Online data fusion method for multi-sensor information

Abstract

This invention discloses an online data fusion method for multi-sensor information, belonging to the field of radar signal processing. Specifically, it first establishes a communication scenario including a sensor end and a fusion center end. Each sensor at the sensor end performs unordered measurements and track trajectories for multiple target UAVs and sends the data to the fusion center end. Preliminary track association is performed through feature extraction. Then, the fusion center end uses an online spatial registration algorithm to perform spatiotemporal registration of the track association information and estimates the measurement and attitude deviations of each sensor to correct the tracks and original measurement values. Next, based on the sequential m-best track association algorithm, multiple tracks formed by different sensors measuring multiple targets are associated to obtain multiple tracks for the same target. Finally, comprehensive track management is performed on the multiple tracks for the same target to obtain the final fusion result of the multi-sensor data for that target. This invention improves the stability and accuracy of track tracking.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing, specifically a multi-sensor online data fusion method to improve the stability and accuracy of tracking tracks. Background Technology

[0002] In distributed information processing systems, track association is generally used to classify multiple tracks from different sensors, thereby eliminating duplicate tracks. Subsequently, fusion algorithms are used to achieve continuous, reliable, and high-precision tracking of multiple targets.

[0003] The trajectory association process can be described as follows: the target's trajectory is represented by position, velocity, variance, and other characteristic values. Each sensor sends its local trajectory to the fusion center. Fusion is performed in two ways at the fusion center: one is to merge the trajectories of the same batch of targets from various sensors according to certain criteria, forming a unified global trajectory; the other is to pair the local trajectories from each sensor with the existing system trajectories in the database to ensure that the paired target state and the state in the system trajectory originate from the same batch of targets.

[0004] Online spatial registration algorithms assume that system bias is a gradually changing value and use filtering methods to obtain real-time estimates of the bias. Sensor online registration algorithms based on unscented filtering outperform algorithms based on extended Kalman filters in both filtering accuracy and robustness, and can be better applied to bias estimation in nonlinear systems.

[0005] The sensor track association and salient target identification algorithm extends the classic Global Nearest Point (GNP) track association algorithm to a multi-sensor scenario. By sequentially performing pairwise GNP track association on multi-sensor data sorted by the number of observed targets from least to most, the algorithm finds the "chain" track association results of all targets from all track association results, i.e., the numbered track association results. At the same time, it finds the salient targets that can be detected by all sensors within the time slice, and re-integrates their measurement data before outputting it.

[0006] The sequential m-best algorithm is a heuristic approach to solving the SD allocation problem. It addresses a series of generalized 2-D allocation problems, where each iteration introduces tracks from the next sensor and correlates these tracks with previous results. To avoid convergence to a local minimum, the m-best solution from each iteration (based on the Murty algorithm) is preserved and used in the next iteration.

[0007] Integrated track management selectively utilizes measurement information from various sensors, resulting in a more stable and accurate integrated track compared to a single sensor tracking track. Summary of the Invention

[0008] This invention addresses the low accuracy problem of traditional single-sensor registration and tracking methods by proposing an online data fusion method based on multi-sensor information. This method, based on sequential m-best track association, ultimately improves the stability and accuracy of track tracking.

[0009] Step 1: Set up a communication scenario including the sensor end and the fusion center end. Each sensor on the sensor end performs unordered measurement and trajectory tracking on multiple targets, and sends its own trajectory information and measurement values ​​to the fusion center end.

[0010] The measurement data transmitted back to the fusion center by each sensor includes the target number it detected, the target track position information, and the target attitude.

[0011] Different sensors may assign different numbers to the same target.

[0012] Step 2: Extract features from the track information and measurement values ​​transmitted back to the fusion center and perform preliminary track association.

[0013] The specific steps are as follows:

[0014] Step 201, for the same time slice T arri The sensors are arranged in descending order of the number of targets based on the measurement data from each sensor, resulting in an array S. i i = 1, 2, ..., s, select sensor array S i The sensor ranked first in the middle serves as the benchmark.

[0015] Each time slice corresponds to a reference sensor;

[0016] Step 202: For each sensor in the array, use the GNP algorithm to convert the target location information uploaded by each sensor to the geocentric coordinate system;

[0017] For two adjacent sensors A and B in the array, the conversion formula is as follows:

[0018]

[0019]

[0020] in, This provides the three-dimensional position information of the i-th target observed and tracked by sensor A in the geocentric-ground-fixed coordinate system. X represents the three-dimensional position information of the j-th target observed and tracked by sensor B in the geocentric-ground-fixed coordinate system; truea(i) X represents the true position of the i-th target on sensor A at the current measurement moment; trueb(j)Let G(·) be the true position of the j-th target of sensor B at the current measurement moment; G(·) is the noise transformed from the measurement coordinate system to the common coordinate system; P i and Q j These are the covariance matrices of the noise measured by the two sensors, respectively. and These are the system biases for sensor A and sensor B, respectively.

[0021] Step 203, for the sensor array S i In the current stage, measurement data is allocated to the track, and the optimal allocation scheme with the highest total cost is calculated.

[0022] The formula for calculating the allocation cost is:

[0023]

[0024] R is the system bias covariance matrix, and a(i) represents the allocation status. a(i) > 0 indicates that the i-th target in sensor A is associated with the a(i)-th target in sensor B, and a(i) = 0 indicates that the i-th target in sensor A is not assigned. δx is the difference in deviations between the two sensors A and B, i.e., the difference in system bias. i For the calculation process, m is the number of targets detected by the sensor, and M is the arc between the allocation stages.

[0025] g is the set empirical threshold value, and the initial threshold value is:

[0026]

[0027] The threshold value is updated each time:

[0028]

[0029] Where, d min Let be the residual matrix.

[0030] Step 204: Save the sensor track matching information obtained from the optimal allocation scheme, combined with the track number, into the matching matrix, and extend it from point matching to track association.

[0031] Step 205: The matching matrices corresponding to each time slice are integrated using the "chain-type co-sensor method" to generate a "chain" of paths for the same target under different sensor observations, thus completing the expansion from dual-sensor to multi-sensor.

[0032] Step 3: The fusion center uses an online spatial registration algorithm to perform spatiotemporal registration of the track association information, and estimates the measurement and attitude deviations of each sensor to correct the track and the original measurement values.

[0033] The specific steps are as follows:

[0034] Step 301, for N s In a scenario where multiple sensors measure the position of a target in the same space, for the k-th time slice, the target position measurement value z obtained by sensor i is... i,k Perform modeling;

[0035] as follows:

[0036] z i,k =h i (x k )+β i,k +v i,k

[0037] Among them, h i (x k ) indicates a transformation of the target's coordinates in a common coordinate system, β i,k v represents the inherent system bias of the sensor with time slot number i in the k-th time slice. i,k This represents the sensor noise at time slot i, which is the kth time slice.

[0038] Step 302: Let the parameter to be estimated be ρ = [X, β], and obtain the result based on the measurement set. The likelihood function is simplified and then re-obtained using Taylor expansion:

[0039]

[0040] C is the normalization constant term;

[0041] For the initial measurements in the common reference coordinate system, β0 is the initial value of the bias vector estimate. Where X = x K This is the set of the target's true locations over the latest K time slices. To obtain the true value of the system deviation for time set K, we take the derivative with respect to β and set the result to zero to obtain the deviation estimate.

[0042]

[0043] β is the estimated measurement bias of the sensor system.

[0044] Step 4: After correction, the track is processed based on the sequential m-best track association algorithm to associate multiple tracks formed by different sensors for multi-target measurements, thereby obtaining multiple tracks for the same target.

[0045] For two adjacent sensors A and B, if sensor A has N1 tracks and sensor B has N2 tracks, and the associated events of different tracks are independent of each other, then the 2-D assignment problem of track association between the two sensors is described as the optimal solution of the following formula.

[0046]

[0047] c ij χ is the result of taking the negative logarithm of the likelihood function. ij Let i be a binary variable indicating whether track i and track j correspond to the same target.

[0048] The constraints are

[0049]

[0050] The initial iteration correlates the tracks from the first two sensors, and each iteration introduces the tracks from the next sensor and correlates these tracks with the previous results.

[0051] To avoid convergence to a local minimum, the m-Best value (based on the Murty algorithm) of each iteration is retained and used in the next iteration.

[0052] The general solution to the two-dimensional assignment problem is as follows:

[0053]

[0054]

[0055]

[0056] The second iteration incorporates the next sensor and performs a two-dimensional assignment of the results from the previous iteration using the tracking files from the sensor, generating the following solution:

[0057]

[0058] The best solution to the 2-D assignment problem is:

[0059]

[0060]

[0061]

[0062]

[0063] Step 5: Perform integrated track management on multiple tracks for the same target to obtain the final fusion result of the multi-sensor data for the target;

[0064] The advantages of this invention are:

[0065] 1) A multi-sensor information online data fusion method. Compared with the traditional extended Kalman filter-based method, this invention utilizes unscented filtering to obtain real-time estimation of the bias, avoiding the linear approximation process, thereby significantly improving accuracy and convergence speed.

[0066] 2) An online data fusion method based on multi-sensor information solves a series of generalized 2-D assignment problems. Each iteration of the algorithm introduces tracks from the next sensor and correlates these tracks with previous results. To avoid convergence to a local minimum, the m-optimal solution (based on the Murty algorithm) of each iteration is retained and used in the next iteration.

[0067] 3) A multi-sensor online data fusion method. Compared with the traditional single-sensor multi-target tracking problem, the integrated trajectory management of the present invention selectively utilizes the measurement information of each sensor, so that the final integrated trajectory is more stable and more accurate than the tracking trajectory of a single sensor. Attached Figure Description

[0068] Figure 1 This is a flowchart of an online data fusion method for multi-sensor information according to the present invention;

[0069] Figure 2 This is a diagram of the Global Nearest Point (GNP) algorithm based on the Global Nearest Point Model of this invention.

[0070] Figure 3 This is a diagram showing the trajectory association results based on sensor trajectory association and salient target recognition in this invention.

[0071] Figure 4 This is a flowchart of the algorithm based on integrated track management in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0073] This invention provides an online data fusion method for multi-sensor information, such as... Figure 1 As shown, the specific steps are as follows:

[0074] Step 1: Set up a communication scenario including the sensor end and the fusion center end. Each sensor on the sensor end performs unordered measurement and trajectory tracking on multiple targets, and sends its own trajectory information and measurement values ​​to the fusion center end.

[0075] The measurement data transmitted back to the fusion center by each sensor includes the target number it detected (different sensors may have different numbers for the same target), target trajectory and position information, and target attitude.

[0076] Different sensors may assign different numbers to the same target.

[0077] Step 2: Extract features from the track information and measurement values ​​transmitted back to the fusion center and perform preliminary track association.

[0078] The specific steps are as follows:

[0079] Step 201, for the same time slice T arri The sensors are arranged in descending order of the number of targets based on the measurement data from each sensor, resulting in an array S. i i = 1, 2, ..., s, select sensor array S i The sensor ranked first in the middle serves as the benchmark for finding a unified path to the same target in subsequent searches.

[0080] Each time slice corresponds to a reference sensor;

[0081] Step 202: For each sensor in the array, use the Global Nearest Model-based trajectory association (GNP) algorithm to transform the target location information uploaded by each sensor to the geocentric coordinate system.

[0082] like Figure 2 As shown, for two adjacent sensors A and B in the array, assume that these two sensors simultaneously track a set of unknown numbers of targets in this space, where sensor A tracks a subset N of the targets. A ={1,...,n A The subset of targets tracked by sensor B is N. B ={1,...,n B Without loss of generality, assume that there are n... A ≤n B ≤N.

[0083] The conversion formula is as follows:

[0084]

[0085]

[0086] in, This provides the three-dimensional position information of the i-th target observed and tracked by sensor A in the geocentric-ground-fixed coordinate system. X represents the three-dimensional position information of the j-th target observed and tracked by sensor B in the geocentric-ground-fixed coordinate system; truea(i)X represents the true position of the i-th target on sensor A at the current measurement moment; trueb(j) Let G(·) be the true position of the j-th target of sensor B at the current measurement moment; truea and trueb are the indices of the same target corresponding to the actual measurements of sensors A and B. G(·) is the noise transformed from the measurement coordinate system to the common coordinate system, and P is the noise. i and Q j These are the covariance matrices of the noise measured by the two sensors, respectively. and These are the system biases for sensor A and sensor B, respectively.

[0087] Step 203, for the sensor array S i In the current stage, measurement data is allocated to the track, and the optimal allocation scheme with the highest total cost is calculated.

[0088] Randomly generate K path association schemes and calculate their total cost J. K After fully allocating the first target of sensor A, the total cost of all current allocations is calculated. If the total cost is less than the current cost, allocation continues. At each stage, the allocation scheme with the highest cost is calculated, and the K schemes with the highest cost are updated at each time step until all allocations are performed. The scheme with the highest total cost is then identified as the optimal allocation scheme, and the sensor's measurement data is allocated to the track.

[0089] The formula for calculating the allocation cost is:

[0090]

[0091] R is the systematic deviation covariance matrix, calculated using the following formula:

[0092]

[0093]

[0094] S i =P i +Q a(i)

[0095] a(i) represents the allocation status. a(i) > 0 indicates that the i-th target in sensor A is associated with the a(i)-th target in sensor B, and a(i) = 0 indicates that the i-th target in sensor A is not allocated. The same target cannot be allocated twice, so let a(i) ≠ a(j). This is the difference in deviation between the two sensors A and B, i.e., the difference in system deviation.

[0096] δx i These are intermediate variables in the calculation process; m is the number of targets detected by the sensor, and M is the arc between allocation stages.

[0097] When the i-th target in sensor A is not assigned, g is a set empirical threshold value. Setting this threshold value is to filter out factors such as false alarms; the initial threshold value is:

[0098]

[0099] The threshold value is updated each time:

[0100]

[0101] Where d min The residual matrix The smallest determinant for all i and j.

[0102] Step 204: Save the sensor point matching information obtained from the optimal allocation scheme, combined with the track number, into the matching matrix, and extend the point matching to track association.

[0103] Step 205: For each time slice, the matching matrix is ​​used to integrate the trajectory association results of all sensors using the "chain-type co-sensor method" to generate a "chain" of paths for the same target under different sensor observations, thus completing the expansion from dual-sensor to multi-sensor.

[0104] The point matching information obtained in the above steps is combined with the track number and stored in a matching matrix. The i-th row of the matching matrix shows the matching situation between the i-th sensor and the (i+1)-th sensor, thus determining the matching situation between two adjacent sensors S. i and S i+1 The process of establishing a unified relationship between multiple target data contained in each data set is called "chaining".

[0105] The unity relationship can be expressed by the trajectory association result a(p) = q, which means that target p in sensor i and target q in sensor i+1 are the same target. To find the tracking path of a single target across all sensors, the trajectory association results are organized into a network form. At this point, the data of target j in a single sensor i is considered as a node V. ij The unity relation a(p) = q is a node V ip With V (i+1)q The edge E between i,p,q Therefore, this diagram is called a "multi-sensor, multi-target unified recognition network". Figure 3 The image shown is a diagram illustrating the trajectory association results.

[0106] Step 3: The fusion center uses an online spatial registration algorithm to perform spatiotemporal registration of the track association information, and estimates the measurement and attitude deviations of each sensor to correct the track and the original measurement values.

[0107] Due to asynchronous sampling and communication delays between sensors, asynchronous data from different sensors is generated. This data is then converted into synchronous data using an online spatial registration algorithm to obtain a real-time estimate of the deviation.

[0108] The specific steps are as follows:

[0109] Step 301, for N s In a scenario where multiple sensors measure the position of a target in the same space, for the k-th time slice, the target position measurement value z obtained by sensor i is... i,k Perform modeling;

[0110] as follows:

[0111] z i,k =h i (x k )+β i,k +v i,k

[0112] Among them, h i (x k ) indicates a transformation of the target's coordinates in a common coordinate system, β i,k v represents the inherent system bias of the sensor with time slot number i in the k-th time slice. i,k This represents the sensor noise at time slot i, which is the kth time slice.

[0113] Step 302: The fusion center obtains all radar position measurement sets for the first k time slices. Proceed with further processing. Assume we are currently at the k-th stage since system startup. c In each time slice, the target's trajectory measurements will be used for spatial registration. Clearly, after the system has been running for a sufficiently long time, if all the target's trajectory measurements... As input to the maximum likelihood space registration algorithm, it will be due to k c The value becomes too large to be easily calculated. Therefore, measurements on certain time slices of the target are selected as inputs for spatial registration, and these selected times are denoted as the time set K. In this method, the latest K time slices are selected as elements of the time set K, i.e. The filtered measurement set is denoted as Z. This approach avoids the uncontrolled increase in computational load for spatial registration over time and ensures that the impact of erroneous track association results on the spatial registration results is limited.

[0114] Let the parameter to be estimated be ρ=[X,β], where X=x K This is the set of the target's true locations over the latest K time slices. To obtain the true value of the system deviation at time set K, we can obtain the value based on the measurement set. Likelihood function:

[0115]

[0116] So, among them are...

[0117]

[0118]

[0119] k∈K, the same below

[0120] According to the measurement z i,k relative to the target's true position x k The interrelationship can be obtained Its in υ i,k The Taylor expansion at =0 can be expressed as:

[0121]

[0122] in Based on the above equation, the likelihood function can be rewritten as follows:

[0123]

[0124] in, For the measurement set in a common coordinate system, C is the normalization constant term;

[0125] Step 303: Expanding the above equation yields...

[0126]

[0127] Using matrix identities

[0128] x T Ax-2x T B+B T A -1 B = (xA) -1 B) T A(xA -1 B)

[0129] Where A is a symmetric matrix, the likelihood function can be simplified to:

[0130] The likelihood function can be simplified to:

[0131]

[0132] in,

[0133] In the above formula, when Time-likelihood function Therefore, to obtain the maximum value, This is the maximum likelihood estimate of the target state at time k. The right-hand side of the above equation can be restated as:

[0134]

[0135] in

[0136] The symbol {·} i,j This indicates that the matrix is ​​in the range i,j=1,2,…,N S The subarray at the location.

[0137] Step 304, Order Then the likelihood function in the above equation can be rewritten as:

[0138]

[0139] consider Due to β i The changes caused by the disturbance, the measurements after the disturbance, and the deviations are respectively and β 0i Using Taylor expansion, we can obtain In this case, we can obtain:

[0140]

[0141] The above formula can be used to iteratively calculate the deviation estimate. Extending this formula to multiple sensor types, we get:

[0142]

[0143] in

[0144] Step 305 Let β0 be the initial measurement in a common reference coordinate system, and β0 be the initial bias estimate. Substituting the above equation into the likelihood function, we can obtain the following:

[0145]

[0146] Taking its derivative with respect to β and setting the result to zero, we can obtain the bias estimate as follows:

[0147]

[0148] The above equation provides a method for estimating the deviation when the disturbance is known. When the disturbance is unknown, β0 can be initialized to 0, and the deviation can be estimated iteratively using the above equation until it converges; β is the estimated measurement deviation of the sensor system.

[0149] The above-obtained deviation estimates In reality, this is an offline estimation of the radar system bias at time set K. However, by continuously updating time set K, the bias estimate... It will also be continuously updated, but because the systematic bias β changes slowly in practical applications, the obtained bias estimate will be... Closer to k c The system deviation β at time t. This deviation estimate is then used in the output and application. Previously, the reliability of this value should also be quantified. Record the deviation estimate. The credibility parameter is calculate The formula is:

[0150]

[0151] Calculate Then, its value can be used to compare the deviation estimates calculated before and after. and Reliability and like Explain the newly obtained deviation estimate If the confidence level is too low, there may be some errors, and the previous estimate of the system bias should be discarded and used instead; otherwise, the updated bias estimate should be adopted. As the output of the algorithm.

[0152] Step 4: After correction, the track is processed based on the sequential m-best track association algorithm to associate multiple tracks formed by different sensors for multi-target measurements, thereby obtaining multiple tracks for the same target.

[0153] The specific steps are as follows:

[0154] Step 401: For N sensor arrays S1, S2, ..., S N State estimation for the same objective: x1, x2, ..., x N and the corresponding covariance matrices P1, P2, ..., P N Based on the maximum likelihood criterion, assuming that the paths are uncorrelated, the cross-covariance P of each sensor is... i,j =0, i≠j, the fused target state estimate x f,N and the corresponding covariance estimate P f,N for

[0155]

[0156]

[0157] Step 402, if using T i j Let represent the tracking result of sensor j on target i, which includes the target state estimate and its covariance matrix. Then the tuple is:

[0158] Γ={T i j |j=1,2,,N s}

[0159] This represents the tracking results of all sensors for the same target i.

[0160] The track association algorithm needs to determine whether two tracks from different sensors belong to the same target. For two sensors, a hypothesis test can be formed between targets i and j.

[0161]

[0162] If the null hypothesis H0 holds, then the two tracks represent the same real object. In multi-target, multi-sensor scenarios, a global correlation hypothesis H containing tuples Γ is created, with the goal of finding the most probable hypothesis among all global hypothesis sets. A method based on the maximum likelihood criterion is given below. First, the likelihood function of the tuples can be expressed as...

[0163]

[0164] The most likely global hypothesis can be obtained by finding the tuple that maximizes the likelihood function:

[0165]

[0166] Since the true value of the target state is unknown, an accurate likelihood function cannot be obtained. In a two-sensor scenario, this is not a problem, because the true state of the target can be obtained through the difference Δ between the two tracks. ij =x i -x j To replace it. If these two paths belong to the same real target, then Δ ij The two objectives follow a zero-mean Gaussian distribution with covariance P. The likelihood functions of these two objectives are:

[0167]

[0168] If the two matched tracks are unrelated, the likelihood function of the two targets is:

[0169]

[0170] Assuming there is no correlation between the tracks, the formula simplifies as follows:

[0171]

[0172] Step 403: For a simplified scenario of two adjacent sensors A and B, if sensor A has N1 tracks and sensor B has N2 tracks, and the events associated with different tracks are independent of each other, then the 2-D assignment problem of track association between the two sensors can be described as the optimal solution of the following formula.

[0173]

[0174] c ij Let c be the result of taking the negative logarithm of the likelihood function, and c ij =-lnΛ(Γ); χ ij Let i be a binary variable indicating whether track i and track j correspond to the same target.

[0175] The constraints are

[0176]

[0177] Some algorithms, such as the JVC algorithm, the Hungarian algorithm, and the auction algorithm, can be used to solve this constrained problem. However, when the number of sensors is greater than two, the 2-D allocation problem transforms into an SD allocation problem. This problem can be described as the following constraint optimization...

[0178]

[0179] The constraints are

[0180]

[0181] in for:

[0182]

[0183] Step 404: In each iteration, introduce tracks from the next sensor and correlate these tracks with previous results. To avoid convergence to a local minimum, retain the m-Best result (based on the Murty algorithm) from each iteration and use it in the next iteration.

[0184] The SD allocation problem given above is an NP-hard problem, and a suboptimal solution can be obtained through the Lagrange relaxation method or the m-Best allocation algorithm. This embodiment will introduce a sequential m-Best allocation algorithm, which solves a series of generalized 2-D allocation problems. The initial iteration correlates the tracks from the first two sensors. The general solution to the 2D allocation problem is as follows:

[0185]

[0186]

[0187]

[0188] Where Γ 2,i Let T be the i-th tuple, k be the number of solutions to the tuple, and T be the number of solutions to the tuple. i j Γ represents the state estimate of the j-th sensor for the i-th target. 2,1 ={T1 1 T3 2} indicates that the information of track 1 received by the first sensor matches the information of track 3 received by the second sensor, Γ 2,2 ={T2 1 T1 2} indicates that the information of track 2 received by the first sensor matches the information of track 1 received by the second sensor, Γ 2,3 ={T3 1 T2 2} indicates that the information of track 3 received by the first sensor matches the information of track 2 received by the second sensor.

[0189] Step 405: The second iteration merges the next sensor and performs a two-dimensional assignment on the results of the previous iteration using the tracking files from the sensor, generating the following solution:

[0190]

[0191] The best solution to the 2-D assignment problem is:

[0192]

[0193]

[0194]

[0195]

[0196] The above results indicate that the information of track 1 received by the first sensor matches the information of track 3 received by the second sensor and the information of track 1 received by the third sensor; the information of track 2 received by the first sensor matches the information of track 1 received by the second sensor and the information of track 3 received by the third sensor; and the information of track 3 received by the first sensor matches the information of track 2 received by the second sensor and the information of track 2 received by the third sensor.

[0197] Step 5: Perform integrated track management on multiple tracks for the same target to obtain the final fusion result of the multi-sensor data for the target;

[0198] like Figure 4 As shown, the specific steps are as follows:

[0199] Step 501: Perform single-moment trajectory association based on sequential m-best on the multi-sensor target tracks after salient target identification and spatiotemporal registration.

[0200] Step 502: Perform optimal trajectory fusion on the correlation results at a single moment. At this point, the multi-sensor multi-target tracking problem can be transformed into a single-sensor multi-target tracking problem. In order to connect the optimal fusion results at a single moment in time intervals to form a stable comprehensive track, the following steps are taken.

[0201] Step 503: Extend the track-oriented MHT algorithm to distributed multi-sensor target tracking, and use the MHT algorithm to track the optimal fused point track to form the final comprehensive track.

Claims

1. A multi-sensor information online data fusion method, characterized in that, The specific steps are as follows: Step 1: Set up a communication scenario including the sensor end and the fusion center end. Each sensor on the sensor end performs unordered measurement and trajectory tracking for the multi-target UAV, and sends its own trajectory information and measurement values ​​to the fusion center end. Step 2: Extract features from the track information and measurement values ​​transmitted back to the fusion center, and perform preliminary track association; The specific steps are as follows: Step 201, for the same time slice The sensors are arranged in descending order of the number of targets based on their measurement data, resulting in an array. Select sensor array The sensor ranked first in the middle is used as a benchmark; Each time slice corresponds to a reference sensor; Step 202: For each sensor in the array, use the GNP algorithm to convert the target location information uploaded by each sensor to the geocentric coordinate system; For two adjacent sensors A and B in the array, the conversion formula is as follows: in, The first sensor A observes and tracks The three-dimensional position information of the target in the Earth-centered Earth-fixed coordinate system. The first one observed and tracked by sensor B Three-dimensional position information of a target in a geocentric coordinate system; To determine the first value of sensor A at the current measurement moment. The true location of the target; To determine the first value of sensor B at the current measurement moment. The true location of the target; This is noise transferred from the measurement coordinate system to the common coordinate system. and These are the covariance matrices of the noise measured by the two sensors, respectively. and These are the system biases for sensor A and sensor B, respectively. Step 203, for the sensor array In the current stage, measurement data is allocated to the track, and the optimal allocation scheme with the highest total cost is calculated. The formula for calculating the allocation cost is: It is the system bias covariance matrix. Indicates the allocation status. Indicates the first in sensor A The target and sensor B, the first One target association, Indicates the first in sensor A One target has not been assigned; This is the difference in deviation between the two sensors A and B, i.e., the difference in system deviation. These are intermediate variables in the calculation process. This represents the number of targets detected by the sensor. For the arc between allocation phases; The initial threshold value is set as the experience threshold. The threshold value is updated each time: in The residual matrix; Step 204: Save the sensor track matching information obtained from the optimal allocation scheme, combined with the track number, into the matching matrix, and extend it from point matching to track association; Step 205: The matching matrices corresponding to each time slice are integrated using the "chain-type co-sensor method" to generate a "chain" of paths for the same target under different sensor observations, thus completing the expansion from dual-sensor to multi-sensor. The "chain-type co-sensor method" is used to integrate the trajectory correlation results of all sensors; The matching matrix of the first Behavior No. The sensor and the first The matching status of each sensor is used to determine the matching status of two adjacent sensors. and The process of establishing a unified relationship between multiple target data contained in each data set is called "chaining". Step 3: The fusion center uses an online spatial registration algorithm to perform spatiotemporal registration of the track association information, and estimates the measurement and attitude deviations of each sensor to correct the track and the original measurement values. The specific steps are as follows: Step 301, targeting In a scenario where multiple sensors measure the position of a target in the same space, for the first... A time slice, numbered as The target position measurement value obtained by the sensor Perform modeling; as follows: in, This indicates a transformation of the target's coordinates in a common coordinate system. Indicates the first The time slice is numbered as follows The inherent system bias of the sensor, Indicates the first The time slice is numbered as follows Sensor noise; Step 302: Record the parameter to be estimated as... , to obtain based on measurement set The likelihood function is simplified and then re-obtained using Taylor expansion: , It is the normalized constant term; , For initial measurements in a common reference coordinate system, Initial values ​​for the bias vector estimation ,in For the latest The set of actual target locations for each time slice. For time set The true value of the systematic deviation; for By taking the derivative and setting the result to zero, we obtain the bias estimate. This is the estimated measurement deviation of the sensor system; Step 4: After correction, the track is processed based on the sequential m-best track association algorithm to associate multiple tracks formed by different sensors for multi-target measurements, and obtain multiple tracks for the same target. Step 5: Perform integrated track management on multiple tracks for the same target to obtain the final fusion result of the multi-sensor data for the target.

2. The online data fusion method for multi-sensor information as described in claim 1, characterized in that, In step one, The measurement data transmitted back to the fusion center by each sensor includes the target number it detected, the target track position information, and the target attitude; different sensors may have different numbers for the same target.

3. The online data fusion method for multi-sensor information as described in claim 1, characterized in that, Step four specifically involves: For two adjacent sensors A and B, if sensor A has A flight path, sensor B has Given multiple tracks, and that the associated events of different tracks are independent of each other, the 2-D assignment problem of track association between two sensors can be described as the optimal solution of the following equation: This is the result of taking the negative logarithm of the likelihood function. For the track and flight path Whether they correspond to the same target (binary variables) ; The constraints are The initial iteration correlates the tracks from the first two sensors, and each iteration introduces the tracks from the next sensor and correlates these tracks with the previous results; To avoid convergence to a local minimum, the m-Best value from each iteration is retained and used in the next iteration; The general solution to the two-dimensional assignment problem is as follows: The second iteration incorporates the next sensor and performs a two-dimensional assignment of the results from the previous iteration using the tracking files from the sensor, generating the following solution: The best solution to the 2-D assignment problem is: 。

Images