A radar-infrared asynchronous sequential fusion target tracking method and system

Abstract

The application discloses a radar-infrared asynchronous sequential fusion target tracking method and system, radar measurement sequences and infrared measurement sequences are sorted according to time stamps from small to large, and when any measurement arrives, the following is executed: based on the last time fusion state estimation, time prediction is carried out through a state transition model to obtain the current time prior state; the radar / infrared error vector and the error entropy of each dimension thereof are calculated according to the current prior state, and the comprehensive error entropy is calculated; the state posterior of the radar / infrared under the current noise condition and the noise precision parameter posterior are updated by using variational iteration reasoning, and after convergence, the updated fusion state estimation and the noise parameter estimation of each dimension are determined; the above steps are cyclically executed until all asynchronous measurements are processed, and the target tracking trajectory and the error entropy of each dimension of the radar / infrared are obtained. The application adopts an asynchronous sequential processing mechanism, and does not need to strictly synchronize and register the measurements of each sensor, thereby improving real-time performance and engineering realizability.

Description

Technical Field

[0001] This invention belongs to the field of multi-sensor information fusion, target tracking and intelligent perception technology, specifically involving a radar-infrared asynchronous sequential fusion target tracking method and system. Background Technology

[0002] In complex electromagnetic environments, weak target detection environments, and under conditions of strong target maneuverability, single-sensor target tracking systems are easily affected by factors such as time-varying measurement noise, unknown noise statistical characteristics, missing observations, and incomplete dimensions. This can lead to reduced tracking accuracy, increased trajectory fluctuations, and in severe cases, even filter divergence. With the development of heterogeneous sensor collaborative sensing technology, leveraging the complementary advantages of multiple sensors to improve target localization and tracking performance has become an important research direction in the field of information fusion.

[0003] In existing technologies, most multi-sensor fusion methods are based on the minimum mean square error criterion, the Kalman filter framework, or the Gaussian noise assumption. These methods typically rely on prior knowledge of the mean and covariance of measurement noise and utilize second-order statistics to describe error uncertainty. However, in practical applications, radar measurement errors are susceptible to clutter, multipath propagation, obstruction, maneuvering coupling, and environmental disturbances, while infrared measurement errors are easily affected by background thermal noise, target scintillation, line-of-sight obstruction, imaging distortion, and changes in environmental radiation. This results in errors often exhibiting non-Gaussian, heavy-tailed, skewed, and time-varying characteristics. Describing these errors solely through the mean and variance is insufficient to fully reflect the true error distribution.

[0004] The minimum error entropy method measures estimation error uncertainty from an information theory perspective, utilizing higher-order statistical information of the error distribution. Compared to the minimum mean square error method, it exhibits better robustness and resistance to outliers in non-Gaussian noise scenarios. However, traditional minimum error entropy methods typically assume that all sensors have the same measurement dimension, or directly calculate the overall error entropy based on a unified error vector. This is unsuitable for heterogeneous observation scenarios such as radar and infrared sensors, where measurement dimensions are inherently inconsistent. In particular, when infrared sensors lack a range dimension, constructing a unified-dimensional error vector for error entropy calculation introduces physically meaningless range error terms, causing distortion in kernel density estimation and deviations in subsequent fusion results.

[0005] Meanwhile, in asynchronous multi-sensor systems, the arrival times of measurements from different sensors are inconsistent. If synchronous batch processing or forced time registration methods are used, complex interpolation and time-series reconstruction are required for asynchronous observations, which not only involves a large amount of computation but also easily introduces additional modeling errors, reducing the system's real-time performance and engineering adaptability.

[0006] Therefore, there is an urgent need for a sequential fusion method that can adapt to the asynchronous observation characteristics of active radar and passive infrared, allow for the missing infrared range dimension, and operate stably under unknown non-Gaussian noise conditions. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, the radar-infrared asynchronous sequential fusion target tracking method and system provided by this invention can calculate the error entropy for the effective observation dimensions of radar and infrared separately, and perform asynchronous sequential fusion by combining variational inference to adaptively estimate noise parameters. This solves the problems of existing unified error modeling methods being difficult to accurately fuse when the measurement dimensions of active radar and passive infrared are inconsistent, the lack of robustness of traditional fusion methods based on second-order statistics when the statistical characteristics of multi-sensor measurement noise are unknown, non-Gaussian and time-varying, and the need for complex time alignment and high synchronization processing costs under asynchronous multi-sensor measurement arrival conditions.

[0008] To achieve the aforementioned objectives, the present invention employs the following technical solution: a radar-infrared asynchronous sequential fusion target tracking method, which sorts the radar measurement sequence and the infrared measurement sequence according to their timestamps from smallest to largest, and performs the following steps when any measurement arrives: S100. Based on the fusion state estimate of the previous time step, time prediction is performed through the state transition model to obtain the prior state of the current time step. S200: In response to the current arrival measurement coming from active radar / passive infrared, calculate the measurement error of active radar / passive infrared and its error entropy in each dimension based on the current prior state, and then calculate the comprehensive error entropy; S300. Based on the prior state at the current moment, the likelihood of active radar / passive infrared measurement and its corresponding comprehensive error entropy, the radar / infrared optimization problem at the current moment is constructed. Variational iterative inference is used to update the posterior state of radar / infrared and the posterior noise accuracy parameter under the current noise condition. After convergence, the mean of the posterior state is used as the updated fused state estimate, and the posterior expectation of the noise accuracy parameter is used as the updated noise parameter estimate for each dimension. S400, and S100~S300 are executed repeatedly until all asynchronous measurement processing is completed, obtaining the target tracking trajectory and the error entropy of each dimension of radar / infrared.

[0009] Furthermore, in S100, the state transition model is: In the formula, For a moment The state vector, For a moment The state vector, For a moment The corresponding state transition matrix, For a moment Corresponding process noise; The prior state at the current moment is obtained based on the state transition model. for: In the formula, For the previous update time Propagation to the current arrival time The state transition matrix, For a moment Post-instance verification of the state after asynchronous measurement fusion is completed.

[0010] Furthermore, in S200, in response to the current arrival measurement coming from the active radar: Based on the prior state at the current moment Calculate radar predictive measurement Then, the measurement error of the active radar is calculated. ; The comprehensive error entropy corresponding to the active radar for: In the formula, For active radar measurement functions, For active radar at all times The measurement achieved They are time points Measurement errors of active radar in the range, azimuth, and elevation dimensions. , and These are the radar range-dimensional error entropy, the radar azimuth-dimensional error entropy, and the radar elevation-dimensional error entropy, respectively. , and They are respectively , and The weighting coefficients, The total number of samples, , and These are the radar range dimension and bandwidth parameters. Radar azimuth angle core bandwidth parameters Radar elevation angle and bandwidth parameters The kernel function below, , and These are the first two numbers used for kernel density estimation. The errors include distance measurement error, radar azimuth measurement error, and radar elevation measurement error. , and These are the first two numbers used for kernel density estimation. The distance measurement error, radar azimuth measurement error, and radar elevation measurement error are indicated by the superscript. This represents the transpose operator.

[0011] Furthermore, in S200, in response to the current arrival measurement from passive infrared: Based on the prior state at the current moment Calculate infrared predictive measurement Then, the measurement error of passive infrared is calculated. ; The comprehensive error entropy corresponding to the passive infrared for: In the formula, For passive infrared measurement functions, For passive infrared at all times The measurement achieved They are time points The measurement errors of passive infrared in the range and elevation dimensions. and These are the infrared azimuth angle error entropy and the infrared elevation angle error entropy, respectively. and They are respectively and The weighting coefficients of the score The total number of samples, and These are the infrared azimuth angle core bandwidth parameters. and infrared elevation angle and bandwidth parameters The kernel function below, and These are the first two numbers used for kernel density estimation. One infrared azimuth angle measurement error and one infrared elevation angle measurement error. and These are the first two numbers used for kernel density estimation. The infrared azimuth and elevation measurement errors are indicated by the superscript. This represents the transpose operator.

[0012] Furthermore, in S300, a variational posterior formula is established for the state posterior and the radar / infrared noise accuracy parameter posterior to form the radar / infrared optimization problem at the current time. Inverse gamma prior and inverse Wishart prior are adopted respectively, and variational inference is used to alternately update the state posterior and the noise accuracy parameter posterior to achieve adaptive estimation under unknown noise conditions. The variational posterior formula is: In the formula, The joint variational posterior distribution of the current state and radar / infrared noise accuracy parameters. For the posterior state, For the a posteriori of the noise accuracy parameters of active radar, The noise accuracy parameters for passive infrared are a posteriori. For a moment Active radar noise accuracy parameters at that time For a moment Passive infrared noise accuracy parameters at that time This is the reciprocal of the range-dimensional noise variance of the active radar. This is the reciprocal of the noise variance in the azimuth dimension of the active radar. It is the reciprocal of the noise variance in the elevation angle dimension of the active radar. This is the reciprocal of the variance of the passive infrared directional angular dimension noise. This is the reciprocal of the noise variance in the passive infrared elevation angle dimension. The active radar noise accuracy entropy weighting adjustment coefficient is determined by the active radar synthesis error entropy. The passive infrared noise accuracy entropy weighting adjustment coefficient is determined by the passive infrared integrated error entropy. These are matrix operation functions.

[0013] Furthermore, in S300, during the alternating update of the state posterior and noise accuracy parameter posterior, the corresponding noise accuracy parameter is updated according to the currently arrived measurement type, and the posterior of the noise accuracy parameter that has not yet arrived retains the update result of the previous moment and passes it forward.

[0014] Furthermore, in S300, based on the current measurement type, variational iterative reasoning is performed according to the minimization objective function to alternately update the current state posterior and the noise accuracy parameter posterior.

[0015] Furthermore, in S300: When the current arrival measurement comes from active radar, To optimize the target, the state posterior updated by the current active radar is alternately optimized. Post-hoc noise accuracy parameters of active radar in various dimensions ; When the current arrival measurement comes from passive infrared, To optimize the objective, alternately optimize the posterior state after the current passive infrared update. Post-hoc test of noise accuracy parameters in various dimensions of passive infrared ; When the current moment is simultaneously measured by both active radar and passive infrared, perform two optimization updates in sequence according to the preset priority; in, and These are the target functions for active radar and passive infrared, respectively. As the lower bound of evidence, and They are respectively and The corresponding balance coefficient, and They are time points The combined error entropy corresponding to active radar and passive infrared.

[0016] A target tracking system, comprising: Measurement sorting module: used to sort radar measurement sequences and infrared measurement sequences by timestamp from smallest to largest; State prediction module: Used to perform time prediction based on the fused state estimate of the previous time step and the state transition model to obtain the prior state at the current time step; Sequential update module: Based on the current measurement arrival type, it calls the corresponding measurement model to calculate the error vector and its error entropy in each dimension based on the prior state at the current time, and performs sequential update based on the comprehensive error entropy until all asynchronous measurement processing is completed, and obtains the target tracking trajectory and the error entropy in each dimension of radar / infrared. The joint noise estimation module is used to iteratively update the state posterior and noise accuracy parameter posterior during the sequential update process using variational inference. After convergence, the mean of the state posterior is used as the updated fused state estimate, and the expected value of the noise accuracy parameter posterior is used as the updated noise parameter estimate for each dimension.

[0017] Furthermore, the sequential update module includes a radar error entropy calculation unit and an infrared error entropy calculation unit; The radar error entropy calculation unit calculates the radar predicted measurement based on the prior state at the current moment, calculates the radar error vector, and then calculates the radar comprehensive error entropy; the error entropy term in the radar comprehensive error entropy includes radar range dimension error entropy, radar azimuth dimension error entropy, and radar elevation dimension error entropy; The infrared error entropy calculation unit calculates the infrared prediction measurement based on the prior state at the current moment, calculates the infrared error vector, and then calculates the radar comprehensive error entropy; the error entropy term in the infrared comprehensive error entropy includes the infrared azimuth angle error entropy and the infrared elevation angle error entropy. In the sequential update module, during the sequential update process, variational iterative inference is performed based on the current arrival measurement type and the minimization objective function to alternately update the current state posterior and the noise accuracy parameter posterior.

[0018] The beneficial effects of this invention are as follows: (1) This invention addresses the problem of inconsistent measurement dimensions between radar and infrared, especially the lack of distance dimension in infrared. It constructs error entropy for active radar three-dimensional observation and passive infrared two-dimensional observation respectively. In essence, it models differentiated information based on the sensor's observable space, thus avoiding the false distance term caused by the traditional unified dimension error construction from the source. Therefore, it can reduce the kernel density estimation bias, improve the consistency between error entropy evaluation and the actual observation error distribution, and enhance the physical rationality and calculation accuracy of error modeling.

[0019] (2) The present invention introduces the minimum error entropy criterion into the radar-infrared asynchronous fusion process, which can make full use of the higher-order statistical information of the error distribution. While maintaining the computability of the posterior distribution, it can use higher-order statistical information to suppress the influence of outliers, heavy-tailed perturbations and local abnormal measurements, thereby improving the robustness of state estimation in non-Gaussian, heavy-tailed and abnormal perturbation environments.

[0020] (3) The present invention introduces variational reasoning to adaptively estimate noise parameters of different sensors and different observation dimensions. It can dynamically adjust the fusion weight under different sensor working states, background interference and maneuvering states, thereby enhancing the system's environmental adaptability.

[0021] (4) The present invention adopts an asynchronous sequential processing mechanism, which does not require strict synchronous registration of each sensor measurement, thus improving real-time performance and engineering feasibility.

[0022] (5) The final output of this invention includes not only the fusion state results, but also the corresponding dimension error entropy results of active radar and passive infrared and the noise parameter estimation results, which facilitates subsequent sensor health assessment, adaptive weight allocation and resource scheduling. Attached Figure Description

[0023] Figure 1 The flowchart of the radar-infrared asynchronous sequential fusion target tracking method provided by the present invention is shown. Detailed Implementation

[0024] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0025] Example 1:

[0026] This invention provides an asynchronous sequential fusion method that can calculate the error entropy for the effective observation dimensions of radar and infrared separately, and perform adaptive estimation of noise parameters by combining variational inference, thereby realizing target tracking under asynchronous sequential radar-infrared fusion.

[0027] refer to Figure 1 A radar-infrared asynchronous sequential fusion target tracking method sorts the radar measurement sequence and the infrared measurement sequence in ascending order of timestamps. When any measurement arrives, the following steps are performed: S100. Based on the fusion state estimate of the previous time step, time prediction is performed through the state transition model to obtain the prior state of the current time step. S200: In response to the current arrival measurement coming from active radar / passive infrared, calculate the measurement error of active radar / passive infrared and its error entropy in each dimension based on the current prior state, and then calculate the comprehensive error entropy; S300. Based on the prior state at the current moment, the likelihood of active radar / passive infrared measurement and its corresponding comprehensive error entropy, the radar / infrared optimization problem at the current moment is constructed. Variational iterative inference is used to update the posterior state of radar / infrared and the posterior noise accuracy parameter under the current noise condition. After convergence, the mean of the posterior state is used as the updated fused state estimate, and the posterior expectation of the noise accuracy parameter is used as the updated noise parameter estimate for each dimension. S400, and S100~S300 are executed repeatedly until all asynchronous measurement processing is completed, obtaining the target tracking trajectory and the error entropy of each dimension of radar / infrared.

[0028] In S100 of this embodiment, the fused state estimate at the previous time step refers to the posterior state estimate obtained after the previous asynchronous measurement update. The prior state at the current moment refers to the state transition model used to... The predicted state is propagated to the current measurement arrival time; the state transition model is as follows: In the formula, For a moment The state vector, For a moment The state vector, For a moment The corresponding state transition matrix, For a moment The corresponding process noise, These are the position components of the target in three-dimensional space. These are the velocity components of the target in three-dimensional space, indicated by superscript. This is the transpose operator.

[0029] At the moment when the first measurement is reached, the fusion state estimate and initial covariance are given through initialization. The initial value can be obtained from the radar initial acquisition result, the infrared initial orientation result, the track initial estimate or external prior information. The posterior state estimate output by each subsequent asynchronous update is used as the starting point for the next time prediction.

[0030] Furthermore, the prior state at the current moment is obtained based on the state transition model. for: Its corresponding predictive covariance for: In the formula, For the previous update time Propagation to the current arrival time The state transition matrix, The covariance of the previous time step. For a moment Post-fusion state verification after asynchronous measurement completion The process noise covariance matrix is ​​given by the prior state and covariance at the first arrival measurement time. For the first arrival measurement time, the prior state and covariance at the current time can be given by the target prior track, system calibration parameters or target initial acquisition results.

[0031] In S200 of this embodiment of the invention, the active radar and the passive infrared are at time... The achieved measurement is: In the formula, For active radar measurement, For passive infrared measurement, For radar noise measurement, These are the target range, radar azimuth angle, and radar measurement noise, respectively. The azimuth and elevation angles are respectively achieved using infrared sensors; however, since the passive infrared sensor does not directly provide range observations, the above measurement model does not include a range dimension. The measurement function... Used to convert Cartesian state variables into range, azimuth, and elevation observation domains.

[0032] Based on the current measurement, in response to the current arriving measurement from active radar: Based on the prior state at the current moment Calculate radar predictive measurement Then, the measurement error of the active radar is calculated. .

[0033] Based on the error vector, for active radar, error sample sets are constructed in the range, azimuth, and elevation dimensions, respectively. The error entropy of each dimension is calculated based on kernel density estimation, thereby determining the comprehensive error entropy corresponding to the active radar. for: In the formula, For active radar measurement functions, For active radar at all times The measurement achieved They are time points Measurement errors of active radar in the range, azimuth, and elevation dimensions. , and These are the radar range-dimensional error entropy, the radar azimuth-dimensional error entropy, and the radar elevation-dimensional error entropy, respectively. , and They are respectively , and The weighting coefficients, The total number of samples, , and These are the radar range dimension and bandwidth parameters. Radar azimuth angle core bandwidth parameters Radar elevation angle and bandwidth parameters The kernel function below, , and These are the first two numbers used for kernel density estimation. The errors include distance measurement error, radar azimuth measurement error, and radar elevation measurement error. , and These are the first two numbers used for kernel density estimation. The distance measurement error, radar azimuth measurement error, and radar elevation measurement error are indicated by the superscript. This represents the transpose operator.

[0034] In this invention, the aforementioned active radar integrated error entropy is a core formula designed around the "heterogeneous error entropy". It no longer forcibly constructs a unified dimension error vector for different sensors, but only models and weights the three dimensions that are actually observable by the active radar.

[0035] Response to current arrival measurement from passive infrared: Based on the prior state at the current moment Calculate infrared predictive measurement Then, the measurement error of passive infrared is calculated. ; Based on the error vector, for passive infrared, error sample sets are constructed in the azimuth and elevation dimensions respectively, and the error entropy of each dimension is calculated based on kernel density estimation, thereby determining the comprehensive error entropy corresponding to passive infrared. for: In the formula, For passive infrared measurement functions, For passive infrared at all times The measurement achieved They are time points The measurement errors of passive infrared in the range and elevation dimensions. and These are the infrared azimuth angle error entropy and the infrared elevation angle error entropy, respectively. and They are respectively and The weighting coefficients of the score The total number of samples, and These are the infrared azimuth angle core bandwidth parameters. and infrared elevation angle and bandwidth parameters The kernel function below, and These are the first two numbers used for kernel density estimation. One infrared azimuth angle measurement error and one infrared elevation angle measurement error. and These are the first two numbers used for kernel density estimation. The infrared azimuth and elevation measurement errors are indicated by the superscript. This represents the transpose operator.

[0036] In this invention, the calculation of the passive infrared integrated error entropy involves the infrared sensor participating in error entropy optimization only within its own truly observable two-dimensional angular space, thereby avoiding meaningless compensation terms when the distance dimension is missing.

[0037] In S300 of this embodiment of the invention, a variational posterior formula is established for the state posterior and the radar / infrared noise accuracy parameter posterior to form the radar / infrared optimization problem at the current time. The inverse gamma prior and the inverse Wishart prior are used respectively. The state posterior and the noise accuracy parameter posterior are updated alternately by variational inference to achieve adaptive estimation under unknown noise conditions. The variational posterior formula is: In the formula, The joint variational posterior distribution of the current state and radar / infrared noise accuracy parameters. For the posterior state, For the a posteriori of the noise accuracy parameters of active radar, The noise accuracy parameters for passive infrared are a posteriori. For a moment Active radar noise accuracy parameters at that time For a moment Passive infrared noise accuracy parameters at that time This is the reciprocal of the range-dimensional noise variance of the active radar. This is the reciprocal of the noise variance in the azimuth dimension of the active radar. It is the reciprocal of the noise variance in the elevation angle dimension of the active radar. This is the reciprocal of the variance of the passive infrared directional angular dimension noise. This is the reciprocal of the noise variance in the passive infrared elevation angle dimension. The active radar noise accuracy entropy weighting adjustment coefficient is determined by the active radar synthesis error entropy. The passive infrared noise accuracy entropy weighting adjustment coefficient is determined by the passive infrared integrated error entropy. These are matrix operation functions.

[0038] in, For the posterior state, the likelihood is measured by active radar. Passive infrared measurement likelihood It was calculated within the Bayesian estimation framework.

[0039] In this embodiment of the invention, during the alternating update of the state posterior and noise accuracy parameter posterior, the corresponding noise accuracy parameter is updated according to the currently arrived measurement type, and the posterior of the noise accuracy parameter that has not yet arrived retains the update result of the previous moment and passes it forward.

[0040] Specifically, when the current measurement comes from active radar, the focus is on updating... and ,and As a historical a posteriori, it is prioritized for updating when the current measurement comes from passive infrared. and ,and The noise accuracy parameters involved in active radar and passive infrared do not remain unchanged; therefore, the noise accuracy parameters involved in active radar and passive infrared are not solved simultaneously at every micro-update moment, but are selectively activated according to the arrival of the measurement source.

[0041] As a preferred implementation, the aforementioned noise accuracy parameters can be updated alternately using ascending variational inference, that is, the state post-hoc is first updated under the current noise parameter conditions. Then, under the current state posterior conditions, update the posterior noise parameters corresponding to the activated sensor until the lower bound of evidence converges or the set number of iterations is reached.

[0042] In this embodiment of the invention, based on the current measurement type, variational iterative inference is performed according to the minimization objective function to alternately update the current state posterior and the noise accuracy parameter posterior.

[0043] Specifically, by minimizing the objective function, we simultaneously achieve posterior state approximation, adaptive estimation of noise parameters, and minimization of error entropy; whereby the objective function is expressed as: Based on the aforementioned update process description, the objective function does not exist independently, but rather serves as a local solution criterion after each asynchronous measurement arrives. Each completed local optimization outputs a new state posterior, the corresponding sensor noise parameter posterior, and the current dimension error entropy, which are then used as input for the next asynchronous measurement update.

[0044] In this embodiment, when the current arrival measurement comes from active radar, with To optimize the target, the state posterior updated by the current active radar is alternately optimized. Post-hoc noise accuracy parameters of active radar in various dimensions Furthermore, the fusion state estimate updated by the active radar will serve as the input for subsequent updates of other measurements at the same macroscopic moment. If no other measurements arrive after that moment, it will be directly used as the final fusion state output at that moment, and further used as the initial condition for time prediction at the next asynchronous moment.

[0045] When the current arrival measurement comes from passive infrared, To optimize the objective, alternately optimize the posterior state after the current passive infrared update. Post-hoc test of noise accuracy parameters in various dimensions of passive infrared Furthermore, the fusion state estimate after passive infrared update has the same function as the active radar update result, that is, it serves as the input state when subsequent asynchronous measurements arrive. If the infrared update is the last update at the current macroscopic moment, it serves as the final fusion state output at that moment, and further serves as the initial condition for the time prediction of the next asynchronous moment.

[0046] When measurements are received simultaneously from both active radar and passive infrared at the current moment, two optimization updates are performed sequentially according to a preset priority. For example, the update is performed first according to the radar measurement, and then the updated intermediate fusion state is used as the prior input for the infrared measurement update. Alternatively, the earlier timestamp can be prioritized. The final posterior state obtained after the two sub-updates still belongs to the asynchronous sequential fusion framework of this invention.

[0047] in, and These are the target functions for active radar and passive infrared, respectively. As the lower bound of evidence, and They are respectively and The corresponding balance coefficient, and They are time points The combined error entropy corresponding to active radar and passive infrared.

[0048] This invention achieves radar-infrared asynchronous sequential fusion target tracking for target state estimation in an asynchronous heterogeneous multi-sensor system composed of an active radar sensor and a passive infrared sensor under unknown non-Gaussian noise conditions, by using variational inference and minimum error entropy joint modeling.

[0049] Example 2:

[0050] This invention provides a target tracking system, which is a further extension of Embodiment 1. It is implemented based on the radar-infrared asynchronous sequential fusion target tracking method in Embodiment 1. Other contents not mentioned refer to Embodiment 1 or related prior art descriptions.

[0051] The target tracking system in this embodiment of the invention includes: Measurement sorting module: used to sort radar measurement sequences and infrared measurement sequences by timestamp from smallest to largest; State prediction module: Used to perform time prediction based on the fused state estimate of the previous time step and the state transition model to obtain the prior state at the current time step; Sequential update module: Based on the current measurement arrival type, it calls the corresponding measurement model to calculate the error vector and its error entropy in each dimension based on the prior state at the current time, and performs sequential update based on the comprehensive error entropy until all asynchronous measurement processing is completed, and obtains the target tracking trajectory and the error entropy in each dimension of radar / infrared. The joint noise estimation module is used to iteratively update the state posterior and noise accuracy parameter posterior during the sequential update process using variational inference. After convergence, the mean of the state posterior is used as the updated fused state estimate, and the expected value of the noise accuracy parameter posterior is used as the updated noise parameter estimate for each dimension.

[0052] In this embodiment of the invention, the sequential update module includes a radar error entropy calculation unit and an infrared error entropy calculation unit; The radar error entropy calculation unit calculates the radar predicted measurement based on the prior state at the current moment, calculates the radar error vector, and then calculates the radar comprehensive error entropy; the error entropy term in the radar comprehensive error entropy includes radar range dimension error entropy, radar azimuth dimension error entropy, and radar elevation dimension error entropy; The infrared error entropy calculation unit calculates the infrared prediction measurement based on the prior state at the current moment, calculates the infrared error vector, and then calculates the radar comprehensive error entropy; the error entropy term in the infrared comprehensive error entropy includes the infrared azimuth angle error entropy and the infrared elevation angle error entropy. In the sequential update module, during the sequential update process, based on the currently arrived measurement type, variational iterative inference is performed according to the minimized objective function to alternately update the current state posterior and the noise accuracy parameter posterior. During the alternate update of the state posterior and the noise accuracy parameter posterior, the corresponding noise accuracy parameter is updated according to the currently arrived measurement type. For the noise accuracy parameter posterior that has not yet arrived, the update result of the previous time step is maintained and passed on to the future.

[0053] Among them, the radar error entropy calculation unit calculates the comprehensive error entropy corresponding to the active radar. for: Comprehensive error entropy corresponding to passive infrared for: In this invention, through the above calculation method, the active radar only participates in the error entropy calculation in the three dimensions of its own observable range, azimuth angle and elevation angle, while the passive infrared only participates in the error entropy calculation in the two effective dimensions of azimuth angle and elevation angle, thereby avoiding the introduction of invalid error terms due to the lack of range dimension in infrared.

[0054] In this embodiment of the invention, during the variational iterative inference process, when the current measurement comes from active radar, the focus is on updating... and ,and As a historical a posteriori, it is prioritized for updating when the current measurement comes from passive infrared. and ,and It remains unchanged.

[0055] Furthermore, the variational iterative update process based on the objective function is as follows: When the current arrival measurement comes from active radar, To optimize the target, the state posterior updated by the current active radar is alternately optimized. Post-hoc noise accuracy parameters of active radar in various dimensions ; When the current arrival measurement comes from passive infrared, To optimize the objective, alternately optimize the posterior state after the current passive infrared update. Post-hoc test of noise accuracy parameters in various dimensions of passive infrared ; When the current moment is simultaneously measured by both active radar and passive infrared, two optimization updates are performed sequentially according to the preset priority.

[0056] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.

[0057] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0058] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A target tracking method using radar-infrared asynchronous sequential fusion, characterized in that, Sort the radar measurement sequence and infrared measurement sequence in ascending order of timestamp. When any measurement arrives, perform the following steps: S100. Based on the fusion state estimate of the previous time step, time prediction is performed through the state transition model to obtain the prior state of the current time step. S200: In response to the current arrival measurement coming from active radar / passive infrared, calculate the measurement error of active radar / passive infrared and its error entropy in each dimension based on the current prior state, and then calculate the comprehensive error entropy; S300. Based on the prior state at the current moment, the likelihood of active radar / passive infrared measurement and its corresponding comprehensive error entropy, the radar / infrared optimization problem at the current moment is constructed. Variational iterative inference is used to update the posterior state of radar / infrared and the posterior noise accuracy parameter under the current noise condition. After convergence, the mean of the posterior state is used as the updated fused state estimate, and the posterior expectation of the noise accuracy parameter is used as the updated noise parameter estimate for each dimension. S400, repeat S100~S300 until all asynchronous measurement processing is completed, and obtain the target tracking trajectory and the error entropy of each dimension of radar / infrared; In S200, in response to the current arrival measurement coming from the active radar: Based on the prior state at the current moment Calculate radar predictive measurement Then, the measurement error of the active radar is calculated. ; The comprehensive error entropy corresponding to the active radar for: In the formula, For active radar measurement functions, For active radar at all times The measurement achieved They are time points Measurement errors of active radar in the range, azimuth, and elevation dimensions. , and These are the radar range-dimensional error entropy, the radar azimuth-dimensional error entropy, and the radar elevation-dimensional error entropy, respectively. , and They are respectively , and The weighting coefficients, The total number of samples, , and These are the radar range dimension and bandwidth parameters. Radar azimuth angle core bandwidth parameters Radar elevation angle and bandwidth parameters The kernel function below, , and These are the first two numbers used for kernel density estimation. The errors include distance measurement error, radar azimuth measurement error, and radar elevation measurement error. , and These are the first two numbers used for kernel density estimation. The distance measurement error, radar azimuth measurement error, and radar elevation measurement error are indicated by the superscript. This represents the transpose operator; Response to current arrival measurement from passive infrared: Based on the prior state at the current moment Calculate infrared predictive measurement Then, the measurement error of passive infrared is calculated. ; The comprehensive error entropy corresponding to the passive infrared for: In the formula, For passive infrared measurement functions, For passive infrared at all times The measurement achieved They are time points The measurement errors of passive infrared in the directional and elevation dimensions. and These are the infrared azimuth angle error entropy and the infrared elevation angle error entropy, respectively. and They are respectively and The weighting coefficients, The total number of samples, and These are the infrared azimuth angle core bandwidth parameters. and infrared elevation angle and bandwidth parameters The kernel function below, and These are the first two numbers used for kernel density estimation. One infrared azimuth angle measurement error and one infrared elevation angle measurement error. and These are the first two numbers used for kernel density estimation. The infrared azimuth and elevation measurement errors are indicated by the superscript. This represents the transpose operator.

2. The target tracking method of radar-infrared asynchronous sequential fusion according to claim 1, characterized in that, In S100, the state transition model is as follows: In the formula, For a moment The state vector, For a moment The state vector, For a moment The corresponding state transition matrix, For a moment Corresponding process noise; The prior state at the current moment is obtained based on the state transition model. for: In the formula, For the previous update time Propagation to the current arrival time The state transition matrix, For a moment Post-instance verification of the state after asynchronous measurement fusion is completed.

3. The target tracking method of radar-infrared asynchronous sequential fusion according to claim 1, characterized in that, In S300, a variational posterior formula is established for the state posterior and the radar / infrared noise accuracy parameter posterior to form the radar / infrared optimization problem at the current time. The inverse gamma prior and the inverse Wishart prior are adopted respectively. The state posterior and the noise accuracy parameter posterior are updated alternately by variational inference to achieve adaptive estimation under unknown noise conditions. The variational posterior formula is as follows: In the formula, The joint variational posterior distribution of the current state and radar / infrared noise accuracy parameters. For the posterior state, For the a posteriori of the noise accuracy parameters of active radar, The noise accuracy parameters for passive infrared are a posteriori. For a moment Active radar noise accuracy parameters at that time For a moment Passive infrared noise accuracy parameters at that time This is the reciprocal of the range-dimensional noise variance of the active radar. This is the reciprocal of the noise variance in the azimuth dimension of the active radar. It is the reciprocal of the noise variance in the elevation angle dimension of the active radar. This is the reciprocal of the variance of the passive infrared directional angular dimension noise. This is the reciprocal of the noise variance in the passive infrared elevation angle dimension. The active radar noise accuracy entropy weighting adjustment coefficient is determined by the active radar synthesis error entropy. The passive infrared noise accuracy entropy weighting adjustment coefficient is determined by the passive infrared integrated error entropy. These are matrix operation functions.

4. The target tracking method of radar-infrared asynchronous sequential fusion according to claim 3, characterized in that, In S300, during the alternating update of state posterior and noise accuracy parameter posterior, the corresponding noise accuracy parameter is updated according to the currently arrived measurement type, and the posterior of the noise accuracy parameter that has not yet arrived retains the update result of the previous moment and passes it on to the next step.

5. The target tracking method of radar-infrared asynchronous sequential fusion according to claim 1, characterized in that, In step S300, based on the current measurement type, variational iterative reasoning is performed according to the minimized objective function to alternately update the current state posterior and the noise accuracy parameter posterior.

6. The target tracking method of radar-infrared asynchronous sequential fusion according to claim 5, characterized in that, In S300: When the current arrival measurement comes from active radar, To optimize the target, the state posterior updated by the current active radar is alternately optimized. Post-hoc noise accuracy parameters of active radar in various dimensions ; When the current arrival measurement comes from passive infrared, To optimize the objective, alternately optimize the posterior state after the current passive infrared update. Post-hoc test of noise accuracy parameters in various dimensions of passive infrared ; When the current moment is simultaneously measured by both active radar and passive infrared, perform two optimization updates in sequence according to the preset priority; in, and These are the target functions for active radar and passive infrared, respectively. As the lower bound of evidence, and They are respectively and The corresponding balance coefficient, and They are time points The combined error entropy corresponding to active radar and passive infrared.

7. A target tracking system, implemented based on the radar-infrared asynchronous sequential fusion target tracking method according to any one of claims 1 to 6, characterized in that, include: Measurement sorting module: used to sort radar measurement sequences and infrared measurement sequences by timestamp from smallest to largest; State prediction module: Used to perform time prediction based on the fused state estimate of the previous time step and the state transition model to obtain the prior state at the current time step; Sequential update module: Based on the current measurement arrival type, it calls the corresponding measurement model to calculate the error vector and its error entropy in each dimension based on the prior state at the current time, and performs sequential update based on the comprehensive error entropy until all asynchronous measurement processing is completed, and obtains the target tracking trajectory and the error entropy in each dimension of radar / infrared. The joint noise estimation module is used to iteratively update the state posterior and noise accuracy parameter posterior during the sequential update process using variational inference. After convergence, the mean of the state posterior is used as the updated fused state estimate, and the expected value of the noise accuracy parameter posterior is used as the updated noise parameter estimate for each dimension.

8. The target tracking system according to claim 7, characterized in that, The sequential update module includes a radar error entropy calculation unit and an infrared error entropy calculation unit. The radar error entropy calculation unit calculates the radar predicted measurement based on the prior state at the current moment, calculates the radar error vector, and then calculates the radar comprehensive error entropy; the error entropy term in the radar comprehensive error entropy includes radar range dimension error entropy, radar azimuth dimension error entropy, and radar elevation dimension error entropy; The infrared error entropy calculation unit calculates the infrared prediction measurement based on the prior state at the current moment, calculates the infrared error vector, and then calculates the infrared comprehensive error entropy. The error entropy term in the infrared integrated error entropy includes infrared azimuth angle error entropy and infrared elevation angle error entropy. In the sequential update module, during the sequential update process, variational iterative inference is performed based on the current arrival measurement type and the minimization objective function to alternately update the current state posterior and the noise accuracy parameter posterior.

Images