Maneuvering target association and tracking method under sparse observation based on bimodal particle flow

Abstract

The invention provides a maneuvering target correlation and tracking method under sparse observation based on a bimodal particle flow, and the method comprises the steps: constructing a core layer and exploration layer bimodal particle swarm with differentiated process noise, and employing a particle individual nearest neighbor normalization distance based on measurement noise as a soft gate correlation criterion. And the overall particle distribution is corrected by using a particle flow algorithm after exploration of particle capture observation. According to the method, the problem of deadlock caused by excessive confidence can be solved, accurate tracking of the sudden maneuvering target is always kept, the robustness under sparse observation can be remarkably improved, and breakpoint-free tracking can be realized. Moreover, the method does not need to maintain a plurality of filter models like an IMM algorithm, can achieve the coverage of a plurality of motion modes only through the adjustment of the particle noise distribution, is suitable for the real-time implementation of an engineering, and is more suitable for GPU parallelism.

Description

Technical Field

[0001] This application belongs to the field of target tracking technology, specifically relating to a method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow. Background Technology

[0002] Radar target tracking is a core function of modern defense systems. With the advent of hypersonic vehicles and highly maneuverable unmanned aerial vehicles (UAVs), target maneuverability has significantly increased. Simultaneously, to balance multi-target search and reduce the probability of interception, radars often employ intermittent scanning or low data rate operating modes (i.e., sampling intervals). (Larger). Under this "sparse observation" condition, the displacement of the target between two frames is extremely nonlinear, which poses a severe challenge to the prediction and correlation capabilities of the filter.

[0003] Currently, the mainstream nonlinear filtering and correlation techniques mainly include: (1) Gate correlation based on Mahalanobis distance (traditional method), this scheme usually calculates the observation values Compared with the predicted mean Mahalanobis distance , is represented as: .in , To predict covariance.

[0004] (2) Interactive Multiple Model (IMM): This scheme defines multiple model sets such as constant speed (CV) and cooperative turning (CT) and uses Markov chains to transfer model probabilities.

[0005] (3) Particle Flow Filter (PFF): This scheme proposes a filter based on logarithmic homotopy flow, which smoothly migrates particles from the prior distribution to the posterior distribution by solving partial differential equations.

[0006] However, the above solution has the following drawbacks: The gate correlation method based on Mahalanobis distance, when the filter converges, When the covariance matrix approaches zero, the gate range shrinks. If the target undergoes strong maneuvers, the distance between the actual observation point and the predicted mean will far exceed the covariance range, leading to association failure. Interactive multi-model model set design relies on prior knowledge, and under sparse observations, model probability updates are not timely, causing tracking divergence. Particle flow filtering algorithms typically assume data association is already complete and lack a mechanism for actively capturing observations in complex clutter or strong maneuvering environments. Summary of the Invention

[0007] To address the aforementioned problems in related technologies, this application provides a method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow. The technical problem to be solved by this application is achieved through the following technical solution: A method for correlation and tracking of maneuvering targets under sparse observations based on dual-modal particle flow includes: S100: Acquire the observation values ​​obtained from radar observation of the maneuvering target and set the maneuvering range of multiple particles dispersed to cover the maneuvering target; S200, the particles within the maneuvering range are divided into core particles and exploration particles, and a one-step state prediction is performed on all particles so that the exploration particles disperse in all directions and the core particles maintain their inertia to obtain the current state of the particles. S300, based on the normalized measurement distance between the observed value and each particle in the current state, perform an association determination on the observed value to obtain an association determination result; S400, Update the positions of all particles according to the association determination result to obtain the updated position coordinates of all particles; S500 determines the position of the maneuvering target based on the updated positions of all particles after filtering.

[0008] Beneficial effects: This application proposes a method for association and tracking of maneuvering targets under sparse observations based on dual-modal particle flow. It constructs a dual-modal particle swarm with a core layer and an exploration layer containing differentiated process noise. The method employs the nearest neighbor normalized distance of individual particles based on measurement noise as a soft-wave gate association criterion, and uses a particle flow algorithm to correct the overall particle distribution after the exploration particle capture observation. This application can solve the deadlock problem caused by overconfidence, maintain accurate tracking of sudden maneuvering targets, and significantly improve robustness under sparse observations, enabling uninterrupted tracking. Furthermore, unlike the IMM algorithm, this application does not require maintaining multiple filter models; it only needs to adjust the particle noise distribution to cover multiple motion modes, making it suitable for real-time engineering implementation. The particle flow filtering method is also more suitable for GPU parallelization.

[0009] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the method for correlation and tracking of maneuvering targets based on sparse observations of dual-modal particle flow provided in this application. Figure 2 This is the flowchart of the procedure provided in this application; Figure 3 This is a schematic diagram of the scene motion trajectory provided in this application; Figure 4This is a schematic diagram of the gate correlation index provided in this application; Figure 5 This is a schematic diagram comparing the position tracking accuracy (RMSE) provided in this application. Detailed Implementation

[0011] The present application will be described in further detail below with reference to specific embodiments, but the implementation of the present application is not limited thereto.

[0012] like Figure 1 As shown, this application provides a method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow, including: S100: Acquire the observation values ​​obtained from radar observation of the maneuvering target and set the maneuvering range of multiple particles dispersed to cover the maneuvering target; S200, the particles within the maneuvering range are divided into core particles and exploration particles, and a one-step state prediction is performed on all particles so that the exploration particles disperse in all directions and the core particles maintain their inertia to obtain the current state of the particles. S300, based on the normalized measurement distance between the observed value and each particle in the current state, perform an association determination on the observed value to obtain an association determination result; S400, Update the positions of all particles according to the association determination result to obtain the updated position coordinates of all particles; S500 determines the position of the maneuvering target based on the updated positions of all particles after filtering.

[0013] This application calculates the mean coordinates of the updated positions of all particles to obtain the mean coordinates, and uses these mean coordinates as the position of the maneuvering target estimated by filtering.

[0014] In one specific embodiment of this application, S200 includes: S210, to form a particle swarm of particles within the said mobility range. Division of core particle swarm And exploring particle swarm The core particle swarm contains the following number of core particles: Its noise is very small. The number of exploratory particles contained in the exploratory particle swarm is Its noise is enormous. ; S220, state prediction is performed on all particles using the state transition equation to ensure that the exploring particles disperse in all directions and the core particle maintains its inertia. The state transition equation is expressed as:

[0015] In the formula, Indicates the first Individual particles The state at any given moment, Represents the state transition matrix. Indicates the first Individual particles The state at any given moment, Indicates the first Individual particles The noise term at time step, adjusted according to particle type, is represented as follows:

[0016] In the formula, The variance of the velocity component is set to of Doubled to cover sparse sampling intervals The maximum possible range of maneuverability within the area.

[0017] In one specific embodiment of this application, S300 includes: S310 calculates the normalized measurement distance between the observation and each particle in the current state.

[0018] This application discards Traditional gates are proposed, based on measurement noise. Particle-level measurement.

[0019] For observation , define the first The normalized measured distance of each particle is expressed as:

[0020] In the formula, Indicates in Observations at time and Normalized measured distance between them Represents the nonlinear measurement function with respect to the first Predicted observations of the state of each particle. The superscript represents the inverse of the measurement noise covariance matrix. Indicates transpose; The denominator here contains only Not included This means that the criterion depends solely on the sensor's own measurement accuracy, and is unrelated to the filter's confidence level.

[0021] S320, based on the normalized measurement distance between the observed value and each particle in the current state, the observed value is correlated to obtain a correlation determination result.

[0022] In one specific embodiment of this application, S320 includes: S321, Calculate the global minimum particle distance index based on the normalized measurement distance between the observed value and each particle in the current state, expressed as:

[0023] S322, The observed values ​​are correlated using the global minimum distance index to obtain the correlation determination result.

[0024] In one specific embodiment of this application, S322 includes: S3221, For each particle in the current state, if the global minimum distance index calculated between the particle and the observation value is not greater than the threshold, then the association determination result of the observation value and the particle being successfully associated is obtained. S3222, if the global minimum distance index is greater than the threshold, then the observed value is obtained as the clutter observation value or the correlation determination result of the lost maneuvering target.

[0025] Association determination is: if If it is determined to be clutter or lost tracking, only prediction output is performed. The association is determined to be successful, where, This represents the threshold.

[0026] In one specific embodiment of this application, S400 includes: S410, if the association determination result is that the observation value is successfully associated with the particle, then the position of all particles is updated using the observation value to obtain the updated position coordinates of all particles. This application identifies the successfully associated observations and particles in the association determination results, and transmits the position information of the particle to the other particles so that they move closer to the particle to update their own positions, until the global minimum distance index between the particle and the observation is less than a threshold, thus obtaining the updated position coordinates of all particles.

[0027] Global particles update their positions using the Daum-Huang homotopy equation, when When the threshold is reached, the observation is considered valid. The homotopic flow equation is expressed as:

[0028] In the formula, This is a pseudo-time step, ranging from 0 to 1. This represents the flow gain matrix at the pseudo-time step. Represents a nonlinear measurement function. This represents the observed value at the current moment. This represents the particle state vector.

[0029] The update process generates a flow field that pulls the originally deviated core particles along the probability gradient direction toward the high likelihood region discovered by the exploration particles.

[0030] S420, if the correlation determination result is the observation value of clutter or the loss of maneuvering target, the CV uniform linear motion model is used to predict and update the position of all particles to obtain the updated position coordinates of all particles.

[0031] This application assumes the number of particles is... The state dimension is The measurement dimension is This application requires only a single particle flow update. The particle prediction complexity is O(n). The computational complexity of distance calculation is The update complexity of the stream is .

[0032] To verify the effectiveness of the correlation and tracking method based on dual-modal particle flow proposed in this application under sparse observation and strong maneuvering conditions, the following simulation scenario was constructed, and specific experimental parameters were set for comparative verification.

[0033] This application constructs a scenario of "sudden strong maneuvering under sparse observation" to simulate the extreme case of radar tracking a highly maneuvering target in low data rate mode, as follows: Target kinematics setting: Initial state: The target starts from the origin. Start, initial speed is , ,along It flies at high speed in the positive direction of the axis.

[0034] Exercise phase: Stable phase ( The target maintains uniform linear motion (CV model) to simulate cruise.

[0035] Sharp turn phase ( The target suddenly makes an extremely strong coordinated turn (CT model), with the turning angular velocity set to... (about ).

[0036] Emergency stop phase ( The target speed decreases to 10% of its original speed with each step.

[0037] Challenges of the scenario: high turning speed, rapid speed decay during emergency stops, and large radar sampling intervals. The target's displacement direction changes dramatically between two frames (with extremely high nonlinearity), making it difficult for traditional linear prediction-based filters to capture this sudden change.

[0038] Observation environment settings: Sensor location: The radar is deployed at coordinates At a certain point, there is a certain geometric angle between the target trajectory and the target trajectory, so as to introduce a nonlinear measurement relationship.

[0039] Sparse observation: The radar sampling interval is set to This means that within the 0.5-second blind zone when the target makes a sharp turn, the filter receives no data updates and must rely entirely on predictive capabilities.

[0040] Measurement noise: To simulate real sensor error, the standard deviation of the ranging is set to... The standard deviation of the angle measurement is set as .

[0041] For comparative verification, two sets of control parameters were set up for the method in this application (dual-modal particle flow PFF) and a related technique (traditional mean gate filtering). Except for the core algorithm logic, the basic environmental parameters of both are completely identical, as shown in the table below:

[0042] This application includes simulation result analysis, and the scene motion trajectory is as follows: Figure 3 As shown, in Figure 3 The blue legend represents the traditional gate screening correlation method. When the target begins a sharp turn, due to the large sampling interval, the actual target position deviates significantly from the predicted position based on linear extrapolation. Because the traditional method's gate is constructed based on the predicted covariance, it cannot cover the actual position after the abrupt change, causing the filter to refuse to update. The tracking trajectory then flies off in a straight line along the tangential direction, resulting in complete loss of the target. Subsequently... At the moment of sudden braking, since the target has been completely lost, there is no way to respond to the target's stopping action. The red diagram illustrates the method used in this application. Thanks to the wide-area coverage mechanism of the "exploration particles," some exploration particles successfully capture the actual observations that deviate from the predicted mean at the instant of the turn. After successful correlation is determined through particle-level soft-gating, the particle flow algorithm quickly corrects the overall particle distribution. As shown in the figure, the red trajectory closely follows the actual turning path and... It then accurately converged near the stopping point, achieving stable tracking without any breaks throughout the entire process.

[0043] Gate correlation indicators such as Figure 4 As shown, in Figure 4 The blue legend represents the traditional gate screening correlation method. At the first observation point afterward, the Normalized Innovation Squared (NIS) spiked to over 150, far exceeding the threshold. The red legend illustrates the method used in this application. Throughout the tracking process, including sharp turns and sudden stops, the "Nearest Neighbor Particle Normalized Distance" calculated in this application consistently remained below the threshold (close to 0). This demonstrates that even if the predicted mean deviates, there are always effective particles in the exploratory particle swarm constructed in this application that can interpret the observation data, thus ensuring the continuity of data correlation.

[0044] Position tracking accuracy RMSE such as Figure 5 As shown, in Figure 5 The blue legend represents the traditional gate screening correlation method. After the turn, the RMSE diverged exponentially, rapidly exceeding 50m, indicating a complete failure of tracking. The red legend represents the method of this application. During sharp turns, the RMSE showed only minor fluctuations (approximately 5m) before rapidly converging. During emergency braking, the RMSE remains low (around 5m) despite the drastic change in velocity magnitude.

[0045] This application can solve the deadlock problem caused by overconfidence, because in related technologies, the longer the filter runs, the lower the covariance. The smaller the value, the more resistant to mutations. This application utilizes measurement noise. (The sensor's inherent properties mean it will not shrink over time) serve as a benchmark, and combined with the exploration of particles, it maintains constant sensitivity to sudden maneuvers.

[0046] This application significantly improves robustness under sparse observations. Experiments demonstrate that, At low refresh rates, traditional methods are completely ineffective for 90-degree turns, while this application can achieve breakpoint-free tracking.

[0047] This application is computationally efficient. Unlike the IMM algorithm, it does not require maintaining multiple filter models. It can cover multiple motion modes simply by adjusting the particle noise distribution, making it suitable for real-time engineering implementation. The particle flow filtering method is more suitable for GPU parallel processing.

[0048] It is worth noting that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0049] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of this application and should not be construed as limiting the specific implementation of this application to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of this application, and all such modifications or substitutions should be considered within the scope of protection of this application.

Claims

1. A method for maneuvering target association and tracking under sparse observation based on bimodal particle flow, characterized in that, include: S100: Acquire the observation values ​​obtained from radar observation of the maneuvering target and set the maneuvering range of multiple particles dispersed to cover the maneuvering target; S200, the particles within the maneuvering range are divided into core particles and exploration particles, and a one-step state prediction is performed on all particles so that the exploration particles disperse in all directions and the core particles maintain their inertia to obtain the current state of the particles. S300, based on the normalized measurement distance between the observed value and each particle in the current state, perform an association determination on the observed value to obtain an association determination result; S400, Update the positions of all particles according to the association determination result to obtain the updated position coordinates of all particles; S500 determines the position of the maneuvering target based on the updated positions of all particles after filtering.

2. The method according to claim 1, wherein, S200 includes: S210, a particle group composed of particles within the maneuvering range dividing the core particle group and exploring the particle group ; wherein the number of core particles contained in the core particle group is , and the noise thereof is micro noise ; the number of exploration particles contained in the exploration particle group is , and the noise thereof is macro noise ; S220 uses state transformation equations to predict the state of all particles, so that the exploration particles can disperse in all directions and the core particles can maintain inertia.

3. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 2, characterized in that, The state transition equation is expressed as: In the formula, Indicates the first Individual particles The state at any given moment, Represents the state transition matrix. Indicates the first Individual particles The state at any given moment, Indicates the first Individual particles The noise term at time step, adjusted according to particle type, is represented as follows: In the formula, The variance of the velocity component is set to of times.

4. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 3, characterized in that, The S300 includes: S310, calculate the normalized measurement distance between the observation and each particle in the current state, expressed as: In the formula, Indicates in Observations at time and Normalized measured distance between them Represents the nonlinear measurement function with respect to the first Predicted observations of the state of each particle. The superscript represents the inverse of the measurement noise covariance matrix. Indicates transpose; S320, based on the normalized measurement distance between the observed value and each particle in the current state, the observed value is correlated to obtain a correlation determination result.

5. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 4, characterized in that, The S320 includes: S321, Calculate the global minimum particle distance index based on the normalized measurement distance between the observed value and each particle in the current state, expressed as: S322, The observed values ​​are correlated using the global minimum distance index to obtain the correlation determination result.

6. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 5, characterized in that, S322 includes: S3221, For each particle in the current state, if the global minimum distance index calculated between the particle and the observation value is not greater than the threshold, then the association determination result of the observation value and the particle being successfully associated is obtained. S3222, if the global minimum distance index is greater than the threshold, then the observed value is obtained as the clutter observation value or the correlation determination result of the lost maneuvering target.

7. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 1, characterized in that, The S400 includes: S410, if the association determination result is that the observation value is successfully associated with the particle, then the position of all particles is updated using the observation value to obtain the updated position coordinates of all particles. S420, if the correlation determination result is the observation value of clutter or the loss of maneuvering target, the CV uniform linear motion model is used to predict and update the position of all particles to obtain the updated position coordinates of all particles.

8. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 7, characterized in that, The S410 includes: The observations and particles that are successfully associated in the association determination results are identified, and the position information of the particle is transmitted to the other particles so that they move closer to the particle to update their own positions, until the global minimum distance index between the particle and the observation is less than the threshold, and the updated position coordinates of all particles are obtained.

9. The method for correlation and tracking of maneuvering targets based on sparse observations of dual-modal particle flow according to claim 8, characterized in that, Global particles update their positions using a homotopic flow equation, which is expressed as: In the formula, This is a pseudo-time step, ranging from 0 to 1. This represents the flow gain matrix at the pseudo-time step. Represents a nonlinear measurement function. This represents the observed value at the current moment. This represents the particle state vector.

10. The method for correlation and tracking of maneuvering targets under sparse observation based on dual-modal particle flow according to claim 1, characterized in that, The S500 includes: The mean coordinates are obtained by averaging the coordinates of all updated particle positions, and this mean coordinate is used as the position of the maneuvering target estimated by filtering.

Images