Decoupling and parallel computing method of kalman filter state for multi-target tracking

Abstract

The application discloses a Kalman filtering state decoupling and parallel computing method for multi-target tracking, and comprises the following steps: S1, generating several two-dimensional Kalman subsystems according to the original high-dimensional state space and column vectors of Kalman filtering hardware; S2, constructing two-dimensional Kalman filtering processing units for each two-dimensional Kalman subsystem, and outputting a high-dimensional state vector; and S3, generating a high-dimensional target state estimation result according to the high-dimensional state vector. When the multi-target tracking algorithm is applied, the parallel two-dimensional Kalman filtering processing array is simultaneously executed in the FPGA, and the processing delay of single filtering is greatly shortened. The soft and hard collaborative architecture completely sinks the heavy matrix operation to the FPGA, significantly reduces the CPU load, enables the CPU to focus on upper logic control, and thus realizes real-time multi-target tracking with high frame rate and low delay on the end-side heterogeneous platform.

Description

Technical Field

[0001] This invention belongs to the field of filtering technology, specifically relating to a Kalman filter state decoupling and parallel computation method for multi-target tracking. Background Technology

[0002] With the development of intelligent sensing and visual analysis technologies, multi-target tracking algorithms have been widely applied in video surveillance, intelligent transportation, and robot perception. Existing multi-target tracking algorithms typically employ Kalman filters to recursively estimate the target's motion state, enabling continuous prediction and updating of the target trajectory.

[0003] In a typical multi-target tracking framework, each valid target trajectory needs to independently maintain a set of Kalman filter states to describe the target's dynamic changes in spatial position, scale, and velocity. As the number of targets increases, the computational load of Kalman filtering grows linearly with the number of targets, becoming a significant computational expense in the tracking system.

[0004] In existing technologies, Kalman filtering is typically based on high-dimensional state vectors for overall modeling and computation. For example, in target tracking algorithms based on linear uniform motion models, high-dimensional state vectors containing position, scale, and velocity components are often used, and prediction and update operations are performed through a unified state transition matrix and covariance matrix. However, such high-dimensional Kalman filtering implementations require frequent large-scale matrix multiplication and matrix update operations, resulting in complex computational structures and high resource consumption.

[0005] In edge or embedded multi-target tracking applications, systems are typically deployed on resource-constrained heterogeneous computing platforms. The CPU, as the main control unit, must simultaneously handle system scheduling, communication management, and upper-level control logic. To alleviate CPU load and improve real-time performance, existing solutions attempt to accelerate Kalman filtering by offloading it to hardware such as FPGAs. However, most of these solutions utilize a high-dimensional overall computational structure. Directly mapping this centralized high-dimensional matrix operation to the FPGA consumes significant amounts of DSP slices and storage resources. Furthermore, resource overhead increases rapidly and non-linearly with the number of targets, limiting system scalability. Some existing solutions employ approximation methods such as sparse matrix approximation, low-precision quantization, or simplified motion models to reduce computational load. While these methods reduce the load to some extent, they compromise filtering consistency, leading to decreased tracking accuracy and failing to meet the demands of high-precision tracking.

[0006] Existing implementations ignore the inherent structural independence between state components in the target motion model and force high-dimensional coupled calculations, resulting in a large number of invalid "zero element" operations in the hardware, which wastes computing resources. Summary of the Invention

[0007] To address the problems of high computational complexity, low hardware resource utilization, and decreased tracking accuracy caused by approximate optimization in the deployment of Kalman filtering on heterogeneous end-side platforms in the prior art, this invention proposes a Kalman filtering state decoupling and parallel computation method for multi-target tracking.

[0008] The technical solution of this invention is: a Kalman filter state decoupling and parallel computation method for multi-target tracking, comprising the following steps:

[0009] S1. Based on the original high-dimensional state space and column vectors of the Kalman filter hardware, generate several two-dimensional Kalman subsystems;

[0010] S2. Construct two-dimensional Kalman filtering units for each two-dimensional Kalman subsystem and output high-dimensional state vectors;

[0011] S3. Generate high-dimensional target state estimation results based on the high-dimensional state vector.

[0012] Furthermore, S1 includes the following sub-steps:

[0013] S11. The original high-dimensional state space of the Kalman filter hardware is rearranged in pairs according to the position components and their corresponding velocity components to obtain the rearranged state transition matrix.

[0014] S12. Determine the column vector;

[0015] S13. Based on the rearranged state transition matrix and column vectors, generate several two-dimensional Kalman subsystems.

[0016] Furthermore, in S12, the column vector The expression is:

[0017] ;

[0018] in, The x-coordinate of the target's center. Represents the center ordinate of the target. Indicates the aspect ratio. Indicates altitude, This represents the rate of change of velocity corresponding to the x-coordinate of the target's center. This represents the rate of change of velocity corresponding to the center ordinate of the target. This represents the rate of change of velocity corresponding to the aspect ratio. This represents the rate of change of velocity corresponding to altitude.

[0019] Furthermore, in S13, the expression for the two-dimensional Kalman subsystem is:

[0020] ;

[0021] in, The x-coordinate of the target's center. Represents the center ordinate of the target. Indicates the aspect ratio. Indicates altitude, This represents the rate of change of velocity corresponding to the x-coordinate of the target's center. This represents the rate of change of velocity corresponding to the center ordinate of the target. This represents the rate of change of velocity corresponding to the aspect ratio. This represents the rate of change of velocity corresponding to altitude.

[0022] Furthermore, in S2, the timing phases of the two-dimensional Kalman filter processing unit include the idle phase, data loading and policy branching, posterior update phase, update preparation state, division handshake state, and multiplication start state.

[0023] Furthermore, the data loading and policy branch is used to perform prior predictions on the sub-state vectors, and the expression for updating the prior covariance matrix of the prior prediction is:

[0024] ;

[0025] ;

[0026] in, Represents the state transition matrix. Indicates time interval, Represents the prior state covariance matrix. Let the posterior state covariance matrix at the previous time k-1 be represented. Indicates transpose. Represents the process noise covariance matrix;

[0027] The expressions for each element of the prior covariance matrix are as follows:

[0028]

[0029] in, This represents the (0,0)th element of the predicted state covariance matrix. This represents the variance of the position state at the previous moment. This represents the covariance of the error between the position and velocity at the previous moment. This represents the variance of the velocity state at the previous moment. This represents the noise component corresponding to the position dimension in the process noise covariance matrix. This represents the error covariance between the predicted position and velocity. This represents the variance of the predicted velocity state error. This represents the noise component corresponding to the velocity dimension in the process noise covariance matrix.

[0030] Furthermore, the prepared state is updated to calculate the observation residuals. and innovation covariance Their expressions are as follows:

[0031] ;

[0032] ;

[0033] in, Represents the observation vector. This represents the prior state estimation vector. Represents the prior state covariance matrix. Indicates transpose. Represents the observation matrix. This represents the observation noise covariance matrix.

[0034] Furthermore, in S3, the high-dimensional state vector is reorganized to generate a high-dimensional target state estimation result. Specifically, the low-dimensional posterior state estimate and the covariance matrix are spliced ​​and reorganized according to the dimensional definition order of the original high-dimensional state space to generate a high-dimensional target state estimation result.

[0035] The beneficial effects of this invention are:

[0036] (1) The state-decoupled parallel Kalman filtering method proposed in this invention utilizes the block diagonal sparsity of the state transition matrix in the target motion model to losslessly decompose the high-dimensional state space into multiple low-dimensional subsystems through the permutation matrix, thereby reducing the computational complexity from Reduced to This fundamentally eliminates invalid zero-element operations. In FPGA implementations, this significantly reduces the usage of multipliers and BlockRAM, resulting in a substantial increase in the number of concurrent tracking channels that can be supported with the same on-chip resources.

[0037] (2) Based on the independence of the physical motion model, the present invention performs mathematical equivalent transformation, which ensures that the output posterior state and covariance matrix are highly consistent with the original high-dimensional Kalman filter, thus meeting the high-precision tracking requirements. Through the efficient control of the designed state machine, the regulation of resources and computing power is realized, avoiding the introduction of irregular memory access patterns or complex dynamic sparse detection mechanisms and the accuracy loss and control complexity caused by approximate optimization in the existing technology.

[0038] (3) When the multi-target tracking algorithm is applied, the parallel two-dimensional Kalman filter processing array is executed simultaneously in the FPGA, which greatly shortens the processing delay of a single filter. This hardware-software co-processing architecture completely sinks the heavy matrix operation to the FPGA, significantly reducing the CPU load and allowing it to focus on upper-level logic control, thereby realizing high frame rate and low latency real-time multi-target tracking on the edge heterogeneous platform. Attached Figure Description

[0039] Figure 1 A flowchart of a Kalman filter state decoupling and parallel computation method for multi-target tracking;

[0040] Figure 2 Hardware architecture diagram for Kalman 8-dimensional matrix decomposition;

[0041] Figure 3 A dimensionality reduction diagram for the KF state vector prediction stage;

[0042] Figure 4 A dimension-reduced graph of the covariance matrix permutation;

[0043] Figure 5 Split the noise matrix into a dimensionality-reduced graph;

[0044] Figure 6 This is a Kalman 8-dimensional matrix decomposition diagram;

[0045] Figure 7 This is a two-dimensional Kalman state machine diagram;

[0046] Figure 8 This is a schematic diagram of the data flow during the prediction and update phases. Detailed Implementation

[0047] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0048] like Figure 1 As shown, this invention provides a Kalman filter state decoupling and parallel computation method for multi-target tracking, comprising the following steps:

[0049] S1. Based on the original high-dimensional state space and column vectors of the Kalman filter hardware, generate several two-dimensional Kalman subsystems;

[0050] S2. Construct two-dimensional Kalman filtering units for each two-dimensional Kalman subsystem and output high-dimensional state vectors;

[0051] S3. Generate high-dimensional target state estimation results based on the high-dimensional state vector.

[0052] This invention relates to a state-decoupling Kalman filtering method for multi-target tracking. The method includes a state-decoupling modeling approach and a sub-state filtering computation method. The former utilizes the block diagonal sparsity of the state transition matrix in a linear uniform motion model of the target motion model, and losslessly decomposes the high-dimensional target state vector and covariance matrix into multiple independent low-dimensional subsystems through matrix permutation transformation. The latter, based on this, designs a dedicated parallel computing array driven by a finite state machine (FSM). Each low-dimensional subsystem performs prediction and update calculations in parallel in an independent hardware pipeline channel, and outputs a complete high-dimensional state estimate through state recombination at the output stage. While strictly ensuring the mathematical consistency between the filtering result and the original high-dimensional Kalman filter, this method significantly reduces matrix operation complexity and on-chip resource consumption, effectively improving the parallel processing efficiency and real-time performance of heterogeneous computing platforms.

[0053] In this embodiment of the invention, S1 includes the following sub-steps:

[0054] S11. The original high-dimensional state space of the Kalman filter hardware is rearranged in pairs according to the position components and their corresponding velocity components to obtain the rearranged state transition matrix.

[0055] S12. Determine the column vector;

[0056] S13. Based on the rearranged state transition matrix and column vectors, generate several two-dimensional Kalman subsystems.

[0057] This invention focuses on two core issues: high-dimensional state physical separability modeling and sub-state parallel filtering architecture. It presents a hardware acceleration scheme for Kalman filtering that deeply integrates algorithm modeling and hardware architecture. First, based on the physical characteristics of a linear uniform motion model, a permutation matrix is ​​used to losslessly decouple the high-dimensional state space. Then, a dedicated two-dimensional Kalman filtering processing unit based on a finite state machine (FSM) is constructed. Through a multi-stage pipelined parallel scheduling mechanism, high-throughput, low-latency parallel computation is achieved on an FPGA.

[0058] Firstly, in multi-target tracking algorithms such as ByteTrack, Kalman filtering is based on a linear uniform motion model. In its original 8-dimensional state vector, the target's center x-coordinate, center y-coordinate, aspect ratio, height, and the corresponding velocity change rates of these four components constitute an 8-dimensional column vector. Most existing implementations treat this 8-dimensional state as a whole for modeling and calculation, inevitably introducing large-scale matrix operations and complex data dependencies in hardware implementation.

[0059] To address the aforementioned issues, this invention, starting from the physical separability of target motion, systematically analyzes the dynamic coupling relationships between state dimensions under a linear uniform velocity model. Research shows that, under this model assumption, the target's motion processes in the horizontal, vertical, and scale-change directions are independent of each other. Each state component is dynamically coupled only with its corresponding velocity component, and there is no cross-influence between different dimensions. The remaining initialization matrix vectors also exhibit similar characteristics; the state transition matrix F, process noise matrix Q, and observation noise matrix R are initially presented as block diagonal structures. Based on the above transformation, the 8-dimensional Kalman filter system is losslessly decomposed into four independent 2-dimensional subsystems (e.g., and (This constitutes a subsystem). Each subsystem only needs to process 2... 2-dimensional matrix operations thus avoid directly processing 8 8. The computational complexity and inversion problem brought about by large-scale matrices.

[0060] In this embodiment of the invention, in S12, the column vector The expression is:

[0061] ;

[0062] in, The x-coordinate of the target's center. Represents the center ordinate of the target. Indicates the aspect ratio. Indicates altitude, This represents the rate of change of velocity corresponding to the x-coordinate of the target's center. This represents the rate of change of velocity corresponding to the center ordinate of the target. This represents the rate of change of velocity corresponding to the aspect ratio. This represents the rate of change of velocity corresponding to altitude.

[0063] In this embodiment of the invention, in S13, the expression of the two-dimensional Kalman subsystem is:

[0064] ;

[0065] in, The x-coordinate of the target's center. Represents the center ordinate of the target. Indicates the aspect ratio. Indicates altitude, This represents the rate of change of velocity corresponding to the x-coordinate of the target's center. This represents the rate of change of velocity corresponding to the center ordinate of the target. This represents the rate of change of velocity corresponding to the aspect ratio. This represents the rate of change of velocity corresponding to altitude.

[0066] In this embodiment of the invention, the timing phases of the two-dimensional Kalman filter processing unit in S2 include an idle phase, a data loading and policy branching phase, a post-update phase, an update preparation state, a division handshake state, and a multiplication start state.

[0067] After achieving lossless decoupling of high-dimensional states, this invention further proposes a sub-state parallel filtering method tailored to the parallel computing characteristics of FPGAs. Unlike traditional implementations based on general matrix operation instruction sets, this invention abandons the general computing architecture and designs a dedicated two-dimensional Kalman filtering processing unit (Processing Element, PE) for the fixed computing mode of the two-dimensional Kalman subsystem, using it as the basic building block of the parallel computing array.

[0068] Each processing unit is driven by a specific finite state machine, mapping the standard Kalman filtering algorithm into five discrete temporal phases: Idle (IDLE); Data Loading and Policy Branching (PRED); A posteriori Update (UPDATE); Update Ready (UPD1); Division Handshake (WAIT_DIV); and Multiplication Start (UPD2). Each phase corresponds to a specific mathematical operation in Kalman filtering.

[0069] In the Pred stage, the processing unit processes the sub-state vectors based on a two-dimensional linear uniform motion model. Perform prior prediction,

[0070] In this embodiment of the invention, the data loading and policy branch are used to perform prior prediction on the sub-state vector, and the prior covariance matrix update expression for the prior prediction is:

[0071] ;

[0072] ;

[0073] in, Represents the state transition matrix. Indicates time interval, Represents the prior state covariance matrix. Let the posterior state covariance matrix at the previous time k-1 be represented. Indicates transpose. Represents the process noise covariance matrix;

[0074] For the matrix calculations described above, this embodiment does not use a general matrix multiplier. Instead of matrix multiplication, it utilizes the state transition matrix. Only the diagonal elements and certain non-zero elements have a value of 1 or 0. Based on the sparse structure characteristics, the covariance prediction formula is expanded and reconstructed.

[0075] The expressions for each element of the prior covariance matrix are as follows:

[0076] ;

[0077] in, This represents the (0,0)th element of the predicted state covariance matrix. This represents the variance of the position state at the previous moment. This represents the covariance of the error between the position and velocity at the previous moment. This represents the variance of the velocity state at the previous moment. This represents the noise component corresponding to the position dimension in the process noise covariance matrix. This represents the error covariance between the predicted position and velocity. This represents the variance of the predicted velocity state error. This represents the noise component corresponding to the velocity dimension in the process noise covariance matrix.

[0078] No multiplication and addition operations related to zero elements need to be calculated, thus significantly reducing the number of multipliers and operation clock cycles during the prediction phase.

[0079] In this embodiment of the invention, during the update preparation phase (UPD1), the processing unit calculates the observation residuals and innovation covariance, including the residual vector and the innovation covariance.

[0080] Update the prepared state for calculating the observation residuals. and innovation covariance Their expressions are as follows:

[0081] ;

[0082] ;

[0083] in, Represents the observation vector. This represents the prior state estimation vector. Represents the prior state covariance matrix. Indicates transpose. Represents the observation matrix. This represents the observation noise covariance matrix.

[0084] Where the observation matrix The corresponding calculation process is simplified to the selection and accumulation of a single element in the prior covariance matrix.

[0085] In this embodiment of the invention, in S3, the high-dimensional state vector is reorganized to generate a high-dimensional target state estimation result. Specifically, the low-dimensional posterior state estimate and the covariance matrix are spliced ​​and reorganized according to the dimensional definition order of the original high-dimensional state space to generate a high-dimensional target state estimation result.

[0086] Overall hardware architecture such as Figure 2 As shown, the system comprises four parallel two-dimensional Kalman sub-filter modules. Each module obtains its corresponding sub-state input through data routing logic and integrates a dedicated multiply-accumulate arithmetic unit and a pipelined divider interface. Each sub-module can start computation simultaneously within a single cycle. The final low-dimensional output results are summarized and recombined in the OutputCombination of the top-level module to output a complete high-dimensional state vector. Furthermore, the data between modules is completely independent, and all four modules are executed on an FPGA. The state recombination output refers to the process where, after all computation channels have completed filtering updates, the low-dimensional posterior state estimates and covariance matrices output by each channel are concatenated and recombined strictly according to the dimensional definition order of the original high-dimensional state space. This results in a high-dimensional target state estimate with complete motion information, used for subsequent trajectory association and lifecycle management.

[0087] The following description is based on specific embodiments.

[0088] To address the challenge of parallel implementation of high-dimensional Kalman filtering on embedded hardware platforms such as FPGAs, this invention analyzes the physical coupling relationships between state variables in a linear uniform motion model. The original high-dimensional state space is rearranged in pairs according to the position components and their corresponding velocity components. This results in a block diagonal form for the rearranged state transition matrix, where each dimension of the 8-dimensional column vector is dynamically coupled only to its corresponding velocity component. Based on the above discussion and state equation formulas, the state components and their velocities constitute an independent subsystem, thus enabling the prediction stage to become... Figure 3 As shown.

[0089] The covariance matrix, process noise matrix, and observation noise matrix corresponding to the state vector are all synchronously rearranged according to the same dimensional mapping relationship. After this processing, the above matrices are structurally decomposed into multiple low-dimensional submatrices, as shown below. Figure 4 and Figure 5 As shown. This synchronous rearrangement ensures that each two-dimensional subsystem is mathematically strictly equivalent to the original high-dimensional system, while eliminating matrix coupling terms across subsystems.

[0090] The decomposition of the observation matrix and observation noise is the same as the technical details described above, and therefore will not be repeated. Based on this, this invention uses a state decoupling method to rearrange and decompose the original 8-dimensional Kalman filter (KF) into four independent 2-dimensional subsystems. This reconstruction method, while maintaining the consistency of the original Kalman filter estimation results, explicitly eliminates the computational dependencies between cross-dimensional states, forming a system-level decoupling structure oriented towards hardware implementation.

[0091] Based on this decoupled system architecture, this invention further maps the four two-dimensional subsystems onto four independent Kalman filter processing units on an FPGA platform to achieve parallel execution, such as... Figure 6 As shown, this parallel mapping method significantly reduces the computational complexity and data interaction overhead of a single processing unit, providing a feasible path for high-throughput hardware implementation in multi-target real-time tracking scenarios.

[0092] On the FPGA side, each KF2-dimensional sub-cell contains a dedicated data path driven by an FSM, and the specific process is as follows:

[0093] As a fundamental building block of parallel computing arrays, the design of the two-dimensional Kalman sub-filter directly determines the resource efficiency and tracking accuracy of the overall architecture. This paper abandons the general matrix operation instruction set architecture and instead adopts an application-specific instruction set design approach, constructing a dedicated data path driven by a finite state machine. The standard Kalman filter algorithm is mapped to five discrete time-series stages: idle, data loading, prior prediction, Kalman gain calculation, and posterior update. The state machine is as follows: Figure 7 As shown.

[0094] The computational design employs a multi-stage pipeline scheduling strategy. During the system's standby phase, the input register set is used to cache the state vector from the previous time step. Covariance matrix and the current observation scalar The Pred phase is the policy branch point of the state machine. It determines whether there is an observation update based on the value of do_update. When there is no observation input, the system only performs prior prediction calculation and returns directly, thereby avoiding unnecessary update calculation and reducing dynamic power consumption. When there is an observation update, it sequentially enters the subsequent gain calculation and state update phases.

[0095] The Update phase is divided into two states: UPD1 (update preparation state, WAIT_DIV division handshake state) and UPD2 (multiplication start state). UPD1 calculates the intermediate variables required for the Kalman gain, waits for the custom trigger to complete the calculation (since the gain needs to be calculated twice), and then uses the obtained gain to calculate the update equation. The data flow for the prediction and update phases is as follows: Figure 8 As shown.

[0096] 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 method for state decoupling and parallel computation of Kalman filtering in multi-target tracking, characterized in that, Includes the following steps: S1. Based on the original high-dimensional state space and column vectors of the Kalman filter hardware, generate several two-dimensional Kalman subsystems; S2. Construct two-dimensional Kalman filtering units for each two-dimensional Kalman subsystem and output high-dimensional state vectors; S3. Generate high-dimensional target state estimation results based on the high-dimensional state vector.

2. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 1, characterized in that, S1 includes the following sub-steps: S11. The original high-dimensional state space of the Kalman filter hardware is rearranged in pairs according to the position components and their corresponding velocity components to obtain the rearranged state transition matrix. S12. Determine the column vector; S13. Based on the rearranged state transition matrix and column vectors, generate several two-dimensional Kalman subsystems.

3. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 2, characterized in that, In S12, the column vector The expression is: ； in, The x-coordinate of the target's center. Represents the center ordinate of the target. Indicates the aspect ratio. Indicates altitude, This represents the rate of change of velocity corresponding to the x-coordinate of the target's center. This represents the rate of change of velocity corresponding to the center ordinate of the target. This represents the rate of change of velocity corresponding to the aspect ratio. This represents the rate of change of velocity corresponding to altitude.

4. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 2, characterized in that, In S13, the expression for the two-dimensional Kalman subsystem is: ； in, The x-coordinate of the target's center. Represents the center ordinate of the target. Indicates the aspect ratio. Indicates altitude, This represents the rate of change of velocity corresponding to the x-coordinate of the target's center. This represents the rate of change of velocity corresponding to the center ordinate of the target. This represents the rate of change of velocity corresponding to the aspect ratio. This represents the rate of change of velocity corresponding to altitude.

5. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 1, characterized in that, In S2, the timing phases of the two-dimensional Kalman filter processing unit include an idle phase, data loading and strategy branching, posterior update phase, update preparation state, division handshake state, and multiplication start state.

6. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 5, characterized in that, The data loading and policy branch is used to perform prior prediction on the sub-state vector, and the prior covariance matrix update expression for the prior prediction is: ； ； in, Represents the state transition matrix. Indicates time interval, Represents the prior state covariance matrix. Let the posterior state covariance matrix at the previous time k-1 be represented. Indicates transpose. Represents the process noise covariance matrix; The expressions for each element of the prior covariance matrix are as follows: in, This represents the (0,0)th element of the predicted state covariance matrix. This represents the variance of the position state at the previous moment. This represents the covariance of the error between the position and velocity at the previous moment. This represents the variance of the velocity state at the previous moment. This represents the noise component corresponding to the position dimension in the process noise covariance matrix. This represents the error covariance between the predicted position and velocity. This represents the variance of the predicted velocity state error. This represents the noise component corresponding to the velocity dimension in the process noise covariance matrix.

7. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 5, characterized in that, The updated readiness state is used to calculate the observation residuals. and innovation covariance Their expressions are as follows: ； ； in, Represents the observation vector. This represents the prior state estimation vector. Represents the prior state covariance matrix. Indicates transpose. Represents the observation matrix. This represents the observation noise covariance matrix.

8. The Kalman filter state decoupling and parallel computation method for multi-target tracking according to claim 1, characterized in that, In step S3, the high-dimensional state vector is reorganized to generate a high-dimensional target state estimation result. Specifically, the low-dimensional posterior state estimate and the covariance matrix are spliced ​​and reorganized according to the dimensional definition order of the original high-dimensional state space to generate a high-dimensional target state estimation result.

Images