Self-adaptive detection and tracking method based on Bernoulli filter

Abstract

The invention provides an adaptive detection and tracking method based on a Bernoulli filter, and the method comprises the steps: employing a random finite set to model a target state, and describing the Markov process dynamic characteristics of a target based on a state transition probability; calculating a detection probability and a false alarm rate by using a Newman Pearson criterion, and modeling measurement as a random finite set; according to the existence probability and the space state of the target at the previous moment, combining the birth probability, the survival probability and the state transition model of the target, predicting the existence probability and the space state of the target at the current moment; predicting the GOSPA performance of the Bernoulli filter, modeling a target adaptive detection and tracking problem into an optimization problem taking a false alarm rate as an optimization parameter, and achieving the minimization of the GOSPA performance predicted by the Bernoulli filter by adjusting the false alarm rate; a golden proportion constant is calculated by initializing a false alarm rate search interval and an allowable error, GOSPA performance is iteratively optimized by using a golden section method, and an optimal false alarm rate and a corresponding detection threshold value are obtained; a detection result is obtained by using an optimal detection threshold value, then the detection result is used as input, the target existence probability and the spatial state are recursively updated, the calculation complexity is controlled by cutting low weight components and combining similar components, and efficient and accurate self-adaptive detection and tracking are achieved.

Description

Technical Field

[0001] This invention belongs to the field of target detection and tracking technology, specifically relating to an adaptive detection and tracking method based on a Bernoulli filter. Background Technology

[0002] Target tracking typically employs a "detect-then-track" framework, where a detector determines the presence of a target, and a tracker estimates the target's state based on the detection results. Since the detector and tracker are usually developed independently, the resulting system performance may be suboptimal. To overcome this limitation, a tracking-aware adaptive detection method has been proposed, which improves tracking performance by optimizing the detection threshold. In recent years, Bernoulli filters have been introduced to recursively estimate target presence and state. Explicitly modeling target presence not only improves tracking robustness but also affects the construction of the adaptive detection and tracking problem. The adaptive detection and tracking problem based on Bernoulli filters has not yet been systematically studied. The key challenge lies in how to efficiently use the performance prediction of the Bernoulli filter as the cost function for optimizing the detection threshold, and further solve the resulting optimization problem to determine the optimal detection threshold. Summary of the Invention

[0003] The present invention aims to at least partially solve one of the technical problems in the related art.

[0004] Therefore, the first objective of this invention is to propose an adaptive detection and tracking method based on a Bernoulli filter.

[0005] The second objective of this invention is to propose an adaptive detection and tracking device based on a Bernoulli filter.

[0006] To achieve the above objectives, a first aspect of the present invention proposes an adaptive detection and tracking method based on a Bernoulli filter, comprising: S1, using a random finite set The target state is modeled, and the dynamic characteristics of the target's Markov process are described based on the state transition probability; the detection probability and false alarm rate are calculated using the Newman-Pearson discrimination criterion, and the measurement is modeled as a random finite set. S2, based on the target existence probability and spatial state obtained in the previous moment, predict the target existence probability and spatial state at the current moment through the target birth probability, survival probability and state transition model; S3. Based on the predicted existence probability and spatial state of the target, predict the GOSPA performance of the Bernoulli filter, model the target adaptive detection and tracking problem as an optimization problem with the false alarm rate as the optimization parameter, and minimize the GOSPA performance by adjusting the false alarm rate. S4. By initializing the false alarm rate search interval and allowable error and calculating the golden ratio constant, the GOSPA performance is iteratively compared within the interval based on the golden section method and the search range is gradually narrowed until the convergence condition is met, thereby obtaining the optimal false alarm rate and obtaining the corresponding detection threshold according to the constraint relationship. S5: Obtain the detection result using the optimal detection threshold, and use the detection result as input to recursively deduce the target existence probability and single target state density. In implementation, the computational complexity is controlled by pruning low-weight components and merging similar components.

[0007] In one embodiment of the present invention, S1 includes: S11, when a random finite set When it is an empty set, it means that the target does not exist in the monitoring area; when Contains a single element When, it indicates that the target exists within the monitoring area, where for Single-target status; S12, Measurement of random finite sets The cardinality and element position are random variables: when the target does not exist, the measurement only contains clutter and the spatial distribution is uniform; when the target exists, the measurement is described by the signal generated by the target and the clutter, and its spatial distribution is jointly determined by the measurement likelihood generated by the target and the clutter distribution.

[0008] In one embodiment of the present invention, S2 includes: S21, the probability of the target prediction's existence and spatial state Calculated using the following formula:

[0009]

[0010] in, Here is the state transition matrix. Let be the noise covariance matrix of the target motion process.

[0011] In one embodiment of the present invention, S3 includes: S31 predicts the GOSPA performance of the Bernoulli filter by jointly averaging the uncertainties of both the target and the measurement, normalizing the measurement information, and using a threshold technique. S32, the optimization parameter of the adaptive detection and tracking problem is the false alarm rate. Under the given signal-to-noise ratio, by establishing the correspondence between the false alarm rate, detection probability and the average number of false alarms, an optimization problem is constructed with the goal of minimizing the prediction GOSPA metric, and the optimal false alarm rate or detection threshold is obtained by solving the problem.

[0012] In one embodiment of the present invention, S4 includes: S41, based on the golden ratio constant Calculate interior points and And compare in each iteration and The value is updated by updating the search range. Until the convergence condition is met ,in Tolerance; S42, outputs the optimal false alarm rate and by constraints Calculate the corresponding detection threshold .

[0013] To achieve the above objectives, a second aspect of the present invention provides an adaptive detection and tracking device based on a Bernoulli filter, comprising: The random finite set modeling module is used to model random finite sets. The target state is modeled, and the dynamic characteristics of the target's Markov process are described based on the state transition probability; the detection probability and false alarm rate are calculated using the Newman-Pearson discrimination criterion, and the measurement is modeled as a random finite set. The existence probability and spatial state prediction module is used to predict the existence probability and spatial state of the target at the current moment based on the target existence probability and spatial state obtained at the previous moment, through the target birth probability, survival probability and state transition model; The GOSPA performance prediction and optimization problem establishment module is used to predict the GOSPA performance of the Bernoulli filter based on the predicted existence probability and spatial state of the target. It models the target adaptive detection and tracking problem as an optimization problem with the false alarm rate as the optimization parameter, and minimizes the GOSPA performance by adjusting the false alarm rate. The module for optimizing the false alarm rate using the golden section method is used to initialize the false alarm rate search interval and allowable error and calculate the golden ratio constant. Based on the golden section method, it iteratively compares the GOSPA performance within the interval and gradually narrows the search range until the convergence condition is met, thereby obtaining the optimal false alarm rate and obtaining the corresponding detection threshold according to the constraint relationship. The target existence probability update and computational complexity control module is used to obtain the detection result using the optimal detection threshold, and use the detection result as input to recursively deduce the target existence probability and single target state density. In implementation, the computational complexity is controlled by pruning low-weight components and merging similar components.

[0014] The beneficial effects of this invention are: This invention uses the GOSPA metric as the optimization function, which comprehensively evaluates three types of errors in target tracking: the positioning error of correctly tracking the target, the error introduced by missing targets and false tracks. Therefore, the resulting adaptive detection and tracking system can comprehensively optimize the target existence estimation and state estimation.

[0015] This invention utilizes target state prediction information provided by the tracker and optimizes the detection threshold through closed-loop feedback to achieve optimal detection and tracking performance. Compared with the fixed threshold method, it has better tracking accuracy and target presence judgment ability.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an adaptive detection and tracking method based on a Bernoulli filter according to an embodiment of the present invention; Figure 2 This is an architecture diagram of an adaptive detection and tracking method based on a Bernoulli filter according to an embodiment of the present invention; Figure 3 According to an embodiment of the present invention, GOSPA is based on the false alarm rate. Change curves (different) Schematic diagram (under SNR conditions); Figure 4 The optimal false alarm rate according to the embodiment of the present invention is... A diagram illustrating the changes in SNR; Figure 5 The optimal GOSPA according to the embodiment of the present invention A diagram illustrating the changes in SNR; Figure 6 This is a schematic diagram illustrating the evolution of GOSPA metric, target presence probability, and false alarm rate over time according to an embodiment of the present invention. Figure 7 This is a schematic diagram comparing the performance over time under fixed and optimized threshold conditions (single Gaussian implementation, SNR=20 dB) according to an embodiment of the present invention. Figure 8 This is a schematic diagram comparing the average performance of a Bernoulli filter under fixed threshold and optimized threshold (single Gaussian implementation) conditions according to an embodiment of the present invention; Figure 9This is a schematic diagram illustrating the performance index of Gaussian mixture realization under SNR=20 dB conditions as a function of time according to an embodiment of the present invention. Figure 10 This is a schematic diagram comparing the average performance indicators under the conditions of a fixed detection threshold and an optimized detection threshold (Gaussian mixture implementation) according to an embodiment of the present invention; Figure 11 This is a structural diagram of an adaptive detection and tracking device based on a Bernoulli filter according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] An adaptive detection and tracking method and apparatus based on a Bernoulli filter, according to an embodiment of the present invention, is described below with reference to the accompanying drawings.

[0021] Figure 1 This is a flowchart of an adaptive detection and tracking method based on a Bernoulli filter according to an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: S1, using a random finite set The target state is modeled, and the dynamic characteristics of the target's Markov process are described based on the state transition probability; the detection probability and false alarm rate are calculated using the Newman-Pearson discrimination criterion, and the measurement is modeled as a random finite set. S2, based on the target existence probability and spatial state obtained in the previous moment, predict the target existence probability and spatial state at the current moment through the target birth probability, survival probability and state transition model; S3. Based on the predicted existence probability and spatial state of the target, predict the GOSPA performance of the Bernoulli filter, model the target adaptive detection and tracking problem as an optimization problem with the false alarm rate as the optimization parameter, and minimize the GOSPA performance by adjusting the false alarm rate. S4. By initializing the false alarm rate search interval and allowable error and calculating the golden ratio constant, the GOSPA performance is iteratively compared within the interval based on the golden section method and the search range is gradually narrowed until the convergence condition is met, thereby obtaining the optimal false alarm rate and obtaining the corresponding detection threshold according to the constraint relationship. S5: Obtain the detection result using the optimal detection threshold, and use the detection result as input to recursively deduce the target existence probability and single target state density. In implementation, the computational complexity is controlled by pruning low-weight components and merging similar components.

[0022] The following describes in detail an adaptive detection and tracking method based on a Bernoulli filter according to an embodiment of the present invention, with reference to the accompanying drawings.

[0023] This invention discloses an efficient prediction method for Bernoulli filter performance based on the GOSPA metric, such as... Figure 2 As shown, the following implementation methods are included.

[0024] S10: Time Using random finite sets The target state is modeled, and the dynamic characteristics of the target's Markov process are described based on the state transition probabilities of a random finite set. The Newman-Pearson discrimination criterion is used for target detection, and the computational probability and false alarm rate are calculated based on the detection threshold. The detected measurements are modeled as a random finite set.

[0025] In the Bernoulli model, time... The target state is composed of a random finite set Indicates. When When, it indicates that the target is not within the monitoring area; when When, it indicates that the target exists within the monitored area, where express A single-objective state in dimensionality. The dynamic characteristics of the Bernoulli Markov process are determined by the RFS transfer density. What is depicted. When At that time, its transfer density is

[0026] in, For the probability of birth, For a moment Single-target birth density, The cardinality of the state set. When At that time, its transfer density is

[0027] in, For the probability of survival, It is a single-objective transfer density function.

[0028] Signal and Measurement Model. The signal received by the sensor is divided into multiple resolution units, and the existence of the target is checked within each resolution unit. Under the Swerling-I target fluctuation model, the... The amplitude of each resolution unit follows an exponential distribution.

[0029] in, Indicates the first The amplitude of each resolution unit , To distinguish the total number of units, This represents the signal-to-noise ratio of the target. Assume... This indicates that there is no target within the resolution cell. This indicates that a target exists within the resolution unit.

[0030] The hypothesis testing problem described above can be solved using the Newman-Pearson criterion, thus yielding a constant false alarm rate (CFAR) detector. This detector aims to maximize the detection probability while constraining the false alarm probability, typically by adjusting the signal amplitude. With detection threshold This is achieved through comparison.

[0031] For a CFAR detector, the detection probability and false alarm rate are respectively , Furthermore, the number of false alarms follows the parameter: The Poisson distribution, in which To monitor the volume of the area, To detect and distinguish the volume of the unit; the false alarm locations follow a uniform distribution over the monitored area, and their probability density function is: To simplify symbolic representation, and without causing confusion, , and In Dependencies can be omitted.

[0032] At any moment The measurement after thresholding by the CFAR detector is modeled as RFS. Its base and measurement space Elements The positions of all variables are random variables. Measurement vector The dimension. When When only clutter generates the measurement probability density function, it is:

[0033] in, Let be the spatial probability density function for false alarms. At that time, the measurement probability density function jointly generated by clutter and the target is:

[0034] in, Indicates the state Target generation measurement The likelihood function.

[0035] S20: Based on the target existence probability and spatial state obtained in the previous moment, predict the target existence probability and spatial state at the current moment through the target birth probability, survival probability and state transition model.

[0036] The probability of the target existing at the previous moment was The spatial density is given by the Gaussian mixture form, i.e., , in, The existence probability and spatial density of the predicted target are obtained by the prediction step of the Bernoulli filter, where, and They are respectively

[0037]

[0038] in, Indicates the probability of a target newborn. Indicates time Single-target birth density, It is the state transition matrix. It is the noise covariance matrix of the target motion process. The mean is The covariance is The Gaussian distribution.

[0039] S30: Based on the predicted existence probability and spatial state of the target, predict the GOSPA performance of the Bernoulli filter, model the target adaptive detection and tracking problem as an optimization problem with the false alarm rate as the optimization parameter, and minimize the GOSPA performance by adjusting the false alarm rate.

[0040] Measuring random finite sets Essentially dependent on the detection threshold, this means the GOSPA metric is also a function of the detection threshold. In the CFAR detector, the false alarm rate... Based on detection threshold Explicitly determined. Therefore, direct optimization. This approach can more meaningfully reflect the GOSPA metric in theory. Therefore, this paper defines the optimization variable as the false alarm rate. Therefore, the adaptive detection and tracking problem can be rigorously formulated as the following optimization task.

[0041] Note that performance prediction of the GOSPA metric requires joint averaging of uncertainties in both the target and the measurement.

[0042] S40: By initializing the search interval and allowable error and calculating the golden ratio constant, the objective function value is iteratively compared within the interval based on the golden section method, and the search range is gradually narrowed until the convergence condition is met, thereby obtaining the optimal false alarm rate and obtaining the corresponding detection threshold according to the constraint relationship.

[0043] Initialize search interval and tolerance And calculate the golden ratio constant. Calculate two interior points based on the golden ratio: , .exist and Calculate the objective function at each location. Compare function values ​​and narrow the search interval: If Then let Otherwise, Repeat the above calculations until the condition is met. Output the optimal false alarm rate. and by constraints Calculate the corresponding detection threshold .

[0044] S50: Obtain the detection result using the optimal detection threshold, and use the detection result as input to recursively deduce the target existence probability and single target state density. In implementation, the computational complexity is controlled by pruning low-weight components and merging similar components.

[0045] The update equation is

[0046]

[0047]

[0048] in

[0049] To maintain manageable complexity, in practice, components with smaller weights need to be pruned and components that are close together need to be merged.

[0050] Furthermore, the GOSPA index was analyzed in relation to the false alarm rate. The changing pattern. The probability of target prediction taking [a certain value]. The predicted spatial density is represented by two Gaussian components, with parameters as follows: as well as The measurement model uses a two-dimensional linear model, and the noise is white Gaussian noise. The signal-to-noise ratio is set to... GOSPA metric parameters are taken as follows: , The threshold is . Figure 3 Showing different And the GOSPA index under SNR The evolution of the CFAR detector can be observed. It can be seen that there exists an optimal false alarm rate that minimizes the GOSPA index, and this optimal false alarm rate varies with... It changes with SNR. As expected, when SNR decreases, the optimal GOSPA index value and the corresponding optimal false alarm rate both increase, resulting in a higher detection probability. It should be noted that when... Furthermore, when SNR=10dB, the GOSPA index exhibits a sudden change with the false alarm rate, which is due to the shift in target presence estimation from "no target" to "target present".

[0051] Figure 4 and Figure 5 The optimal GOSPA index value and the optimal false alarm rate are given respectively. And the changes in SNR. When At this time, the optimal GOSPA index value remains small because the filter tends to infer that the target does not exist; within this range, the optimal false alarm rate also remains at a low level, thus effectively suppressing false alarms. When the value is increased to 0.5, the optimal GOSPA index value increases because the existence of the target is difficult to determine, and the optimal false alarm rate also increases, in order to improve the detection probability and improve the target existence estimation. At that time, further improve This will cause the optimal GOSPA index value to decrease, because the existence of the target can be estimated more accurately at this time; at the same time, the optimal false alarm rate will also decrease, so as to effectively suppress false alarms.

[0052] The focus is on comparing the performance differences between Bernoulli filters using a fixed false alarm rate and those using an optimized false alarm rate. Consider a two-dimensional monitoring area with a range of [area missing]. There can be at most one target within this region. The target state is denoted as... , which includes location and speed The target is generated by a single Bernoulli random finite set with a generation probability of . The generation density is ,in, , The total simulation time is 100 seconds, with a step size of 1 second. The sensor returns a position measurement. Likelihood function , , The signal resolution is 1m in both the X and Y directions. A constant false alarm rate (CFAR) detector is used for detection. Bernoulli filters are implemented using single Gaussian and Gaussian mixture filters, respectively. The pruning threshold for the mixture components is set to... The component merging was performed using a chi-square test with a threshold of 4 and a maximum number of components of 100. The gating probability was set to 99.99%. The existence of the target was determined when the estimated existence probability was greater than 0.5, and its state estimate was the mean of the merged posterior components. All simulations in this paper were performed on 500 random implementations. The performance metrics compared included: (1) GOSPA metric; (2) cardinality error; (3) root mean square error of position; and (4) computation time.

[0053] Figure 6 This paper demonstrates the time-varying GOSPA metric, target presence probability, and false alarm rate of the Bernoulli filter under a single Gaussian implementation with an SNR of 20 dB. Compared to the fixed detection threshold method, the optimized method increases the detection threshold to suppress false alarm trajectories and thus reduce the false alarm rate when the target presence probability is low; conversely, it decreases the detection threshold to enhance the detection probability when the target presence probability increases, maintaining reliable tracking while moderately increasing the false alarm rate. This dynamic adjustment effectively balances the trade-off between detection probability and false alarm rate, thereby improving performance.

[0054] Figure 7 The performance of a Bernoulli filter implemented with a single Gaussian threshold was compared under the condition of SNR=20dB, with fixed and optimized false alarm rates. The results show that the Bernoulli filter with optimized false alarm rate outperforms the fixed threshold scheme in terms of GOSPA metric and cardinality error. To further evaluate the robustness of the algorithm, Figure 8 The changes in GOSPA metric, cardinality error, RMSE, and computation time are shown when SNR = 10, 15, 20, 25, and 30 dB. The results show that as SNR increases, GOSPA metric, cardinality error, and position RMSE all decrease significantly, verifying the positive correlation between SNR and tracking performance. Furthermore, at the same SNR, the performance of the filter optimizing the false alarm rate is superior to the fixed threshold method; moreover, the running time of the optimized algorithm only increases by 3-4 times.

[0055] Figure 9Performance comparisons of Bernoulli filters implemented using a Gaussian mixture with a fixed false alarm rate and those with an optimized false alarm rate are presented at an SNR of 20 dB. The results show that the Bernoulli filter with the optimized false alarm rate also exhibits advantages in terms of GOSPA metric and cardinality error. To further evaluate performance under different SNR conditions, Figure 10 The results show how the SNR changes from 10 dB to 30 dB. The results consistently demonstrate that the Bernoulli filter with optimized false alarm rate outperforms the Bernoulli filter with a fixed false alarm rate, and the running time of the optimized algorithm only increases by 3-4 times.

[0056] To achieve the above embodiments, such as Figure 11 As shown, this embodiment also provides an adaptive detection and tracking device 10 based on a Bernoulli filter. The device 10 includes a random finite set modeling module 100, an existence probability and spatial state prediction module 200, a GOSPA performance prediction and optimization problem establishment module 300, a golden section method optimization false alarm rate module 400, and a target existence probability update and computational complexity control module 500.

[0057] The random finite set modeling module is used to model random finite sets. The target state is modeled, and the dynamic characteristics of the target's Markov process are described based on the state transition probability; the detection probability and false alarm rate are calculated using the Newman-Pearson discrimination criterion, and the measurement is modeled as a random finite set. The existence probability and spatial state prediction module is used to predict the existence probability and spatial state of the target at the current moment based on the target existence probability and spatial state obtained at the previous moment, through the target birth probability, survival probability and state transition model; The GOSPA performance prediction and optimization problem establishment module is used to predict the GOSPA performance of the Bernoulli filter based on the predicted existence probability and spatial state of the target. It models the target adaptive detection and tracking problem as an optimization problem with the false alarm rate as the optimization parameter, and minimizes the GOSPA performance by adjusting the false alarm rate. The module for optimizing the false alarm rate using the golden section method is used to initialize the false alarm rate search interval and allowable error and calculate the golden ratio constant. Based on the golden section method, it iteratively compares the GOSPA performance within the interval and gradually narrows the search range until the convergence condition is met, thereby obtaining the optimal false alarm rate and obtaining the corresponding detection threshold according to the constraint relationship. The target existence probability update and computational complexity control module is used to obtain the detection result using the optimal detection threshold, and use the detection result as input to recursively deduce the target existence probability and single target state density. In implementation, the computational complexity is controlled by pruning low-weight components and merging similar components.

[0058] Furthermore, the aforementioned random finite set modeling and detection module 100 is also used for: When random finite set When it is an empty set, it means that the target does not exist in the monitoring area; when Contains a single element When, it indicates that the target exists within the monitoring area, where for Single-target status; Measuring random finite sets The cardinality and element position are random variables: when the target does not exist, the measurement only contains clutter and the spatial distribution is uniform; when the target exists, the measurement is described by the signal generated by the target and the clutter, and its spatial distribution is jointly determined by the measurement likelihood generated by the target and the clutter distribution.

[0059] Furthermore, the aforementioned existence probability and spatial state prediction module 200 is also used for: The probability of the target's existence and spatial state Calculated using the following formula:

[0060]

[0061] in, Here is the state transition matrix. Let be the noise covariance matrix of the target motion process.

[0062] Furthermore, the GOSPA performance prediction and optimization problem-establishing module 300 mentioned above is also used for: By jointly averaging the uncertainties of both the target and measurement aspects, and by normalizing the measurement information and employing a threshold technique, the GOSPA performance of the Bernoulli filter is predicted. The optimization parameter for the adaptive detection and tracking problem is the false alarm rate. Given the signal-to-noise ratio, by establishing the correspondence between the false alarm rate, detection probability and the average number of false alarms, an optimization problem is constructed with the goal of minimizing the prediction GOSPA metric, and the optimal false alarm rate or detection threshold is obtained by solving the problem.

[0063] Furthermore, the aforementioned golden ratio method-optimized false alarm rate module 400 is also used for: According to the golden ratio constant Calculate interior points and And compare in each iteration and The value is updated by updating the search range. Until the convergence condition is met ,in Tolerance; Output the optimal false alarm rate and by constraints Calculate the corresponding detection threshold .

[0064] An adaptive detection and tracking device based on a Bernoulli filter according to an embodiment of the present invention can realize closed-loop joint optimization of detection and tracking, effectively improve the accuracy of target existence judgment and state estimation, and at the same time, efficiently solve the optimal false alarm rate by means of the golden section method, significantly reducing computational complexity.

[0065] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0066] Furthermore, the terms "first" and "second" 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. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A Bernoulli filter based adaptive detection and tracking method, characterized in that, Comprise: S1, adopt random finite set Modeling target state and describing the dynamic characteristics of Markov process of target according to state transition probability; calculating detection probability and false alarm rate by using Neyman-Pearson criterion, and modeling measurement as random finite set; S2, according to the target existence probability and space state obtained at last time, through the target birth probability, survival probability and state transition model, the existence probability and space state of target current time are predicted; S3, according to the existence probability and space state predicted by target, the GOSPA performance of Bernoulli filter is predicted, the adaptive detection and tracking problem of target is modeled as optimization problem with false alarm rate as optimization parameter, and the minimization of GOSPA performance is realized by adjusting false alarm rate; S4, by initializing false alarm rate search interval and allowable error and calculating golden ratio constant, the GOSPA performance is iteratively compared in the interval based on golden section method and the search range is gradually reduced until the convergence condition is met, so that the optimal false alarm rate is obtained and the corresponding detection threshold is obtained according to the constraint relationship; S5, the optimal detection threshold is used to obtain the detection result, and the detection result is used as input to recursively update the target existence probability and single target state density, and the calculation complexity is controlled by cutting low weight components and merging similar components in implementation.

2. The method of claim 1, wherein, Said S1, comprising: S11, when the random finite set is empty, indicates that the target is not present in the surveillance region; when contains a single element , indicates that the target is present in the surveillance region, where is a one-dimensional target state; S12, measure random finite set The cardinality and element locations are random variables: when the target is absent, the measurement only contains clutter and the spatial distribution is uniform; when the target is present, the measurement is jointly described by the signal generated by the target and the clutter, and the spatial distribution is jointly determined by the measurement likelihood generated by the target and the clutter distribution.

3. The method of claim 1, wherein, Said S2, comprising: S21, target predicted presence probability and spatial state is calculated by the following equation: wherein is the state transition matrix, is the target motion process noise covariance matrix.

4. The method of claim 1, wherein, Said S3, further comprising: S31, the GOSPA performance of Bernoulli filter is predicted by jointly averaging the uncertainty of target and measurement, and by normalizing the measurement innovation and adopting threshold technology; S32, the optimization parameter of adaptive detection and tracking problem is false alarm rate, under the condition of given signal-to-noise ratio, the corresponding relationship among false alarm rate, detection probability and false alarm average number is established, the optimization problem with the minimum predicted GOSPA metric as the target is constructed, and the optimal false alarm rate or detection threshold is obtained.

5. The method of claim 1, wherein, Said S4, comprising: S41, according to the golden ratio constant , calculate the internal point and , and compare the value of with in each iteration, update the search interval until the convergence condition is met , where is the allowable error; S42, output the optimal false alarm rate and by the constraint relationship corresponding detection threshold is obtained .

6. A Bernoulli filter based adaptive detection and tracking apparatus, characterized by, Comprise: a stochastic finite set modeling module for modeling the target state as a stochastic finite set a Markov process dynamic characteristic of the target according to the state transition probability; and calculating the detection probability and the false alarm rate by using the Neyman-Pearson criterion, and modeling the measurement as a stochastic finite set; existence probability and space state prediction module, for according to the target existence probability and space state obtained at last time, through the target birth probability, survival probability and state transition model, the existence probability and space state of target current time are predicted; GOSPA performance prediction and optimization problem establishment module, for according to the existence probability and space state predicted by target, the GOSPA performance of Bernoulli filter is predicted, the adaptive detection and tracking problem of target is modeled as optimization problem with false alarm rate as optimization parameter, and the minimization of GOSPA performance is realized by adjusting false alarm rate; Golden section method optimization false alarm rate module, by initializing false alarm rate search interval and allowable error and calculating golden ratio constant, the GOSPA performance is iteratively compared in the interval based on golden section method and the search range is gradually reduced until the convergence condition is met, so that the optimal false alarm rate is obtained and the corresponding detection threshold is obtained according to the constraint relationship; Target existence probability update and calculation complexity control module, for using the optimal detection threshold to obtain the detection result, and using the detection result as input to recursively update the target existence probability and single target state density, and the calculation complexity is controlled by cutting low weight components and merging similar components in implementation.

7. The apparatus of claim 6, wherein, Said random finite set modeling and detection module is also used for: When the random finite set is empty, indicates that the target is not present in the surveillance region; when contains a single element , indicates that the target is present in the surveillance region, where is a one-dimensional target state; Measuring random finite sets The cardinality and element locations are random variables: when the target is absent, the measurements only contain clutter and the spatial distribution is uniform; when the target is present, the measurements are jointly described by the signal generated by the target and the clutter, and the spatial distribution is jointly determined by the measurement likelihood generated by the target and the clutter distribution.

8. The apparatus of claim 6, wherein, Said existence probability and space state prediction module is also used for: Targeted prediction of presence probability and spatial state is calculated by the following equation: wherein is the state transition matrix, is the target motion process noise covariance matrix.

9. The apparatus of claim 6, wherein, Said GOSPA performance prediction and optimization problem establishment module is also used for: By jointly averaging the uncertainties of target and measurement, and by normalizing the measurement innovation and using threshold technique, the GOSPA performance of Bernoulli filter is predicted; The optimization parameter of the adaptive detection and tracking problem is the false alarm rate. Under a given signal-to-noise ratio, by establishing the corresponding relationship among the false alarm rate, the detection probability and the average number of false alarms, an optimization problem is constructed to minimize the predicted GOSPA metric, and the optimal false alarm rate or detection threshold is obtained.

10. The apparatus of claim 6, wherein, The golden section method optimization false alarm rate module is also used for: According to the golden ratio constant , the internal points and are calculated and the values of and are compared in each iteration, by updating the search interval until the convergence condition is met, where is the tolerance error; Output optimal false alarm rate and by the constraint relation the corresponding detection threshold is obtained.

Images