Track-before-detect algorithm based on multi-feature knowledge-assisted particle filtering

Abstract

The invention relates to the technical field of underwater target detection and tracking, in particular to a track-before-detect algorithm based on multi-feature knowledge-assisted particle filtering, which can effectively cope with frogman detection and tracking tasks under different sea conditions and noise conditions and remarkably improve the marine safety monitoring capability. The method comprises the following steps of 1, initializing a target state and sonar observation parameters; 2, constructing a frogman multidirectional motion and state transition model; and step 3, particle filtering based on multi-feature assistance. According to the method, a frogman scattering model and a multidirectional motion model are combined as priori knowledge to optimize a particle state transition process, and particle filtering and a track-before-detect algorithm are combined, so that the detection and tracking precision of the underwater frogman is effectively improved. The method is especially suitable for real-time detection and tracking of underwater frogmen and small moving targets similar to the frogmen.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater target detection and tracking, and relates to a tracking-before-detection algorithm based on multi-feature knowledge-assisted particle filter (MFKA-PF-TBD) and its application in diver detection and tracking. The algorithm proposed by the present invention is particularly applicable to complex and changeable underwater environments, can effectively cope with diver detection and tracking tasks under different sea conditions and noise conditions, and significantly improves the maritime safety monitoring ability. Background Art

[0002] Underwater small target detection and tracking algorithms play an important role in the field of ocean exploration, and their development process has experienced a transformation from traditional signal processing methods to modern deep learning technologies. Traditional methods such as Kalman Filter (KF) and Particle Filter (PF) are widely used in target tracking due to their simple theoretical basis and implementation. However, Kalman Filter is mainly applicable to linear systems and is difficult to give accurate estimates for non-linear dynamic changes and multi-target mixed motion scenarios; while particle filter can better handle non-linear problems and estimate the target state through random sampling and weight recursion, but in complex underwater environments, when the signal echo background noise interference is severe, it often faces limitations such as high computational complexity and insufficient real-time performance. Especially for divers, due to their diverse movement patterns and drastic posture changes, the detection and tracking problems are more complex, and it is urgent to introduce multi-feature information to improve the target identification rate and tracking stability.

[0003] In response to the above problems, in recent years, a new technical route - Tracking Before Detect (TBD) has emerged in the field of underwater target detection and tracking. The traditional "detection - tracking" mode is prone to losing weak targets in real complex marine environments, while the TBD method makes a decision based on the cumulative statistic by first tracking the target trajectory, effectively alleviating the detection difficulty when the signal-to-noise ratio is low. Common TBD algorithms mainly include Dynamic Programming TBD (DP-TBD), Hough Transform TBD (HF-TBD), and Particle Filter TBD (PF-TBD), etc.

[0004] The DP-TBD algorithm uses dynamic programming to optimize the multi-target tracking problem, and is committed to improving the detection and tracking accuracy in complex environments. Although the algorithm optimizes the discretization problem of the multi-target state space, it has the problem of large computational complexity. The HF-TBD algorithm uses the Hough transform to map the target trajectory to the parameter space, which is suitable for linear motion targets, but has limitations in tracking frogmen in complex curved motion. The PF-TBD algorithm uses a particle filter to simulate the target trajectory and recursively estimate the target state. It shows high flexibility in nonlinear and multi-target scenarios, but still needs to be improved in dealing with non-Gaussian noise and extremely low signals.

[0005] Based on this, the present invention proposes a tracking before detection algorithm based on multi-feature knowledge-assisted particle filtering (MFKA-PF-TBD). Aiming at the complexity and variability of frogman's movement, the algorithm makes full use of the target's motion characteristics and scattering characteristics as knowledge-assisted information to achieve a more accurate and robust tracking effect. At the same time, through the idea of tracking before detection, using cumulative trajectory statistics and threshold decision mechanism, the problem of easy failure of traditional detection algorithms is effectively solved. Summary of the invention

[0006] The purpose of this invention is to consider the motion characteristics and scattering characteristics of frogmen, and propose a detection-pre-tracking algorithm based on multi-feature knowledge assisted particle filtering (MFKA-PF-TBD). The algorithm combines the frogman scattering model and the multi-directional motion model as prior knowledge to optimize the particle state transfer process, combines the particle filter with the detection-pre-tracking algorithm, and improves the detection and tracking accuracy of underwater small targets. The specific steps are as follows:

[0007] Step 1: Initialization of target state and sonar observation parameters

[0008] This step lays the foundation for the entire tracking process and determines the state model of the frogman target and the parameters of the sonar measurement model. By clarifying the position, speed and measurement relationship of the target, a solid foundation is ensured for the subsequent signal processing and target state prediction.

[0009] First, determine the target state and sonar measurement model parameters. Assume that the target is in the far field and its position coordinates are The corresponding speeds are and The target state is thus determined to be

[0010]

[0011] In terms of measurement relationships, sonar obtains target information by emitting sound waves and receiving their reflected signals. Ideally, when the effects of noise are ignored, the distance and angle measurements of sonar are closely related to the target position.

[0012]

[0013] The sonar system consists of a transmitting array element and a uniform linear array. The received signal r(t) is represented in vector form. The transmitted signal s(t) propagates in water and interacts with the target to generate a reflected signal. The expression of the received signal vector is

[0014]

[0015] In this expression is a vector, and its elements reflect the phase differences of the signals received by different array elements; τ m (θ t ) = (m - 1)dsinθ t / c takes into account the propagation time delay caused by different target positions and array element angles when sound waves propagate in water, where c is the speed of sound; A is the reflection coefficient of the target, which characterizes the reflection ability of the target to sound waves; ω is the carrier angular frequency, which determines the frequency characteristics of the signal; represents the total propagation time delay from transmission to the target and then to reception; v(t) is a vector representing noise, including various interference factors such as environmental noise and electronic noise.

[0016] Usually, for the convenience of digital signal processing, the complex carrier does not carry useful information. Therefore, in practical applications, mainly the complex baseband signal is considered. At this time, the discrete form of the received signal is

[0017] r(k) = a(θ k )As(k - τ tt+tr ) + v(k)

[0018] where k represents the k-th sampling time.

[0019] In addition, the measurement area has a clear definition. Taking the sonar position as the origin, the observed area of interest is limited in distance within the range of [R min , R max , and this range is further divided into N r distance cells; in azimuth, it is limited within the range of [θ min , θ max , and is also divided into N b azimuth cells. The number of distance cells N r is

[0020]

[0021] where F s is the sampling frequency in array signal processing. N b can be determined by the direction resolution unit Δθ.

[0022] At this time, the measurement unit Δθ contains N at time k r ×N b data

[0023]

[0024] where g(·) is the mapping between the signal y(k) and the measurement, and its form affects the specific form of the measurement. And w k is the measured value of the noise and clutter of the system at time k

[0025] Step 2: Construction of the multi-directional movement and state transition model for frogmen

[0026] This step describes the movement trajectory of the frogman target, establishes a multi-directional movement model to simulate its complex movement behavior underwater, and ensures the accuracy of particle state transition. By reasonably setting the movement models in different directions and speed ranges, the actual movement characteristics of the frogman can be better reflected

[0027] Set the parameters of the target movement model and construct a multi-directional movement model. This model consists of 8 directions and 16 uniform linear movements, as Figure 1 shown

[0028] In the representation of the model, the position of the frogman target at the current moment is used as the reference point and marked with a circle. Eight arrows extend from this point, and the direction pointed by each arrow represents the possible movement direction of the frogman target. These eight directions are not randomly set, but are based on an in-depth analysis of the frogman's movement behavior underwater. In each direction, two uniform linear movement models are further subdivided. The speeds of these two models are set to 0.75 times and 1.25 times the speed range of the target at the previous moment respectively. This speed setting method fully considers the variation law of the frogman's movement speed. In the actual underwater environment, although the frogman's speed changes little, there is still a certain fluctuation range. By setting models with different speeds, the change of the frogman's movement state can be simulated more carefully. Assume that the speed range of the frogman target at the previous moment is [v min , v max , then in each predicted direction, there will be uniform linear movement models with speeds from 0.75v min to 0.75v max and from 1.25v min to 1.25v max . Particles can be transferred according to these preset movement patterns, so as to more accurately reflect the real movement trajectory of the frogman target underwater

[0029] From the perspective of the mathematical model, the general expression of the target state transition is

[0030] x k+1 = f(x k, τ k ) + v k

[0031] where x k represents the state of the target at time k, τ k represents the type of the target motion model, v k is the corresponding process noise, and f(·) is the state transition matrix under different models.

[0032] For the common uniform motion model (τ k = 1), its state transition matrix is

[0033]

[0034] In this model, it is assumed that within the time interval T, the speeds of the target in the x and y directions remain unchanged, and the position update is only based on the current speed and the time interval. For example, the position update in the x direction is as follows, and the y direction is the same.

[0035]

[0036] For the coordinated turn motion model (τ k = 2), its state transition matrix is

[0037]

[0038] where w k is the coordinated turn rate, which is introduced in this model to describe the motion characteristics of the target during the turning process. The updates of the position and speed involve trigonometric operations to more accurately simulate the turning motion trajectory of the target. It is more complex than the uniform motion model but can better adapt to the target motion scenarios with turning behaviors.

[0039] The transition probabilities between different motion modes at time k are represented by a Markov chain as

[0040] P{τ k = j | τ k-1 = i} = p ij , i, j = 1, 2

[0041] In practical applications, through the analysis and statistics of the diver motion data, the transition probability matrix between different motion modes can be determined. For example, through a large number of experimental observations, it is found that the probability of a diver continuing to move in a straight line after straight-line motion is relatively high, while the probability of suddenly changing from straight-line motion to large-angle turning motion is relatively low. These probability information can be filled into the transition probability matrix to provide a basis for state prediction in the particle filtering process.

[0042] Step 3: Particle Filtering Based on Multi-Feature Aids

[0043] The purpose of this step is to track the target using the particle filter algorithm. By combining the target's motion model and measurement data, the position and state of the frogman target are estimated in real time. Through particle filtering, we can handle complex noise and interference and more accurately predict the motion trajectory of the frogman target.

[0044] First, non-parametric kernel density estimation is used to model the measurement data distribution and calculate the likelihood values of the particles to distinguish the target signal from noise interference. In the particle filtering process, the likelihood function is used to calculate the weights of the particles. By comparing the likelihood function values of the measurement data with the target states assumed by different particles, the weights of each particle can be determined, and the particles can be selected or discarded, thus realizing the estimation of the target state. Non-parametric kernel density estimation does not make any assumptions about the distribution of the measurement data, but only models the probability density function based on the sample data itself.

[0045] Suppose z1, z2, …, z n are n measurement data samples, is the kernel density estimation of the sample probability density function, then The expression of

[0046]

[0047] where n is the number of independent and identically distributed samples, K(·) is the kernel function, which determines the role of each sample data point z i , i = 1, …, n in the density estimation of the random variable z. And h is the window width that affects the smoothness of the probability density estimation.

[0048] After obtaining the statistical characteristics of the real echo data, the likelihood function of the measurement resolution unit (i, j) is

[0049]

[0050] where g1(·) represents the statistical properties of the measurement data when the target exists, and g0(·) represents the statistical properties of the measurement data when the target does not exist. p s (t i,j ) is the value of the frogman scattering echo waveform function corresponding to the resolution unit (i, j). Since the target occupies multiple resolution units in the measurement space, the likelihood function is expressed as

[0051]

[0052] where Θ is the range of the target affecting the resolution unit.

[0053] In each iteration, first generate newborn particles according to prior information or the observation area, and use the diver motion model and scattering feature model to estimate the states of the persistent particles. Then, calculate the weights of the particles by combining the statistical characteristics of the measurement data and the waveform characteristics of the diver scattering echo, and perform normalization. Calculate the mixing probability to adjust the weights of the newborn and persistent particles, perform particle resampling to form a complete particle set, and estimate the target state accordingly. The specific sub-steps are as follows:

[0054] Let the particle set be used to describe the posterior probability density at time k - 1 The iterative loop of the algorithm can be expressed as Figure 2 .

[0055] (1) Initialize particles: At the initial moment, only generate newborn particles. If the prior distribution of the target is known, generate particles according to its distribution. If there is no prior information about the target, uniformly sample the samples within the observation area.

[0056] (2) Generate new particles and estimate the states of persistent particles: According to the prior probability distribution, generate N b newborn particles, and use the established diver motion model and scattering feature model as knowledge auxiliary information to estimate the states of N c consecutive particles as follows

[0057]

[0058] (3) Calculate particle weights and normalize: Use the statistical characteristics of the measurement data and the waveform characteristics of the diver scattering echo as knowledge auxiliary information. The weight calculation and normalization of new particles and consecutive particles are as follows

[0059]

[0060] (4) Calculate the mixing probability and normalize: Calculate the mixing probability using the unnormalized weights, and then normalize, as follows

[0061]

[0062] (5) Adjust particle weights: Scale the weights of newborn and consecutive particles according to the mixing probability, as

[0063]

[0064] (6) Particle resampling and target state estimation: New particles and consecutive particles form a complete particle set, and they are resampled to get Then estimate the target state as follows

[0065]

[0066] (7) Trajectory start and end judgment: In this algorithm, target detection and tracking are realized by particles. The start and end of the trajectory are actually related to the start and end of the particles. The particles are generated by importance sampling, so the importance density can be designed to determine the start and end of the trajectory. For each target with N t frames, calculate the sum of all particle weights Σ w to be

[0067]

[0068] If the following formula is satisfied, the target start fails or ends.

[0069] Σ ω < η d

[0070] where η d is the likelihood threshold.

[0071] (8) Output the target tracking result: Generate and output the target tracking result.

[0072] Compared with the prior art, the present invention not only improves the detection and tracking accuracy of frogmen, but also has the following advantages:

[0073] (1) Strong environmental adaptability: By integrating the movement and scattering characteristics of frogmen, it can effectively resist interference in various underwater environments, stably monitor frogman targets, and break through the environmental limitations of traditional algorithms.

[0074] (2) Precise motion monitoring: Construct a multi-directional motion model and combine the scattering characteristics to calculate the state transition, accurately capture the complex motion trajectory of frogmen, and be more accurate and comprehensive than traditional algorithms.

[0075] (3) High-efficiency real-time monitoring: Optimize data processing by combining the characteristics of frogmen, and can quickly respond and adjust the tracking strategy when the target state changes, meeting the real-time requirements of maritime safety monitoring. Description of the Drawings

[0076] Figure 1 Set diagram of the multi-directional motion model of frogmen;

[0077] Figure 2 Structural diagram of particle filter based on multi-feature assistance;

[0078] Figure 3 Reference trajectory diagram of frogman motion;

[0079] Variation of three methods under different SRRs in Fig. 4 (a) RMSE under different SRRs, (b) Stable detection and tracking probability under different SRRs;

[0080] Figure 5 Results graphs of the reference trajectory and tracking trajectory obtained by three algorithms;

[0081] Figure 6 RMSE graph plotted with the simulation results against time;

[0082] Figure 7 Sequence graph of the accurate detection probability of three methods. Specific implementation manner

[0083] To implement and verify the performance of the proposed MFKA-PF-TBD method, we designed a simulation of a diving target performing complex motion in a plane.

[0084] Execute Step 1: Initialization of target state and sonar observation parameters

[0085] For our purpose, we assume that the target appears at 10 seconds and disappears at 120 seconds. The initial state is set as follows

[0086] x0 = [100, 0.5, 150, 0.5]

[0087] The reference trajectory of the diver's motion is as Figure 3 shown.

[0088] The sonar emits a linear frequency modulated (LFM) signal with a sampling interval T = 2s. The measurement range is from R min = 150m to R max = 250m; the angle range is from θ min = 20°, θ max = 80°.

[0089] Taking SNR = 10dB and SNR = -5dB as examples, the influence of the size of the sub-region on the algorithm performance is explained and how to select a suitable sub-region through the trade-off between efficiency and accuracy is presented. Here, SNR is the signal-to-noise ratio, and the smaller its value, the stronger the ocean environmental noise, corresponding to a worse sea condition. These 6 devices represent 6 sub-regions of different sizes. Six experimental scenarios with different sub-region sizes are set up, and the stable detection and tracking probabilities, position root mean square error, and calculation time are compared. The results are shown in Table 1.

[0090] Table 1 Simulation results of different settings

[0091]

[0092] It can be seen from the table that calculating the likelihood ratio using sub-regions instead of the entire region can significantly improve the calculation efficiency. The smaller the sub-region, the shorter the calculation time, but it will affect the algorithm accuracy. Generally speaking, setting 4 achieves a better balance between calculation efficiency and accuracy, and the subsequent experiments calculate the likelihood ratio based on this scenario.

[0093] Execute Step 2: Construction of the multi-directional movement and state transition model for frogmen

[0094] To ensure the fairness of algorithm comparison, the three comparison algorithms (MMPF-TBD, KA-PF-TBD, and the algorithm of the present invention) maintain the same parameter settings: the number of particles is 1500, the particle birth probability is P b = 0.85, and the particle death probability is Pd = 0.15. The birth information serves as the prior information of the target birth and follows a uniform distribution as follows

[0095]

[0096] The transition probability matrix of the common model is as follows

[0097]

[0098] The transition probability matrix of the frogman multi-directional movement model is

[0099]

[0100] Execute Step 3: Particle filter based on multi-feature assistance

[0101] 100 Monte Carlo simulation experiments are carried out. Figures 4(a) and (b) show the changes of RSME and stable detection and tracking probability of the three methods with the decrease of SNR. The MMPF-TBD algorithm has poor detection and tracking performance when SNR is lower than 5dB, because the two traditional state transition models it adopts limit the particle transfer direction, and the prediction result does not conform to the actual movement state of the frogman. The KA-PF-TBD algorithm has better performance than MMPF-TBD, because it uses the frogman multi-directional movement model to assist particles to improve the particle utilization rate. In contrast, the algorithm of the present invention has better detection and tracking performance at low SNR, which benefits from using the original measurement data as input after an efficient adaptive denoising method and combining the knowledge assistance required for detection and tracking with the characteristics of frogman movement and scattered echo.

[0102] Figure 5 Shows the reference trajectory and tracking trajectory obtained using the three algorithms when SRR = 4dB. In the low SRR scenario with uncertain target direction, the MFKA-PF-TBD and KA-PF-TBD algorithms can accurately estimate the target direction and position. In contrast, the MMPF-TBD algorithm has poor performance and can only track the target within a few frames. The first two methods improve the performance by using auxiliary particles and improving the target state transition model, so that the particles move to the high likelihood ratio region. Since both of these methods perform tracking before the detection algorithm takes effect, the detection decision is made after tracking the target, and the tracking will affect the subsequent detection results.

[0103] Figure 6The position RMSE results of three algorithms per frame are compared. The position RMSE of the MFKA-PF-TBD algorithm is much smaller than that of the other two algorithms and has little fluctuation. This indicates that the MFKA-PF-TBD method has a smaller error and thus its tracking is more accurate.

[0104] Figure 7 The accurate detection probability sequences of the three methods are shown. The detection probability of the MFKA-PF-TBD algorithm of the present invention is higher, which is consistent with the tracking results, has less fluctuation, and the algorithm is more robust. In addition, since the accurate detection probability curve of the MFKA-PF-TBD algorithm has less fluctuation, the algorithm has stronger robustness.

[0105] The present invention also provides a computer program including software instructions, and the software instructions, when executed by a computer, implement the performance evaluation method, including but not limited to personal computers, minicomputers, mainframes, workstations, network or distributed computing environments, separate or integrated computer platforms, or communicating with charged particle tools or other imaging devices, etc. Aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer. When the storage medium or device is read by a computer, it can be used to configure and operate the computer to execute the processes described herein. In addition, the machine-readable code, or portions thereof, can be transmitted via a wired or wireless network. When such media includes instructions or programs that implement the above-described steps in combination with a microprocessor or other data processor, the inventions described in the present invention include these and other different types of non-transitory computer-readable storage media. When programmed according to the methods and techniques described in the present invention, the present invention also includes the computer itself.

[0106] As described above, it is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. As long as it achieves the technical effects of the present invention by the same means, it should fall within the protection scope of the present invention. Within the protection scope of the present invention, various different modifications and variations can be made to its technical solutions and / or implementation manners.

Claims

1. Detection-before-tracking algorithm based on multi-feature knowledge-assisted particle filter, characterized in that It includes the following steps: Step 1: Initialize the target state and sonar observation parameters; Step 2: Construct a multi-directional motion and state transition model for frogmen; Step 3: Particle filter based on multi-feature assistance.

2. The detection-before-tracking algorithm based on multi-feature knowledge-aided particle filter according to claim 1, wherein The said Step 1: Initialize the target state and sonar observation parameters, specifically: Assume that the target is in the far field, and its position coordinates are The corresponding speeds are respectively and Thus, the target state is determined to be The relationship between the distance measurement and angle measurement of the sonar and the target position is expressed as The sonar system is composed of a transmitting array element and a uniform linear array, and the received signal r(t) is represented in vector form; the transmitted signal s(t) propagates in water, interacts with the target to generate a reflected signal, and the expression of the received signal vector is In this expression is a vector, the elements of which reflect the signal phase differences received by different array elements; τ m (θ t ) = (m - 1)dsinθ t / c takes into account the propagation time delay caused by different target positions and array element angles when sound waves propagate in water, where c is the speed of sound; A is the reflection coefficient of the target, characterizing the reflection ability of the target to sound waves; ω is the carrier angular frequency, determining the frequency characteristics of the signal; represents the total propagation time delay of the signal from transmission to the target and then to reception; v(t) is a vector representing noise, including various interference factors such as environmental noise and electronic noise; For the convenience of digital signal processing, mainly considering complex baseband signals, the discrete form of the received signal is simplified to r(k) = a(θ k )As(k - τ tt+tr ) + v(k), where k represents the kth sampling time; Taking the sonar position as the origin, the observation area distance is restricted within [R min , R max , and is further divided into N r range cells; the azimuth is limited within [θ min , θ max , and is divided into N b azimuth cells; the number of range cells N r is where F s is the sampling frequency in array signal processing, and N b can be determined by the direction resolution unit Δθ; At this time, the measurement unit Δθ contains N at time k r ×N b data where g(·) is the mapping between the signal y(k) and the measurement, and its form affects the specific form of the measurement; while w k is the measurement value of the noise and clutter of the system at time k.

3. The detection-before-tracking algorithm based on multi-feature knowledge-assisted particle filtering according to claim 1, characterized in that, The said Step 2: Construct a multi-directional motion and state transition model for frogmen, specifically: Suppose the model consists of 8 directions and 16 uniform linear motions; taking the position of the frogman target at the current moment as the reference point, it is marked with a circle; eight arrows are extended from this point, and the direction pointed by each arrow represents the possible motion direction of the frogman target; in each direction, two uniform linear motion models are further subdivided; the speeds of these two models are respectively set to 0.75 times and 1.25 times the target speed range at the previous moment; Suppose the general expression of target state transition is x k+1 = f(x k , τ k ) + v k , where, x k represents the state of the target at time k, τ k represents the type of the target motion model, v k is the corresponding process noise, and f(·) is the state transition matrix under different models; Assume that within the time interval T, the speeds of the target in the x and y directions remain unchanged, and the position update is only based on the current speed and time interval. For example, the position update in the x direction is as follows, and the y direction is the same, For the coordinated turning motion model (τ k = 2), its state transition matrix is where w k is the coordinated turn rate; The transition probability between each motion mode is represented by a Markov chain at time k as P{τ k =j | τ k-1 =i} = p ij , i, j = 1, 2.

4. The detection-before-tracking algorithm based on multi-feature knowledge-assisted particle filter according to claim 1, characterized in that The said Step 3: Particle filter based on multi-feature assistance, specifically: Suppose \(z_1, z_2, \ldots, z\) n are \(n\) measured data samples, and is the kernel density estimate of the sample probability density function, then The expression of is where n is the number of independent and identically distributed samples, K(·) is the kernel function, which determines the role of each sample data point z i , i = 1, …, n in the density estimation of the random variable z; and h is the window width that affects the smoothness of the probability density estimation; After obtaining the statistical characteristics of the real echo data, the likelihood function of the measurement resolution unit (i,j) is where g1(·) represents the statistical properties of the measurement data when the target exists, and g0(·) represents the statistical properties of the measurement data when the target does not exist; p s (t i,j ) is the value of the frogman echo waveform function corresponding to the resolution cell (i, j); since the target occupies multiple resolution cells in the measurement space, the likelihood function is expressed as where Θ is the range of the target affecting the resolution unit; The specific sub-steps are as follows: Let the particle set be available at time k - 1 to describe the posterior probability density (1) Initialize particles: At the initial moment, only generate new particles; if the prior distribution of the target is known, generate particles according to its distribution; if there is no prior information about the target, uniformly sample the samples within the observation area; (2) Generating new particles and estimating the state of persistent particles: According to the prior probability distribution, N b new particles are generated. Using the established diver motion model and scattering feature model as knowledge auxiliary information, the state estimation of N c consecutive particles is as follows (3) Calculate particle weights and normalize: Use the statistical characteristics of measurement data and the waveform characteristics of frogman scattering echoes as knowledge auxiliary information; the weight calculation and normalization of new particles and continuous particles are as follows (4) Calculate the mixing probability and normalize: Calculate the mixing probability using the unnormalized weights, and then normalize, as follows (5) Adjust particle weights: Scale the weights of new and continuous particles according to the mixing probability, as follows (6) Particle resampling and target state estimation: The new particles and the consecutive particles form a complete set of particles, which are resampled to Then the target state is estimated as follows (7) Trajectory start and end judgment: For each target with N t frames, calculate the sum of all particle weights Σ w to be If the following formula is satisfied, the target start fails or terminates, Σ ω <η d , where η d is the likelihood threshold; (8) Output the target tracking result: Generate and output the target tracking result.

5. A computer device, characterized in that, It includes: A memory for storing a computer program; A processor for implementing any one of the methods described in claims 1 to 4 when executing the said computer program.

6. A readable storage medium, characterized in that, A computer program is stored on the said readable storage medium, and when the computer program is executed by the processor, it implements any one of the methods described in claims 1 to 4.

Images