A multi-sensor multi-view fusion method close to ideal random outlier detection

Abstract

A multi-sensor multi-view fusion method that approximates ideal random outlier detection. It belongs to the technical field related to nonlinear motion tracking of multi-sensor multi-view multi-moving targets with different perspectives. The Gaussian jump Markov cardinality probability hypothesis density is used for nonlinear system moving target tracking. The self-designed approximate ideal solution sorting random outlier selection method is used to improve the problem that the stochastic outlier selection (SOS) method needs to manually set the threshold, and the running accuracy of the SOS algorithm in non-co-viewing domains is improved. The posterior intensity function is decomposed into multiple sub-intensities, and the distribution of the number of targets in the corresponding area is characterized by multi-Bernoulli reconstruction cardinality distribution, so as to realize distributed fusion target tracking in non-co-viewing domains of nonlinear systems and improve the influence of perspective differences on nonlinear systems. The method described in the present invention can be applied to technical fields such as vehicle speeding measurement and ship navigation.

Description

Technical Field

[0001] The present invention belongs to the technical field related to nonlinear motion tracking of multi-sensor multi-view multi-moving targets with different viewing angles, and specifically relates to a multi-sensor multi-view fusion method that approximates ideal random outlier detection. Background Art

[0002] Due to the different placement of multiple sensors, limited detection distance, and limited detection angle, the sensors will have different fields of view. Directly using distributed fusion algorithms under non-co-field of view conditions will lead to a series of problems such as underestimating the number of targets and repeatedly calculating the target motion trajectory. How to divide the field of view of sensor network nodes and how to select appropriate distributed fusion methods are the key to ensuring distributed fusion technology under non-co-field of view conditions. In fact, most of the multi-targets are nonlinear in motion, such as vehicle speeding measurement, ship navigation, etc. Nonlinear systems can more accurately describe the motion state of the system, while the GM-PHD and GM-CPHD methods directly act on the multi-target tracking system of nonlinear targets. The fusion effect is even worse under the influence of different perspectives. In the study of filters suitable for nonlinear systems: relevant scholars proposed nonlinear heterogeneous Bayesian decentralized data fusion, and the new heterogeneous fusion method promoted the application of homogeneous covariance intersection algorithm, etc. However, existing research mainly focuses on nonlinear motion tracking of multiple moving targets, and does not consider the fusion tracking problem between multiple sensors. In the research on distributed fusion technology of nonlinear systems, the fusion methods suitable for nonlinear systems are mainly discussed, and the usage conditions in complex situations such as non-isotropic are not taken into consideration. The research on distributed fusion of nonlinear systems with non-isotropic is even rarer, and the original stochastic outlier selection (SOS) method requires manual setting of thresholds, which leads to the problem of low accuracy during operation. Summary of the invention

[0003] In order to solve the problem of low distributed fusion accuracy of nonlinear systems with different viewing angles in the past, the present invention provides a multi-sensor multi-view fusion method that approaches ideal random outlier detection.

[0004] The method comprises the following steps:

[0005] S1, receiving sensor 1- i The detection data is used, and the GM-CPHD filter is used for prediction and update to obtain the Gaussian component posterior probability intensity corresponding to multiple targets;

[0006] S2, judging whether each target belongs to the same cluster of sensor perspective through the posterior probability strength of Gaussian components corresponding to multiple targets;

[0007] S3. If all targets belong to the same sensor perspective cluster, the maximum a posteriori probability criterion is used to estimate the number of targets and calculate the multi-target probability density distribution of this sensor perspective cluster.

[0008] S4. If the targets do not belong to the cluster of the same sensor perspective, the probability density distribution of multiple targets in clusters of different sensor perspectives is obtained by segmentation and fusion processing in turn;

[0009] S5, merging the multi-target probability density distributions of all sensor view clusters obtained in steps S3 and S4, and calculating the merged multi-target probability density distribution to characterize the distribution of the number of targets in the corresponding area.

[0010] Furthermore, the GM-CPHD filter is

[0011] ; Get a sensor k The posterior probability strength of the Gaussian components corresponding to the multiple targets in ,in, represents the target state, Indicates sensor k The number of Gaussian components of ; Indicates sensor No. The weights of the Gaussian components; The variance is , the expectation is Sensors No. Gaussian density function corresponding to the Gaussian components.

[0012] Furthermore, the segmentation process includes, in sequence, performing boundary segmentation processing of difference perspectives and performing SOS clustering processing for approximating an ideal solution.

[0013] Furthermore, the boundary segmentation process of the difference perspective is specifically as follows: sensors with different difference perspectives are divided into non-overlapping regions with different multi-target posterior probability intensities according to the boundaries of the difference perspectives of the sensors.

[0014] Furthermore, the SOS clustering process approaching the ideal solution is specifically as follows:

[0015] S51, data forward processing: divide the data into extremely large indicators, extremely small indicators, intermediate indicators and interval indicators, and divide the probability of abnormal values ​​into extremely small indicators, and divide the number of sensors into interval indicators; perform forward processing on the abnormal probability value to obtain an extremely large indicator;

[0016] S52, data standardization processing: performing standardization processing on the number of sensors and the abnormal probability value, obtaining the standardized number of sensors and the abnormal value probability, and expressing them in the form of a data matrix;

[0017] S53, optimal solution and worst solution processing: find the optimal solution distance in the data matrix and the worst solution distance ;

[0018] S54, calculate relative proximity: calculate the Scoring of goals :

[0019]

[0020] Will Sort from low to high, and the top 30% of the targets are outliers. Use outliers as indicators to divide clusters of different sensor perspectives, and divide the Gaussian components corresponding to the targets into different clusters.

[0021] S55. Cluster the Gaussian components corresponding to the targets to obtain multi-target probability distributions in different clusters.

[0022] Furthermore, the fusion process is specifically performed by:

[0023] ;

[0024] The probability density distribution of multiple targets in different sensor view clusters is obtained; among them, represents the GA fusion result, represents the probability distribution of multiple targets in different clusters, Representation Node The consistency weight of each sensor node , Indicated by A network of sensors ,in , yes A collection of nodes, is a set of links, for any two nodes , if the node From the node Receive data, then there is an edge ,node The internal nodes of .

[0025] The beneficial effects of the method of the present invention are:

[0026] The GM-CPHD (Gaussian Mixture Cardinalized Probability Hypothesis Density) filter is suitable for nonlinear system moving target tracking. The independently designed Technique for Order Preference by Similarity to Ideal Solution of Stochastic Outlier Selection (TOPSIS-SOS) method is used to improve the problem that the stochastic outlier selection (SOS) method needs to manually set the threshold, and the operation accuracy of the SOS algorithm in non-isoscopic domains is improved. The posterior intensity function is decomposed into multiple sub-intensities, and the distribution of the number of targets in the corresponding area is characterized by multi-Bernoulli reconstruction cardinality distribution, so as to realize distributed fusion target tracking in non-isoscopic domains of nonlinear systems and improve the influence of perspective difference on nonlinear systems.

[0027] The method of the present invention can be applied to technical fields such as vehicle overspeed measurement and ship navigation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION

[0029] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the protection scope of the present invention.

[0030] The steps of the present invention are as follows:

[0031] Step 1: Parameterize the Gaussian jump Markov cardinality probability hypothesis density and calculate the prediction and update of the GM-CPHD filter, i.e., the multi-target posterior probability strength.

[0032] pass:

[0033] ;

[0034] Conduct, among which, Represents a sensor k The strength of the posterior probability of multiple targets in the medium; Indicates the target state, and expands the state integral in the traditional CPHD filter into a double integral of mode and state, respectively, for the application variables and express; Indicates sensor k The number of Gaussian components; Indicates sensor No. The weights of the Gaussian components; The variance is , the expectation is Sensors No. Gaussian density function corresponding to Gaussian components (GCs).

[0035] Step 2: Determine whether the Gaussian component belongs to the same sensor perspective cluster

[0036] After implementing step 1 above, the posterior probability strength of the Gaussian component can be obtained. It is also necessary to determine whether they belong to the same cluster. The purpose of grouping is to distinguish which sensors detect the targets.

[0037] Step 2.1: Grouping of Gaussian components into clusters belonging to the same sensor view

[0038] If the target is detected by multiple sensors, it means that all Gaussian components in the group come from multiple sensors, where Indicates grouping When the Gaussian component grouping belongs to the same sensor perspective cluster, the prediction update repeats step 1. According to the GM-CPHD method, the maximum a posteriori probability (MAP) criterion is used to estimate the number of targets and obtain the corresponding The sub-content corresponding to each group is obtained through corresponding methods to obtain the cardinality distribution, such as Bernoulli approximation method, etc.

[0039] Step 2.2: Grouping of Gaussian components does not belong to the same sensor perspective cluster

[0040] When the Gaussian component groups do not belong to the same sensor perspective cluster, they need to be fused through a fusion method.

[0041] Step 3: The fusion process is realized through segmentation and fusion processing, which mainly involves boundary segmentation of different perspectives, SOS clustering close to the ideal, and GA fusion method.

[0042] Step 3.1: Segmentation process

[0043] Step 3.3.1: Boundary segmentation of different perspectives

[0044] After step 1 and step 2, the cardinality distribution collected by multiple sensors at different perspectives is collected for T iterations, and the sensors with different perspectives are segmented.

[0045] According to the multi-target probability distribution, the sensors with different difference perspectives are divided and divided into disjoint multi-target posterior probability intensity regions of the collective potential probability hypothesis density filter (CPHD) according to the boundary of the difference perspective.

[0046] Step 3.3.2: SOS clustering close to the ideal solution (TOPSIS-SOS)

[0047] While segmenting the nodes into non-overlapping nodes according to the boundaries of different perspectives, it is also necessary to improve the accuracy of clustering Gaussian components to ensure that Gaussian components of different sensor nodes are grouped into the same set.

[0048] The SOS clustering method that approaches the ideal solution is to use the probability of outliers (values ​​not in the same viewing area) and the number of sensor coverage as the original data information, assuming are Gaussian components with different groups, and the calculation of , is the relative closeness calculated by the TOPSIS-SOS method, if The ranking position of is not in the first 30%, which means that the node data are assigned to the same category, which can explain that the above Gaussian components are divided into the same cluster.

[0049] The TOPSIS-SOS method consists of four steps: data positive, data standardization, optimal solution and worst solution calculation, and relative proximity calculation:

[0050] (1) Data positive: Divide the data into extremely large indicators (the larger (more) the better), extremely small indicators (the smaller (fewer) the better), intermediate indicators (the closer to a certain value the better), and interval indicators. Find a data set that belongs to extremely small indicators, and divide the number of sensor coverage into interval indicators (falling in a certain area is best). According to the definition of data positive, the smaller the probability of outliers, the better. Then, the outlier probability value is positively processed to obtain extremely large indicators with larger values, the better.

[0051] (2) Data standardization: Standardize the number of sensors and the probability of abnormality to obtain the standardized number of sensors and the probability of abnormality, which are expressed in the form of a data matrix.

[0052] ;

[0053] in, — represents the number of sensors after normalization, — represents the standardized outlier probability, n Indicates the number of targets.

[0054] (3) Calculation of optimal and worst solutions: Take the largest number in each column of the above matrix to form an ideal optimal solution vector. Targets , calculate its distance from the optimal solution:

[0055] ;

[0056] in, represents the optimal solution vector, In the above matrix middle The maximum value of In the above matrix middle The maximum value of Indicates Targets The distance from the optimal solution is The value of is limited to 1 and 2, so and In fact, in order to solve the Targets The intermediate variable when measuring the distance from the optimal solution;

[0057] Take the smallest number in each column from the above matrix to form the ideal worst solution vector. Targets , and calculate its distance from the worst solution:

[0058] ;

[0059] in, represents the worst solution vector, In the above matrix middle The minimum value of In the above matrix middle The minimum value of Indicates Targets The distance from the worst solution is The value of is limited to 1 and 2, so and In fact, in order to solve the Targets The intermediate variable when measuring the distance from the worst solution.

[0060] (4) Calculate relative proximity: Calculate the Scoring of goals :

[0061] ;

[0062] Will Sort from low to high, and the top 30% of the targets are outliers. Use outliers as indicators to divide clusters of different sensor perspectives, and divide the Gaussian components corresponding to the targets into different clusters.

[0063] Through the above four steps, the Gaussian components are reasonably divided into clusters.

[0064] After completing the above TOPSIS-SOS outlier processing, it is necessary to cluster the Gaussian components. is the Gaussian component with different motion trajectories, calculate of ,if The ranking position is not in the first 30%, which means that the node data is assigned to the same category. It can be explained that the above Gaussian components are divided into the same cluster, and the Gaussian components that are not assigned to the same cluster are divided into other clusters.

[0065] According to the results of TOPSIS-SOS cluster analysis, the sensor multi-target posterior probability strength can be improved by step 1. The corresponding sub-strengths are expressed as: ;

[0066] in, Represents a sensor node i , j The local strength at represents the weight of the i-th Gaussian component of the sensor, Represents a mean and covariance The Gaussian probability density function is Indicates Sensors;

[0067] represents the number of Gaussian components of nodes i, j in the cluster, when >0, It indicates the number of Gaussian components in the cluster. Different Gaussian components need to be classified. = 0 means that there is no Gaussian component in the cluster and there is no need to classify different GCs. At this time, the corresponding sub-intensity can be simply expressed as: .

[0068] Based on the decomposition of independent identical distribution (IID) cluster density, the sub-intensities obtained by TOPSIS-SOS cluster analysis replace the sub-intensities obtained in the original state space, so that each cluster has sensor nodes belonging to the cluster to participate in the fusion.

[0069] Under the combined effect of TOPSIS-SOS clustering and boundary segmentation of difference perspective, the number of generated targets changes. At this time, the number of generated targets and spatial target density are calculated.

[0070] Through the above cluster analysis and boundary segmentation according to the difference perspective, the area segmented by calculating the segmentation cardinality distribution needs to be Together with boundary segmentation, :

[0071]

[0072]

[0073] in: represents the set of segmented regions, express Regions, represents the difference perspective segmentation area, M represents the total number of sets, N represents the Nth set, Indicates the segmented area that has been segmented. From 1 to N represents a collection, The corresponding random finite set of the above targets is approximately multi-Bernoulli, and the probability distribution is , there is a probability .

[0074] Step 3.2 Fusion process:

[0075] The GA fusion method is used for fusion, and the expression of multi-target probability distribution is:

[0076] ;

[0077] in, is the multi-target probability distribution, is a single target density, is the probability distribution of the target number, represents the number of clusters, n Indicates the number of targets.

[0078] The drone swarm is composed of A network of sensors ,in , yes A collection of nodes, is a set of links, for any two nodes , , if the node From the node Receive data, then there is an edge ,node The internal nodes of .

[0079] ;

[0080] in, represents the GA fusion result, represents the probability distribution of multiple targets in different regions, Multi-target probability distribution, Representation Node The consistency weight.

[0081] Each sensor node , at every time A given number of consistency (fusion) steps between the local density and the neighbor density are performed, and for each region the cardinality distribution given by the different sensors and the fused target state are computed separately.

[0082] Step 4: Merge the multi-target probability density distributions of all sensor view clusters obtained in step S3, and calculate the merged multi-target probability density distribution to characterize the distribution of the number of targets in the corresponding area.

[0083] In order to obtain the multi-target probability density distribution of GCs in multi-sensor multi-difference perspectives, the multi-target probability density distributions of different regions segmented above should also be merged. This process is to merge all fused sub-IID clustering processes into a single IID clustering process. It is obtained by summing all sub-intensity functions and convolving the corresponding cardinality distributions, and calculating the merged cardinality distribution and the fused target state.

[0084] The main content of the present invention is to improve the influence of non-co-viewing areas on nonlinear systems, improve the consistency of target information represented by the information to be fused collected by sensors in non-co-viewing areas, and solve the problem of the SOS algorithm being affected by the artificially set threshold. A SOS sorting method that approaches the ideal solution is proposed, and the outlier data and the number of sensor coverage from the difference perspective are simultaneously forward processed, standardized, and the optimal and worst solutions are calculated, as well as the relative proximity processing is calculated to improve the outlier detection accuracy from the difference perspective; the posterior intensity function is decomposed into multiple sub-intensities, and the distribution of the number of targets in the corresponding area is characterized by multi-Bernoulli reconstruction cardinality distribution. After segmentation and fusion, the TS-GM -CPHD method only runs the boundary outlier detection once to avoid the problems of missed detection and repeated detection. At the same time, it approaches the ideal random outlier detection algorithm, which can effectively improve the problem of artificially set thresholds in the traditional SOS algorithm and improve the detection accuracy.

Claims

1. A multi-sensor multi-view fusion method for approximating ideal random outlier detection, characterized in that: The method comprises the following steps: S1, receiving sensor 1- i The detection data is used, and the GM-CPHD filter is used for prediction and update to obtain the Gaussian component posterior probability intensity corresponding to multiple targets. i Represents the total number of sensors, using 1- i Number all sensors in sequence; S2, judging whether each target belongs to the same cluster of sensor perspective through the posterior probability strength of Gaussian components corresponding to multiple targets; S3. If all targets belong to the same sensor perspective cluster, the maximum a posteriori probability criterion is used to estimate the number of targets and calculate the multi-target probability density distribution of this sensor perspective cluster. S4. If the targets do not belong to the cluster of the same sensor perspective, the probability density distribution of multiple targets in clusters of different sensor perspectives is obtained by segmentation and fusion processing in turn; S5, merging the multi-target probability density distributions of all sensor view clusters obtained in steps S3 and S4, and calculating the merged multi-target probability density distribution to characterize the distribution of the number of targets in the corresponding area; The segmentation process includes, in sequence, performing boundary segmentation processing of difference perspectives and performing SOS clustering processing for approximating an ideal solution; The SOS clustering process approaching the ideal solution is specifically as follows: S51, data forward processing: divide the data into extremely large indicators, extremely small indicators, intermediate indicators and interval indicators, and divide the probability of abnormal values ​​into extremely small indicators, and divide the number of sensors into interval indicators; perform forward processing on the abnormal probability value to obtain an extremely large indicator; S52, data standardization processing: performing standardization processing on the number of sensors and the abnormal probability value, obtaining the standardized number of sensors and the abnormal value probability, and expressing them in the form of a data matrix; S53, optimal solution and worst solution processing: find the optimal solution distance in the data matrix and the worst solution distance ; S54, calculate relative proximity: calculate the Scoring of goals : Will Sort from low to high, and the top 30% of the targets are outliers. Use outliers as indicators to divide clusters of different sensor perspectives, and divide the Gaussian components corresponding to the targets into different clusters. S55. Cluster the Gaussian components corresponding to the targets to obtain multi-target probability distributions in different clusters.

2. The multi-sensor multi-view fusion method for approximating ideal random outlier detection according to claim 1, characterized in that: GM-CPHD filter passes ; Get a sensor k The posterior probability strength of the Gaussian components corresponding to the multiple targets in ,in, represents the target state, Indicates sensor k The number of Gaussian components of ; Indicates sensor No. The weights of the Gaussian components; The variance is , the expectation is Sensors No. Gaussian density function corresponding to the Gaussian components.

3. The multi-sensor multi-view fusion method for approximating ideal random outlier detection according to claim 2, characterized in that: The boundary segmentation process of the difference perspective is specifically as follows: sensors with different difference perspectives are divided into non-overlapping regions with different posterior probability intensities of multiple targets according to the boundaries of the difference perspectives of the sensors.

4. The multi-sensor multi-view fusion method for approximating ideal random outlier detection according to claim 3, characterized in that: The fusion process is specifically carried out by: ； The probability density distribution of multiple targets in different sensor view clusters is obtained; among them, represents the GA fusion result, represents the probability distribution of multiple targets in different clusters, Representation Node The consistency weight of each sensor node , Indicated by A network of sensors ,in , yes A collection of nodes, is a set of links, for any two nodes , if the node From the node Receive data, then there is an edge ,node The internal nodes of .

5. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 4 are implemented.

6. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

Images