moment to obtain a fused state estimate; the fused state estimate is the state estimation result of the maneuvering target at the current moment; the current moment is taken as the previous moment, and the first computing unit is triggered to continue to perform the corresponding operation until the maneuvering target tracking is completed.

[0173] In one embodiment of the present invention, when constructing the motion model of the corresponding motion form, the construction unit specifically includes:

[0174] If the possible motion forms of the maneuvering target include non-maneuvering motion modes, a non-maneuvering model is constructed for the non-maneuvering motion modes;

[0175] If the possible motion forms of the maneuvering target include at least one maneuvering mode, the following steps are performed for each maneuvering mode: constructing multiple maneuvering models with the same structure for describing the maneuvering mode; determining a maneuvering parameter range according to the maneuvering parameters corresponding to the maneuvering mode; equally dividing the maneuvering parameter range into multiple maneuvering parameter partitions, where the number of partitions is equal to the number of models of the multiple maneuvering models constructed; and assigning the multiple maneuvering parameter partitions to the multiple maneuvering models in a one-to-one correspondence.

[0176] In one embodiment of the present invention, when the first computing unit mixes the overall state estimate of the other model structure with the state estimate of the model at the previous moment to obtain a mixed state estimate, it specifically includes: calculating the mixed probability of each model at the current moment based on the preset model transition probability and the probability of the model at the previous moment; using the mixed probability of the model at the current moment to weight the overall state estimate from the other model structure and the state estimate of the model at the previous moment to obtain a mixed state estimate.

[0177] In one embodiment of the present invention, the overall state estimation of the model structure is obtained by aligning the overall state estimation of the actual output of the model structure with the input dimension of the model, and the alignment method is:

[0178] If the input dimension of the model is lower than the output dimension of the other model structure, the first several dimensions of the overall state estimation from the other model structure at the previous moment are truncated so that the dimension of the overall state estimation after truncation is the same as the input dimension of the model;

[0179] If the input dimension of the model is higher than the output dimension of other model structures, the overall state estimate from the other model structure at the previous moment is expanded so that the dimension of the overall state estimate after expansion is the same as the input dimension of the model.

[0180] In one embodiment of the present invention, when calculating the conditional state estimate of each model structure, the second calculation unit specifically includes:

[0181] The conditional state estimate for the current model structure is calculated as:

[0182] For each model in the current model structure, the following steps are performed: calculate the conditional prediction probability of the model at the current moment, and use the conditional prediction probability of the model at the current moment and the likelihood function of the model to update the conditional probability of the model under the current model structure using the Bayesian formula; calculate the conditional state estimate of the current model structure based on the conditional probability of the model at the current moment and the conditional state estimate of the model.

[0183] In one embodiment of the present invention, when the second computing unit fuses the conditional state estimates of each model structure at the current moment, it specifically includes: calculating the prediction probability of each model structure at the current moment; updating the posterior probability of each model structure based on the prediction probability of each model structure at the current moment and the likelihood function of each model structure; using zero-padding of low-dimensional state estimates, and fusing the conditional state estimates of each model structure based on the updated posterior probability of each model structure to obtain a fused state estimate.

[0184] It is understood that the structure illustrated in the embodiment of the present invention does not constitute a specific limitation on a partially interacting multi-model tracking device based on Gaussian mixture initialization. In other embodiments of the present invention, a partially interacting multi-model tracking device based on Gaussian mixture initialization may include more or fewer components than shown in the figure, or combine certain components, split certain components, or arrange the components differently. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0185] The information interaction, execution process, etc. between the modules in the above-mentioned device are based on the same concept as the embodiment of the method of the present invention. For specific contents, please refer to the description in the embodiment of the method of the present invention and will not be repeated here.

[0186] An embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, a partially interactive multi-model tracking method based on Gaussian mixture initialization in any embodiment of the present invention is implemented.

[0187] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor executes a partially interactive multi-model tracking method based on Gaussian mixture initialization in any embodiment of the present invention.

[0188] Specifically, a system or device equipped with a storage medium can be provided, on which software program codes that implement the functions of any of the above-mentioned embodiments are stored, and a computer (or CPU or MPU) of the system or device can be enabled to read and execute the program codes stored in the storage medium.

[0189] In this case, the program code itself read from the storage medium can realize the function of any one of the above-mentioned embodiments, and thus the program code and the storage medium storing the program code constitute part of the present invention.

[0190] Examples of storage media for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code can be downloaded from a server computer via a communication network.

[0191] In addition, it should be clear that the functions of any of the above embodiments can be achieved not only by executing the program code read by the computer, but also by enabling the operating system operating on the computer to complete part or all of the actual operations based on the instructions of the program code.

[0192] In addition, it can be understood that the program code read from the storage medium is written into a memory provided in an expansion board inserted into the computer or into a memory provided in an expansion module connected to the computer, and then based on the instructions of the program code, a CPU installed on the expansion board or expansion module is enabled to perform part or all of the actual operations, thereby realizing the functions of any of the above embodiments.

[0193] It should be noted that, in this article, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises", "comprising" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprising a ..." do not exclude the presence of other identical factors in the process, method, article or device comprising the elements.

[0194] Those skilled in the art will understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk, etc. Various media that can store program codes.

[0195] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A partially interacting multi-model tracking method based on Gaussian mixture initialization, characterized in that: include: Step 100: Based on possible motion forms of the maneuvering target, motion models of corresponding motion forms are constructed to obtain a motion model set; wherein the motion forms include non-maneuvering motion modes and multiple maneuvering modes. The same maneuvering mode is described by multiple maneuvering models with the same structure, and the multiple maneuvering models with the same structure have one-to-one corresponding maneuvering parameter partitions; the models used to describe the same motion form constitute a model structure; Step 102: Initialize the filtering for each model and each model structure in the motion model set, and after the filtering is initialized, start recursive filtering for each model; Step 104: At each filtering moment, for each model, the following steps are performed: using a Gaussian mixture distribution to initialize the maneuver parameter state components of other model structures having different structures from the model, mixing the overall state estimate of the other model structures with the state estimate of the model at the previous moment to obtain a mixed state estimate; obtaining the current moment measurement, and using the current moment measurement and the mixed state estimate as input to the model, and calculating the conditional state estimate and likelihood function of the model at the current moment after the input; Step 106: Calculate the conditional state estimate of each model structure based on the conditional state estimate and likelihood function of each model at the current moment, and fuse the conditional state estimate of each model structure at the current moment to obtain a fused state estimate; the fused state estimate is the state estimate result of the maneuvering target at the current moment; Step 108: Take the current moment as the previous moment and continue to execute step 104 until the maneuvering target tracking is completed.

2. The method according to claim 1, characterized in that The step of constructing a motion model of the corresponding motion form includes: If the possible motion forms of the maneuvering target include non-maneuvering motion modes, a non-maneuvering model is constructed for the non-maneuvering motion mode; If the possible motion forms of the maneuvering target include at least one maneuvering mode, the following steps are performed for each maneuvering mode: constructing multiple maneuvering models with the same structure for describing the maneuvering mode; determining a maneuvering parameter range according to the maneuvering parameters corresponding to the maneuvering mode; equally dividing the maneuvering parameter range into multiple maneuvering parameter partitions, where the number of partitions is equal to the number of models of the multiple maneuvering models constructed; and assigning the multiple maneuvering parameter partitions to the multiple maneuvering models in a one-to-one correspondence.

3. The method according to claim 1, characterized in that The step of mixing the overall state estimate of the other model structure with the state estimate of the model at the previous moment to obtain a mixed state estimate includes: Calculate the mixed probability of each model at the current moment based on the pre-set model transition probability and the probability of the model at the previous moment; The mixed probability of the model at the current moment is used to weight the overall state estimate from the other model structure and the state estimate of the model at the previous moment to obtain a mixed state estimate.

4. The method according to claim 3, characterized in that The overall state estimate of the model structure is obtained by aligning the overall state estimate of the actual output of the model structure with the input dimension of the model. The alignment method is: If the input dimension of the model is lower than the output dimension of the other model structure, the first several dimensions of the overall state estimation from the other model structure at the previous moment are truncated so that the dimension of the overall state estimation after truncation is the same as the input dimension of the model; If the input dimension of the model is higher than the output dimension of other model structures, the overall state estimate from the other model structure at the previous moment is expanded so that the dimension of the overall state estimate after expansion is the same as the input dimension of the model.

5. The method according to any one of claims 1 to 4, characterized in that: The calculating of the conditional state estimate of each model structure includes: The conditional state estimate for the current model structure is calculated as: For each model in the current model structure, the following steps are performed: calculate the conditional prediction probability of the model at the current moment, and use the conditional prediction probability of the model at the current moment and the likelihood function of the model to update the conditional probability of the model under the current model structure using the Bayesian formula; calculate the conditional state estimate of the current model structure based on the conditional probability of the model at the current moment and the conditional state estimate of the model.

6. The method according to any one of claims 1 to 4, characterized in that: The fusing of the conditional state estimation of each model structure at the current moment includes: Calculate the predicted probability of each model structure at the current moment; Update the posterior probability of each model structure based on the predicted probability of each model structure and the likelihood function of each model structure at the current moment; The low-dimensional state estimation is padded with zeros, and the conditional state estimates of each model structure are fused according to the updated posterior probability of each model structure to obtain a fused state estimate.

7. A partially interacting multi-model tracking device based on Gaussian mixture initialization, characterized in that: include: a construction unit configured to construct motion models of corresponding motion forms according to possible motion forms of a maneuvering target, thereby obtaining a motion model set; wherein the motion forms include non-maneuvering motion modes and multiple maneuvering modes, wherein the same maneuvering mode is described by multiple maneuvering models having the same structure, and the multiple maneuvering models having the same structure have one-to-one corresponding maneuvering parameter partitions; and wherein the models describing the same motion form constitute a model structure; an initialization unit, configured to perform filtering initialization on each model and each model structure in the motion model set, and after filtering initialization, each model starts recursive filtering; The first computing unit is configured to, at each filtering moment, execute, for each model, the following: initializing, using a Gaussian mixture distribution, a maneuver parameter state component of another model structure having a different structure from the model, mixing an overall state estimate of the other model structure with a state estimate of the model at a previous moment to obtain a mixed state estimate; obtaining a current moment measurement, using the current moment measurement and the mixed state estimate as inputs to the model, and calculating a conditional state estimate and a likelihood function of the model at the current moment after the input; The second computing unit is used to calculate the conditional state estimate of each model structure based on the conditional state estimate and likelihood function of each model at the current moment, and to fuse the conditional state estimate of each model structure at the current moment to obtain a fused state estimate; the fused state estimate is the state estimation result of the maneuvering target at the current moment; the current moment is taken as the previous moment, and the first computing unit is triggered to continue to perform the corresponding operation until the maneuvering target tracking is completed.

8. The device according to claim 7, characterized in that The construction unit is specifically configured to: if the possible motion forms of the maneuvering target include a non-maneuvering motion mode, construct a non-maneuvering model for the non-maneuvering motion mode; if the possible motion forms of the maneuvering target include at least one maneuvering mode, execute, for each maneuvering mode, the following steps: construct multiple maneuvering models having the same structure for describing the maneuvering mode, determine a maneuvering parameter range based on maneuvering parameters corresponding to the maneuvering mode, equally divide the maneuvering parameter range into multiple maneuvering parameter partitions, the number of partitions being equal to the number of the multiple maneuvering models constructed, and assign the multiple maneuvering parameter partitions to the multiple maneuvering models in a one-to-one correspondence.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method according to any one of claims 1 to 6.

Images