An IMM track filtering method and system based on D-S evidence theory and multi-stage adaptive wild value detection

Abstract

The application provides an IMM track filtering method and system based on D-S evidence theory and multi-stage adaptive wild value detection, and the method comprises the following steps: step 1, receiving track observation data, performing data integrity check and complement; step 2, performing multi-priority track association on the track observation data; step 3, filtering the track observation data with successful association by using a multi-motion model matching scheme to obtain the final IMM track; and the system is used for realizing the above method. Through the fusion of D-S evidence theory and adaptive wild value detection, the model estimation precision, the system robustness and the track association accuracy are effectively improved, the problems that the prior art is susceptible to wild value interference and poor adaptability are solved, and the application is suitable for the field of aerial target tracking.

Description

Technical Field

[0001] This invention relates to an IMM track filtering method and system, and in particular to an IMM track filtering method and system based on DS evidence theory and multi-level adaptive outlier detection. Background Technology

[0002] In airborne target tracking systems, due to the uncertainty of target motion patterns (such as constant speed, acceleration, turning, etc.), a single-model Kalman filter is difficult to adapt to all motion states. The Interactive Multiple Model (IMM) algorithm significantly improves tracking accuracy by using multiple models in parallel and dynamically adjusting model probabilities based on observation data.

[0003] However, existing IMM algorithms still have the following problems in complex environments:

[0004] First, model selection mainly relies on the observation likelihood function, which is easily affected by outlier observations (i.e., outliers), leading to inaccurate model probability updates.

[0005] Secondly, outlier detection typically uses the Mahalanobis distance method with a fixed threshold, which lacks adaptability to different motion states and noisy environments.

[0006] Secondly, the track association strategy is relatively simple, and it is prone to false association when tracks intersect densely.

[0007] Finally, the lack of an effective automatic completion mechanism for incomplete observation data affects the stability and continuity of the filter. Summary of the Invention

[0008] Purpose of the invention: The technical problem to be solved by the present invention is to provide an IMM track filtering method and system based on DS evidence theory and multi-level adaptive outlier detection, which addresses the shortcomings of the existing technology.

[0009] To address the aforementioned technical problems, this invention discloses an IMM track filtering method and system based on DS evidence theory and multi-level adaptive outlier detection. The method includes the following steps:

[0010] Step 1: Receive track observation data and perform data integrity checks and completion.

[0011] Step 2: Perform multi-priority track association on the track observation data;

[0012] Step 3: For the successfully associated track observation data, a multi-motion model matching scheme is used for filtering to obtain the final IMM track.

[0013] Furthermore, the track observation data mentioned in step 1 includes:

[0014] Receives time, target location, velocity information, and target identification information; among which,

[0015] Target location includes longitude, latitude, or projected coordinates x and y; velocity information includes ground speed. Heading angle Or velocity component , The target identification information includes the flight number, 24-digit address code, and secondary code SSR.

[0016] Furthermore, the multi-priority track association performed on the track observation data in step 2 includes:

[0017] Step 2-1: Perform track association with priority 1, that is, perform precise matching using 24-bit address codes. If no match is found, proceed to step 2-2.

[0018] Step 2-2: Perform track association with priority 2, that is, match precisely by flight number. If no match is found, proceed to step 2-3.

[0019] Steps 2-3 are executed, involving track association at priority level 3, which involves calculating a weighted score based on position, secondary code, heading, and speed. The matching process is as follows:

[0020] Step 2-3-1, calculate the weighted score, as follows:

[0021]

[0022] in, For secondary code matching score, The score is the Euclidean distance of the location. Score based on speed difference. Score for heading angle difference. , , and These are the weighting coefficients;

[0023] Step 2-3-2: Select the track with the lowest weighted score as the best candidate track;

[0024] Step 2-3-3: If the weighted score of the best candidate track is less than the dynamic acceptance threshold, the trimming is successful; otherwise, the current observation is considered to correspond to a new target.

[0025] Furthermore, the IMM filtering using a multi-motion model matching scheme described in step 3 includes:

[0026] Step 3-1 involves preprocessing the successfully correlated track observation data and pre-setting different motion models for IMM filtering; the details are as follows:

[0027] Calculate the sampling time interval T between the currently associated successfully associated track observation data and the valid observation data of the previous moment. If T is negative or the time interval of the same source is less than the minimum valid interval, discard the observation; otherwise, construct the observation vector z and the observation noise matrix R at the same time, and pre-configure the motion model, which includes the uniform velocity model CV, the uniform acceleration model CA, the left turn model CTL, and the right turn model CTR.

[0028] Step 3-2: Use DS evidence theory to estimate the currently dominant motion model and correct the motion model probability of the previous time step;

[0029] Step 3-3: Perform IMM interactive mixing based on the motion model probabilities of the previous time step to obtain the initial mixing state and initial mixing covariance of each motion model.

[0030] Steps 3-4: Calculate the prior predictions of each model based on the mixed initial state and mixed initial covariance of each motion model, and fuse the prior predictions of each motion model to obtain the outlier detection benchmark.

[0031] Steps 3-5: Perform outlier detection based on the outlier detection benchmark;

[0032] Steps 3-6: Based on the outlier detection results, perform intelligent observation fusion and execute Kalman updates for each motion model to obtain the innovation vector and innovation covariance;

[0033] Steps 3-7: Update the motion model probabilities based on the innovation vector and innovation covariance, and fuse the posterior states of each motion model;

[0034] Steps 3-8: The fusion results of the posterior states of each motion model and the outlier detection results of Step 3-5 are written into the intermediate state of the current track, and the fused posterior states are converted into the filtered track output.

[0035] Furthermore, step 3-3, which involves performing IMM interactive mixing based on the motion model probability from the previous moment, includes:

[0036] Let the model probability of the i-th motion model at the previous time step after correction in step 3-2 be... The transition probability of model i to model j is Then the j-th motion model at the current time Predictive model probability for:

[0037]

[0038] Given the current target model j, what is the mixture probability of model i at the previous time step? for:

[0039]

[0040] Based on the mixing probability, the posterior states and posterior covariances of each motion model at the previous time step are weighted and mixed to obtain the mixed initial state of the j-th motion model. and mixed initial covariance The following are examples:

[0041]

[0042]

[0043] in, Let be the predicted model probability of the j-th motion model. Let be the mixture probability of model i at the previous time step, given that the current model is j. and Let be the mixed initial state and mixed initial covariance of the j-th motion model, respectively.

[0044] Furthermore, steps 3-4, which involve calculating the prior predictions of each motion model based on the mixed initial state and mixed initial covariance, and fusing the prior predictions of each motion model to obtain the outlier detection benchmark, include:

[0045] Step 3-4-1, for the j-th motion model, its prior state estimation and prior covariance prediction They are respectively:

[0046]

[0047]

[0048] in, For the prior state estimate of the j-th motion model at the current time k, Let j be the initial state of the mixed motion model. Let J be the state transition matrix of the j-th motion model. Let be the prior state covariance of the j-th motion model. Let be the initial covariance of the j-th motion model. Let J be the process noise covariance of the j-th motion model;

[0049] Step 3-4-2: Map the prior states of motion models with different state dimensions to a unified fusion state space; let the predicted model probability of the j-th motion model at the current time be... The prior state of the j-th motion model after mapping to the unified fusion state space is: If the prior covariance is P_j(k|k-1), then the fused prior states... for:

[0050]

[0051] Fusion of prior covariance for:

[0052]

[0053] in, This represents the prior state after fusion. The prior covariance after fusion Let be the predicted model probability of the j-th motion model at the current time.

[0054] The fusion of prior states and variance serves as the benchmark for outlier detection.

[0055] Furthermore, the fusion of prior predictions from various motion models described in steps 3-5 includes:

[0056] Extract position observations from the current observation vector. And based on the fusion of prior states Constructing the position observation matrix Calculate the location information and the location information covariance;

[0057] If the current observation vector contains a velocity component, then extract the velocity observation. And based on the fusion of prior states Constructing the velocity observation matrix Calculate the velocity information and velocity information covariance;

[0058] Calculate the Mahalanobis distance of the new information based on the new information and its corresponding covariance. :

[0059]

[0060] in, For the innovation vector, The new information covariance matrix;

[0061] Based on the fusion of prior covariance matrix Computational system uncertainty and confidence level The details are as follows:

[0062]

[0063]

[0064] in, Based on confidence level, The uncertainty scaling coefficient. This is the uncertainty scaling factor. To observe the noise covariance matrix, tr(·) represents the trace of the matrix, and clip(·) represents limiting the confidence level to a preset minimum value. and maximum value between;

[0065] Based on confidence level Determine the basic detection threshold The acceptance threshold is obtained by adjusting the data based on historical Mahalanobis distance statistics. and rejection threshold ;

[0066] According to Mahalanobis distance The relationship with the threshold is used to make a three-level decision of acceptance, deweighting, and rejection, as shown below:

[0067]

[0068] in, This indicates that the result is accepted. This indicates that the outlier detection result is "reject". This indicates that the outlier detection result is reduced in weight, and the reduction weight w is calculated based on the position of the Mahalanobis distance D of the new information between the acceptance threshold, buffer threshold, and rejection threshold.

[0069] Furthermore, the intelligent observation fusion based on the outlier detection results and the execution of Kalman updates for each motion model described in steps 3-6 include:

[0070] When the location observation detection result is REJECT, it is considered that the current location observation is unavailable, the location observation will not be used this time, the current observation update will be skipped, and the prior prediction results of each motion model will be directly retained.

[0071] When the position observation results in ACCEPT and the velocity observation results in REJECT, or when the current observation does not contain a velocity component, only the position observation is used to construct the enhanced observation vector. Enhanced observation matrix and enhance the observation noise matrix ;

[0072] When both position and velocity observations are accepted for ACCEPT, an enhanced observation vector is constructed by combining position and velocity observations. Enhanced observation matrix and enhance the observation noise matrix ;

[0073] When the position observation is downweighted and the velocity observation is available, the position observation is still used, and its observation noise is adjusted according to the position downweighting weight; when the position observation is downweighted and the velocity observation is rejected or does not exist, only the downweighted position observation is used for updating; when the velocity observation is downweighted, the velocity observation is still used, and its observation noise is adjusted according to the velocity downweighting weight.

[0074] For the j-th motion model, the system is based on the prior state obtained in steps 3-4. Prior covariance and the constructed enhanced observation vector Enhanced observation matrix and enhance the observation noise matrix Calculate new information New information covariance Kalman gain Updated model state estimation and model covariance They are represented as follows:

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] in, Let J be the Kalman gain of the j-th motion model. It is the identity matrix. and These are the posterior state estimate and posterior covariance of the j-th motion model at the current time, respectively.

[0081] If a valid observation component exists, the Kalman update is not performed, and the prior state and prior covariance of the j-th motion model are used as the output state of the model at the current time.

[0082] Furthermore, steps 3-7, which involve updating the motion model probabilities based on the innovation vector and innovation covariance and fusing the posterior states of each motion model, include:

[0083] For the j-th motion model, calculate its likelihood function:

[0084]

[0085] in, Let be the likelihood function of the j-th motion model at the current time k. For the j-th motion model, the innovation vector is obtained based on enhanced observations. The corresponding new information covariance matrix, This means that the mean is the zero vector. The multivariate Gaussian probability density function of the covariance matrix in The value at;

[0086] Update the model probability at the current time step based on the likelihood function of each motion model and the predicted model probability in step 3-3:

[0087]

[0088] in, Let be the predicted model probability of the j-th motion model at the current time. Let be the posterior model probability of the j-th motion model at the current time.

[0089] The posterior state estimates of each motion model are weighted and fused according to the posterior model probability to obtain the fused posterior state at the current time. , means as follows:

[0090]

[0091] The posterior covariance after fusion is calculated and expressed as follows:

[0092]

[0093] in, and Let be the posterior state estimate and posterior covariance of the j-th motion model at the current time, respectively. and These are the posterior state estimate and posterior covariance after fusion at the current time, respectively.

[0094] This invention also proposes an IMM track filtering system based on DS evidence theory and multi-level adaptive outlier detection to implement the aforementioned method, comprising:

[0095] The track association module performs multi-priority track association.

[0096] The IMM filtering module performs multi-model interaction and fusion.

[0097] The DS evidence theory module estimates the type of the current dominant motion model;

[0098] The outlier detection module performs multi-level adaptive outlier detection.

[0099] The intelligent observation fusion module dynamically selects the observation type and weight based on the outlier detection results.

[0100] The data completion module automatically completes missing track data.

[0101] Beneficial effects:

[0102] 1. This invention integrates instantaneous motion features and historical statistical information through DS evidence theory, reducing the interference of abnormal observations on model probability updates and improving the model recognition accuracy.

[0103] 2. This invention employs a multi-level adaptive outlier detection mechanism, which can dynamically adjust the detection threshold based on system status and historical information, effectively suppressing the influence of outliers of varying intensities.

[0104] 3. This invention employs a multi-priority association strategy combined with dynamic threshold adjustment, which significantly reduces track overlap and false association.

[0105] 4. This invention employs a threshold adjustment and intelligent observation fusion strategy based on time-series statistics, enabling the system to better adapt to environmental changes and sensor noise variations.

[0106] 5. This invention, through its data completion module, can automatically process missing observation data, ensuring the continuity of the filtering process. Attached Figure Description

[0107] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0108] Figure 1 This is a diagram showing the overall architecture of the system of the present invention.

[0109] Figure 2 This is a flowchart of the motion model estimation driven by DS evidence theory in this invention.

[0110] Figure 3 This is a flowchart of the multi-level adaptive outlier detection in this invention.

[0111] Figure 4 This is a decision-making flowchart for the intelligent observation fusion strategy in this invention.

[0112] Figure 5 This is a flowchart of the multi-priority track association in this invention. Detailed Implementation

[0113] This invention provides an IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection, such as... Figure 1 As shown, it includes the following steps:

[0114] Step 1: Receive track observation data and perform data integrity checks and supplementation;

[0115] The trajectory observation data is a single target observation record from ADS-B, radar, or ASTERIX messages, and includes at least the reception time, target position, velocity information, and target identification information. The target position includes longitude, latitude, or projected coordinates (x, y); the velocity information includes ground speed. Heading angle Or velocity component , The target identification information includes one or more of the following: flight number, 24-bit address code, and secondary code SSR.

[0116] The data completion process includes: converting latitude and longitude to projected coordinates when they exist but are missing; performing reverse conversion when they exist but are missing; and calculating ground speed and heading angle when they exist but are missing velocity components. , Calculation when velocity components exist but ground speed and heading angle are missing. And calculate the heading angle in reverse.

[0117] Step 2: As Figure 5 As shown, multi-priority track association is performed; the input of this step is the single track observation data completed in step 1 and the current valid track set maintained internally by the system; each valid track in the current valid track set corresponds to an internal track batch number, the previous time-lapse filtering status, the historical track queue, and the outlier detector.

[0118] The association process determines which existing target a single observation belongs to. When association is successful, the internal track batch number and track context of the matching track are output, and subsequent IMM filtering only updates the state corresponding to that target. When association fails, a new internal track batch number and track context are created for the observation, and it is added to the set as a new valid track.

[0119] Multi-priority track association is performed in the following order:

[0120] Priority 1: Exact 24-bit address code matching. If the currently observed 24-bit address code matches the 24-bit address code of a valid track, it is preferentially determined to be the same target.

[0121] Priority 2: Exact flight number matching. If address code matching fails, it is determined whether the currently observed flight number matches the flight number of a valid track; if they match, they are determined to be the same target.

[0122] If the address code fails to match, then based on the flight number being the same, further analysis is conducted using time interval, predicted position threshold, and speed continuity to determine whether they belong to the same target.

[0123] Priority 3: Comprehensive Match. If neither Priority 1 nor Priority 2 matches, a comprehensive match score is calculated based on position, secondary code, velocity, and heading angle. The comprehensive score formula is:

[0124]

[0125] in, For comprehensive matching score; The score is for a secondary code match, with 0 for a match and 1 for a non-match. The position is scored using Euclidean distance, calculated based on the Euclidean distance between the current observation and the candidate valid tracks. With position threshold The proportion calculation, that is:

[0126]

[0127] Scoring is based on the speed difference, according to the speed difference. Relative to the speed threshold The proportion calculation, that is:

[0128]

[0129] To score the heading angle difference, the minimum heading angle difference is used. Relative to heading threshold The proportion calculation, that is:

[0130]

[0131] , , , These are the weighting coefficients for the secondary code, position, speed, and heading. The system selects the candidate track with the lowest overall matching score as the best candidate track. When the corresponding difference exceeds a threshold, the score for that item can be set as a preset penalty value, or the candidate track can be determined to not meet the matching condition for that item.

[0132] If the best overall match score is less than the dynamic acceptance threshold, the two are considered to belong to the same target; otherwise, the current observation is considered to correspond to a new target. The dynamic acceptance threshold can be adjusted based on secondary code consistency, track stability, the number of most recent successful associations, and track density.

[0133] Step 3: Perform IMM filtering on the successfully associated tracks. This step takes the previous intermediate state of the matched track, the historical track queue, the current observation vector, the observation noise matrix, and the model configuration as input, and outputs the posterior state, posterior covariance, model probability, fusion state, and outlier detection results for each motion model at the current time. Specifically, it includes:

[0134] Step 3-1: Calculate the sampling time interval T between the current observation and the previous valid observation; if T is negative, or the time interval of the same source is less than the minimum valid interval, then discard the observation.

[0135] Simultaneously, an observation vector z and an observation noise matrix R are constructed, and multiple motion models are pre-configured, including a uniform velocity model (CV), a uniform acceleration model (CA), a left-turn model (CTL), and a right-turn model (CTR). The CV, CTL, and CTR models can employ a four-dimensional state vector [x, , y, ] T This indicates that CA uses a state vector [x, , , y, , ] T This indicates that x and y are projected coordinates. and For velocity components, and This represents the acceleration component.

[0136] Since different motion models may have different state dimensions, before IMM interaction, model mixing, and final fusion, the states of each model are mapped to a unified fusion state space. When expanding from four-dimensional to six-dimensional states, the acceleration components can be set to zero or initialized according to the historical rate of change of velocity; when compressing from six-dimensional to four-dimensional states, the acceleration components can be discarded or marginalized.

[0137]

[0138]

[0139] in, Four-dimensional state To the six-dimensional state The extended mapping matrix, This is a compression mapping matrix from six-dimensional states to four-dimensional states. Through this mapping mechanism, various motion models can complete probability-weighted fusion in a unified state space.

[0140] For CTL and CTR turning models, the turning angular velocity is dynamically calculated based on the current rate of change of heading angle. When the rate of change of heading angle is missing but the roll angle and ground speed are available, the turning angular velocity is estimated based on the coordinated turning relationship, and the state transition matrix F and process noise matrix Q of the turning model are updated accordingly.

[0141] Step 3-2: Use DS evidence theory to estimate the current dominant motion model type and correct the model probability of the previous time step before IMM interaction.

[0142] This step is used to provide auxiliary judgment on the probability of motion models by utilizing motion features that are relatively independent of the observation residuals when the current observation may contain outliers. Instantaneous evidence is calculated based on the roll angle, rate of change of heading angle, and ground speed information between the current track and the track at the previous time step. When the rate of change of heading angle is available, it is directly used as the turning rate feature; when the rate of change of heading angle is missing but the roll angle and ground speed are available, the turning rate is estimated based on the coordinated turning relationship. Historical evidence is calculated based on the historical track queue of the same internal track batch to determine the mean acceleration and mean turning rate.

[0143] After merging the two, the basic probability assignments for CV, CA, CTL, CTR, and UNKNOWN (unidentified model) are obtained. The model with the highest trust level is selected as the current dominant motion model. If the maximum trust level is lower than the decision threshold, UNKNOWN is kept to avoid misjudgment.

[0144] The acceleration and turning rate features are mapped to model confidence using the Gaussian membership function, which is:

[0145]

[0146] Where x is the input motion feature value, This represents the feature mean of the corresponding motion model. The feature standard deviation corresponding to the motion model, This represents the membership or confidence level of the motion feature to the corresponding motion model.

[0147] Motion characteristics may include acceleration characteristics and turning rate characteristics The acceleration characteristics are mainly calculated based on the speed changes between historical track points, while the turning rate characteristics can be calculated based on the current rate of change of the heading angle. When the rate of change of the heading angle is missing but the roll angle and ground speed are available, the turning rate can be estimated based on the coordinated turning relationship. and turning rate characteristics The calculation formula is:

[0148]

[0149]

[0150] in, Let the velocity vector be the velocity vector at the current moment. The velocity vector at the previous moment. The heading angle at the current moment. The heading angle at the previous moment. The time interval between two consecutive observations;

[0151] Acceleration features are used to distinguish between CV and CA models, and turning rate features are used to distinguish between straight-line motion models and CTL / CTR turning models. Historical evidence is calculated based on historical track queues within the same internal track batch, with the historical window including at least three track points. Velocity and heading changes are calculated for adjacent track points to obtain average acceleration and average turning rate, which are then mapped to historical confidence levels for CV, CA, CTL, and CTR using configured Gaussian membership parameters. The Dempster-Shafer combination rule is used to fuse instantaneous and historical evidence; the fusion formula is as follows:

[0152]

[0153] in, The basic probability allocation for transient evidence. The basic probability assignments corresponding to historical evidence are defined by A, B, and C, which are elements in the set of motion model propositions. K is the conflict factor, used to characterize the degree of conflict between instantaneous evidence and historical evidence. The set of motion model propositions includes CV, CA, CTL, CTR, and UNKNOWN.

[0154] The system sets an unknown quality allocation ratio to avoid completely deterministic judgments. The UNKNOWN quality is used to express insufficient or conflicting evidence. When the maximum confidence level after fusion does not reach the decision threshold, the system does not force a switch to a specific motion model, but instead outputs UNKNOWN.

[0155] In a preferred embodiment, the DS fusion result is used as a model probability correction factor before IMM interaction to correct the model probability at the previous time step.

[0156]

[0157]

[0158] in, The model probabilities before DS correction; The unnormalized model probabilities are adjusted for DS support. λ represents the support of the DS evidence theory for the j-th motion model; λ is the evidence correction strength coefficient. This is the corrected and normalized model probability before the interaction. When the UNKNOWN quality is high or the conflict factor K exceeds a preset threshold, λ can be reduced or this correction can be omitted.

[0159] Step 3-3: Perform IMM interactive mixing;

[0160] Based on the model transition probability matrix and the model probabilities obtained in step 3-2, calculate the predicted model probability and mixture probability of each target motion model. Let the model probability of the i-th motion model at the previous time step after correction in step 3-2 be... The transition probability of model i to model j is Then the prediction model probability of the j-th motion model at the current time is:

[0161]

[0162] Given the current target model is j, the mixture probability of model i at the previous time step is:

[0163]

[0164] Based on the mixing probability, the posterior states and posterior covariances of each motion model at the previous time step are weighted and mixed to obtain the mixed initial state and mixed initial covariance of the j-th motion model:

[0165]

[0166]

[0167] in, Let be the predicted model probability of the j-th motion model. Let be the mixture probability of model i at the previous time step, given that the current model is j. and Let be the mixed initial state and mixed initial covariance of the j-th motion model, respectively.

[0168] Through the aforementioned interactive mixing, each motion model can absorb state information from other models before prediction, thereby avoiding filtering discontinuities caused by abrupt changes in a single model. All four motion models—CV, CA, CTL, and CTR—participate in the interactive mixing, and subsequent steps perform prior predictions based on their respective state transition matrices and process noise matrices.

[0169] Steps 3-4: As Figure 4 As shown, the prior predictions of each model are calculated and the prior predictions of each motion model are fused to obtain the outlier detection benchmark.

[0170] For the j-th motion model, its prior state estimate and prior covariance prediction are:

[0171]

[0172]

[0173] in, For the prior state estimate of the j-th motion model at the current time k, Let j be the initial state of the mixed motion model. Let J be the state transition matrix of the j-th motion model. Let be the prior state covariance of the j-th motion model. Let J be the initial covariance of the j-th motion model. Let J be the process noise covariance of the j-th motion model;

[0174] Since different motion models may have different state dimensions, we first map the prior states of each motion model to a unified fusion state space. Let the prediction probability of the j-th motion model at the current time be... The prior state of the j-th motion model after mapping to the unified fused state space is: If the prior covariance is P_j(k|k-1), then the fused prior state is:

[0175]

[0176] The fusion of prior covariance is:

[0177]

[0178] in, This represents the prior state after fusion. The prior covariance after fusion Let be the predicted model probability of the j-th motion model at the current time. This fusion prior is not used as the final filtered output, but rather as the prediction benchmark for outlier detection in subsequent position and velocity observations.

[0179] Steps 3-5: Multi-level adaptive outlier detection based on fusion of prior execution position and velocity observations.

[0180] The inputs for this step are the observation type (type), the corresponding innovation vector (e), the innovation covariance matrix (S), and the fused prior covariance matrix. and the observation noise covariance matrix The output includes the detection level (level), Mahalanobis distance (D), and adaptive detection threshold. Confidence level And a weighting factor w. Historical Mahalanobis distance queues are maintained for both position and velocity observations, thus allowing for different adaptive detection thresholds for different observation types.

[0181] For location observations, extract from the current observation vector And based on the fusion of prior states Constructing the position observation matrix Calculate the position information and information covariance; for velocity observations, if the current observation vector contains velocity components, then extract... And based on the fusion of prior states Constructing the velocity observation matrix Calculate the velocity innovation and innovation covariance. If the current observation does not contain a velocity component, only position observation outlier detection is performed.

[0182] The new interest is calculated as follows:

[0183]

[0184] The new information covariance is:

[0185]

[0186] Calculate the Mahalanobis distance of the new information:

[0187]

[0188] in, For the innovation vector, The new information covariance matrix;

[0189] In a preferred embodiment, it is based at least on the fusion of prior covariance matrices. The uncertainty of the calculation system can also be further combined with the observation noise covariance matrix. :

[0190]

[0191]

[0192] in, As an indicator of system uncertainty, Based on confidence level, The uncertainty scaling coefficient is used to adjust the influence of the fused prior covariance matrix and the observation noise covariance matrix on the system uncertainty index u. This is the uncertainty scaling factor. Let be the state covariance matrix. To observe the noise covariance matrix, tr(·) represents the trace of the matrix, and clip(·) represents limiting the confidence level to a preset minimum value. and maximum value The higher the system uncertainty, the lower the confidence level, and the more appropriately the detection threshold should be relaxed to reduce the probability of misclassifying normal maneuvers as outliers. The basic detection threshold is determined based on adaptive confidence. The threshold is adjusted based on historical Mahalanobis distance statistics. When the current Mahalanobis distance D exceeds the sum of the historical mean and three times the standard deviation, the detection threshold is adjusted as follows:

[0193]

[0194] When the historical standard deviation is less than 0.1 times the historical mean, the detection threshold will be reduced by 20%.

[0195]

[0196] Ultimately based on Mahalanobis distance The relationship with the threshold enables three-level decision-making: accept, downweight, and reject.

[0197]

[0198] when When D is in the middle, output DOWNWEIGHT and calculate the weighting w based on the position of D between the acceptance threshold, buffer threshold and rejection threshold.

[0199] Steps 3-6: Perform intelligent observation fusion based on the detection results and execute Kalman update.

[0200] Based on the position and velocity observation results obtained in steps 3-5, the observation components participating in the Kalman update at the current moment and their weights are dynamically determined, and an enhanced observation vector is constructed. Enhanced observation matrix and enhance the observation noise matrix .

[0201] Specifically, when the location observation detection result is REJECT, it is considered that the current location observation is unavailable and the location observation will not be used this time;

[0202] In a preferred embodiment, when a position observation is rejected, the current observation update is skipped, and the prior prediction results of each motion model are directly retained to avoid abrupt changes in track state caused by abnormal position observations.

[0203] When the position observation result is ACCEPT and the velocity observation result is REJECT, or when the current observation does not contain a velocity component, only the position observation is used to construct the augmented observation:

[0204]

[0205]

[0206]

[0207] When the position observation detection result is ACCEPT and the velocity observation detection result is ACCEPT, the position observation and the velocity observation are used jointly to construct an enhanced observation:

[0208]

[0209]

[0210]

[0211] wherein, is the enhanced observation matrix that extracts both the position component and the velocity component, is the enhanced observation noise matrix obtained by combining the position observation noise and the velocity observation noise.

[0212] When the position observation is DOWNWEIGHT and the velocity observation is available, the position observation is still used, and its observation noise is adjusted according to the position down-weighting weight; when the position observation is DOWNWEIGHT and the velocity observation is REJECT or does not exist, only the down-weighted position observation is used for updating; when the velocity observation is DOWNWEIGHT, the velocity observation is still used, and its observation noise is adjusted according to the velocity down-weighting weight. Let the down-weighting weight of the corresponding observation component be w (0<w≤1), then the down-weighting can be achieved by amplifying the corresponding observation noise:

[0213]

[0214] wherein, is the down-weighted observation noise matrix, is the original observation noise matrix.

[0215] Preferably, the enhanced observation noise matrix is constructed by amplifying the observation noise matrix to maintain the unified form of Kalman update.

[0216] For the j-th motion model, the system is based on the priori state obtained in steps 3-4 , the priori covariance , and the enhanced observation vector constructed in this step , the enhanced observation matrix and the enhanced observation noise matrix , to calculate the innovation:

[0217]

[0218] The innovation covariance is:

[0219]

[0220] The Kalman gain is:

[0221]

[0222] The updated model state estimate is:

[0223]

[0224] The updated model covariance is:

[0225]

[0226] in, Let J be the Kalman gain of the j-th motion model. It is the identity matrix. and Let be the posterior state estimate and posterior covariance of the j-th motion model at the current time.

[0227] If no valid observation components exist in this step, the Kalman update is not performed, and the prior state and prior covariance of the j-th motion model are used as the current state output of the model, i.e.:

[0228] =

[0229] =

[0230] Steps 3-7: Update the motion model probabilities and fuse the posterior states of each motion model.

[0231] Based on the innovation vector and innovation covariance obtained from the Kalman update process of each motion model in steps 3-6, calculate the likelihood function of each motion model under the current observation. For the j-th motion model, its likelihood function is:

[0232]

[0233] in, Let be the likelihood function of the j-th motion model at the current time k. For the j-th motion model, the innovation vector is obtained based on enhanced observations. The corresponding new information covariance matrix, This means that the mean is the zero vector. The multivariate Gaussian probability density function of the covariance matrix in The value at that location.

[0234] Based on the likelihood functions of each motion model and the predicted model probabilities obtained in step 3-3, update the model probabilities at the current time step:

[0235]

[0236] in, Let be the predicted model probability of the j-th motion model at the current time. Let be the posterior model probability of the j-th motion model at the current time, with the normalization factor in the denominator.

[0237] Subsequently, the posterior state estimates of each motion model are weighted and fused according to the posterior model probability to obtain the fused posterior state at the current moment:

[0238]

[0239] The post-fusion covariance is:

[0240]

[0241] in, and Let be the posterior state estimate and posterior covariance of the j-th motion model at the current time, respectively. and These are the posterior state estimate and posterior covariance after fusion at the current time, respectively.

[0242] Steps 3-8: Write the posterior state, posterior covariance, posterior model probability, fused posterior state, fused posterior covariance, and outlier detection results obtained in Step 3-5 into the intermediate state of the current track, and convert the fused posterior state into the filtered track output.

[0243] When a valid augmentation observation exists, the model probability is updated based on the likelihood function. When no valid observation component exists, the model probability update based on the observation likelihood is skipped, and the prior predictions of each model and the predicted model probability are used as the posterior results at the current time step in the output.

[0244] Corresponding to the above method, the present invention also provides a system for implementing the IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection, comprising:

[0245] Track association module: Enables multi-priority track association;

[0246] IMM filtering module: performs multi-model interaction and fusion;

[0247] DS Evidence Theory Module: Estimating the type of the current dominant motion model;

[0248] Outlier detection module: Implements multi-level adaptive outlier detection;

[0249] Intelligent observation fusion module: dynamically selects observation type and weight based on detection results;

[0250] Data completion module: Automatically completes missing track data.

[0251] Example 1:

[0252] The specific implementation process of the above solution will be illustrated with a specific embodiment.

[0253] Part 1: System Initialization and Track Processing Procedure. This part describes the system initialization and a complete track processing procedure.

[0254] First, initialize the IMM filter configuration. Create an IMMConfig object, set the model transition probability to 0.95, the process noise to 0.25, and define the relevant parameters for the model, such as constant velocity (CV), constant acceleration (CA), left turn (CTL), and right turn (CTR). For example, set the left turn rate and right turn rate to -0.25° / s and 0.25° / s, respectively, as well as the standard deviations of position and velocity in each direction.

[0255] Next, receive raw track data, such as observation data containing longitude 116.4, latitude 39.9, speed 250.0 m / s, heading angle 45°, secondary code "A1234" and 24-bit address code "7AB456".

[0256] Furthermore, in a set of example inputs, the flight number is B001P, the secondary code is 0237, the 24-bit address code is 7805b8, the receiving timestamp is 1759048752603, the longitude is 116.256°, the latitude is 27.7384°, the ground speed is 129.897 m / s, the altitude is 5090.46 m, the heading angle is 226.671°, the roll angle is -0.015114°, and the heading angle change rate is 0° / s. The system first converts the latitude and longitude into projected coordinates x and y; simultaneously, it calculates the velocity components based on the ground speed and heading angle. =129.897*sin(226.671°), =129.897*cos(226.671°), thus forming the state observation input used uniformly for subsequent correlation, filtering, and outlier detection.

[0257] Part Two: Motion Model Estimation Driven by DS Evidence Theory.

[0258] like Figure 2 As shown, the execution steps of the DS evidence theory module are as follows:

[0259] 1. Extracting instantaneous evidence: Calculate acceleration features based on the current track (currentTrack) and the previous track (previousTrack). and turning rate characteristics :

[0260]

[0261]

[0262] Then, the confidence score for each model is calculated using the Gaussian membership function. For example, the confidence score for the uniform velocity (CV) model is:

[0263] The Gaussian function is in the form of .

[0264] For example, the average acceleration calculated within the historical window is 0.6 m / s². 2 When the average turning speed is 0.1° / s, the CV model corresponds to an acceleration mean of 0 and an acceleration standard deviation of 10. The linear motion turning speed mean is 0 and the turning speed standard deviation is 2. Therefore, the historical confidence level of the CV model can be obtained by multiplying the two Gaussian membership degrees. If the average acceleration increases to 3.0 m / s² within the same window... 2 If the average turning rate is still close to 0, the confidence level of the CA model will increase accordingly; if the average turning rate is close to the preset left or right turn rate, the confidence level of the CTL or CTR model will increase.

[0265] 2. Extract historical evidence: Calculate the average acceleration and average turning rate from the history queue and calculate the confidence level of each model based on these statistical characteristics.

[0266] 3. Evidence Fusion and Decision Making: The Dempster-Shafer combination rule is used to fuse the instantaneous evidence m1 and historical evidence m2. The fusion formula is as follows:

[0267]

[0268] in, The conflict factor is used to obtain the basic probability assignment (BPA) after fusion. Then, the model with the highest confidence is selected as the dominant motion model (dominantType) at the current time step. This result will be used to guide subsequent filter updates.

[0269] Part 3: Multi-level Adaptive Wild Value Detection.

[0270] After completing the IMM interactive mixing and prior prediction of each model, the system first maps the prior states of each model to a unified fusion state space to obtain the fusion prior states. and fusion of prior covariance This fusion prior is only used as a prediction benchmark for outlier detection in position and velocity observations, and is not used as the final filtered output.

[0271] like Figure 3 As shown, the execution steps of the outlier detection module are as follows:

[0272] 1. Calculate the Mahalanobis distance: Using the innovation vector e and the innovation covariance matrix S, calculate the Mahalanobis distance:

[0273]

[0274] 2. Calculate adaptive confidence: Dynamically adjust the confidence based on the traces of the state covariance matrix P and the observation noise covariance matrix R.

[0275]

[0276] The higher the system uncertainty, the lower the confidence level, thus allowing for a relaxation of the detection threshold.

[0277] 3. Adjust the detection threshold: Dynamically adjust the base threshold based on historical Mahalanobis distance statistics (mean, standard deviation). For example, if the current Mahalanobis distance is significantly greater than the historical mean, increase the threshold to avoid misclassifying normal large maneuvers as outliers.

[0278]

[0279] 4. Three-level decision: Based on the relationship between Mahalanobis distance and the adjusted threshold, output ACCEPT, DOWNWEIGHT, or REJECT results:

[0280] In a numerical example, the information for the location observation is e = [80, -40]. T m, where the new information covariance is Then D=(80 2 +40 2 ) / 2500=3.2; If the new information observed at another location is e=[180,120] T m, then D=(180 2 +120 2 ) / 2500=18.72.

[0281] The system calculates system uncertainty based on the fusion of prior covariance and observation noise. And calculate adaptive confidence. When system uncertainty increases, By lowering the threshold, the detection threshold is correspondingly relaxed, thereby reducing the probability of misjudging normal maneuvers as outliers.

[0282] The system is based on to obtain the basic detection threshold , and combines the historical mean of Mahalanobis distances and the standard deviation to adjust the threshold. For example, when the current Mahalanobis distance exceeds +3 , the basic threshold can be adjusted to = ; when the historical standard deviation is less than 0.1 times the historical mean, the basic threshold can be adjusted to = . The adjusted threshold is used to synchronously generate the acceptance threshold and the rejection threshold .

[0283] The three-level decision rule is: when D≤ , output ACCEPT; when <D≤ , output DOWNWEIGHT; when D> , output REJECT. Taking the aforementioned values as an example, if =3.0, =7.0, then D=3.2 corresponds to DOWNWEIGHT, and D=18.72 corresponds to REJECT.

[0284] Part IV. Intelligent Observation Fusion and Kalman Update.

[0285] 1. According to the detection results of the position observation and the velocity observation, the system dynamically constructs the enhanced observation vector , the enhanced observation matrix and the enhanced observation noise matrix . When the position observation is ACCEPT and the velocity observation is REJECT, or the current observation does not contain a velocity component, only the position observation is used ; when both the position observation and the velocity observation are ACCEPT, they are used jointly .

[0286] When a certain observation component is DOWNWEIGHT, the system can still retain this observation component, but amplifies the corresponding observation noise according to its down-weighting weight w, that is , where 0<w≤1. This manner is consistent with the aforementioned Kalman update form, no additional scaling of the corresponding innovation is performed, and repeated down-weighting is avoided.

[0287] 2. For example, when the position observation is ACCEPT with a weight of 1.0 and the velocity observation is DOWNWEIGHT with a weight of 0.4, the enhanced observation vector still adopts ; the position noise remains unchanged, and the velocity observation noise block is amplified by 1 / 0.4 and then written into This allows velocity observations to continue participating in the update, but with reduced impact. If the position observation is REJECT, the current observation update can be skipped, and the prior predictions of each motion model will be retained to avoid abrupt track changes caused by abnormal positions.

[0288] For the j-th motion model, the system is based on enhanced observation calculations. , , and update , .

[0289] Part 5: Multi-priority track association.

[0290] 3. The track association module first performs a precise match of the 24-bit address code; if no match is found, it performs a precise match of the flight number, and combines the time interval, predicted position threshold and speed continuity to determine whether they belong to the same target; if no match is found again, it performs a comprehensive match.

[0291] The overall matching score adopts a normalized scoring method: .in The value is 0 when there is a secondary code match and 1 when there is no match. , , System selection The lowest-scoring track is selected as the best candidate track.

[0292] For example, let =0.4、 =0.3、 =0.2、 =0.1, location threshold =300 m, velocity threshold =50 m / s, heading threshold =40°. If a candidate track's secondary code matches, and =120 m =15 m / s =8°, then =0、 =0.40、 =0.30、 =0.20, Overall Score =0.20. If the dynamic acceptance threshold is 0.60, and is relaxed to 0.90 due to secondary code matching, then 0.20 < 0.90, and the association is considered successful.

[0293] If the secondary code of the other candidate track does not match, and =280 m =45 m / s =32°, then =1、 =0.93、 =0.90、 =0.80, with an overall score of approximately 0.94; at the unrelaxed acceptance threshold of 0.60, 0.94>0.60, the system does not perform correlation and does not create a new track for the observation.

[0294] Part VI. Model Probability Update and Output.

[0295] When there are effective augmented observations, the system calculates the likelihood function of each motion model. And update the posterior model probability based on the prediction model probability: When no valid observation components exist, the system skips the model probability update based on observation likelihood and uses the prior predictions of each model and the predicted model probability as the posterior result at the current time step in the output.

[0296] Finally, the system fuses the posterior states of each model according to the posterior model probability to obtain... And calculate the post-fusion covariance. The output filtered track includes at least the filtered position, filtered velocity, heading angle, internal track batch number, dominant motion model type, model probability distribution, and outlier detection level. In specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium can store a computer program, which, when executed by the data processing unit, can run the invention content and some or all of the steps in various embodiments of the IMM track filtering method and system based on DS evidence theory and multi-level adaptive outlier detection provided by this invention. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0297] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0298] This invention provides an IMM track filtering method and system based on DS evidence theory and multi-level adaptive outlier detection. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. An IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection, characterized in that, Includes the following steps: Step 1: Receive track observation data and perform data integrity checks and completion. Step 2: Perform multi-priority track association on the track observation data; Step 3: For the successfully associated track observation data, a multi-motion model matching scheme is used for filtering to obtain the final IMM track.

2. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 1, characterized in that, The track observation data mentioned in step 1 includes: Receives time, target location, velocity information, and target identification information; among which, Target location includes longitude, latitude, or projected coordinates x and y; velocity information includes ground speed. Heading angle Or velocity component , The target identification information includes the flight number, 24-digit address code, and secondary code SSR.

3. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 2, characterized in that, Step 2, which involves performing multi-priority track association on track observation data, includes: Step 2-1: Perform track association with priority 1, that is, perform precise matching using 24-bit address codes. If no match is found, proceed to step 2-2. Step 2-2: Perform track association with priority 2, that is, match precisely by flight number. If no match is found, proceed to step 2-3. Steps 2-3 are executed, involving track association at priority level 3, which involves calculating a weighted score based on position, secondary code, heading, and speed. The matching process is as follows: Step 2-3-1, calculate the weighted score, as follows: ； in, For secondary code matching score, The score is the Euclidean distance of the location. Score based on speed difference. Score for heading angle difference. , , and These are the weighting coefficients; Step 2-3-2: Select the track with the lowest weighted score as the best candidate track; Step 2-3-3: If the weighted score of the best candidate track is less than the dynamic acceptance threshold, the trimming is successful; otherwise, the current observation is considered to correspond to a new target.

4. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 3, characterized in that, Step 3, which describes using a multi-motion model matching scheme for IMM filtering, includes: Step 3-1 involves preprocessing the successfully correlated track observation data and pre-setting different motion models for IMM filtering; the details are as follows: Calculate the sampling time interval T between the currently associated successfully associated track observation data and the valid observation data of the previous moment. If T is negative or the time interval of the same source is less than the minimum valid interval, discard the observation; otherwise, construct the observation vector z and the observation noise matrix R at the same time, and pre-configure the motion model, which includes the uniform velocity model CV, the uniform acceleration model CA, the left turn model CTL, and the right turn model CTR. Step 3-2: Use DS evidence theory to estimate the currently dominant motion model and correct the motion model probability of the previous time step; Step 3-3: Perform IMM interactive mixing based on the motion model probabilities of the previous time step to obtain the initial mixing state and initial mixing covariance of each motion model. Steps 3-4: Calculate the prior predictions of each model based on the mixed initial state and mixed initial covariance of each motion model, and fuse the prior predictions of each motion model to obtain the outlier detection benchmark. Steps 3-5: Perform outlier detection based on the outlier detection benchmark; Steps 3-6: Based on the outlier detection results, perform intelligent observation fusion and execute Kalman updates for each motion model to obtain the innovation vector and innovation covariance; Steps 3-7: Update the motion model probabilities based on the innovation vector and innovation covariance, and fuse the posterior states of each motion model; Steps 3-8: The fusion results of the posterior states of each motion model and the outlier detection results of Step 3-5 are written into the intermediate state of the current track, and the fused posterior states are converted into the filtered track output.

5. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 4, characterized in that, Step 3-3, which involves performing IMM interactive mixing based on the motion model probability from the previous moment, includes: Let the model probability of the i-th motion model at the previous time step after correction in step 3-2 be... The transition probability of model i to model j is Then the j-th motion model at the current time Predictive model probability for: ； Given the current target model j, what is the mixture probability of model i at the previous time step? for: ； Based on the mixing probability, the posterior states and posterior covariances of each motion model at the previous time step are weighted and mixed to obtain the mixed initial state of the j-th motion model. and mixed initial covariance The following are examples: ； ； in, Let be the predicted model probability of the j-th motion model. Let be the mixture probability of model i at the previous time step, given that the current model is j. and Let be the mixed initial state and mixed initial covariance of the j-th motion model, respectively.

6. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 5, characterized in that, Steps 3-4 describe calculating the prior predictions of each motion model based on the mixed initial state and mixed initial covariance, and fusing the prior predictions of each motion model to obtain the outlier detection benchmark, including: Step 3-4-1, for the j-th motion model, its prior state estimation and prior covariance prediction They are respectively: ； ； in, For the prior state estimate of the j-th motion model at the current time k, Let j be the initial state of the mixed motion model. Let J be the state transition matrix of the j-th motion model. Let be the prior state covariance of the j-th motion model. Let J be the initial covariance of the j-th motion model. Let J be the process noise covariance of the j-th motion model; Step 3-4-2: Map the prior states of motion models with different state dimensions to a unified fusion state space; let the predicted model probability of the j-th motion model at the current time be... The prior state of the j-th motion model after mapping to the unified fused state space is: If the prior covariance is P_j(k|k-1), then the fused prior states... for: ； Fusion of prior covariance for: ； in, This represents the prior state after fusion. The prior covariance after fusion Let be the predicted model probability of the j-th motion model at the current time. The fusion of prior states and variance serves as the benchmark for outlier detection.

7. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 6, characterized in that, The fusion of prior predictions from various motion models described in steps 3-5 includes: Extract position observations from the current observation vector. And based on the fusion of prior states Constructing the position observation matrix Calculate the location information and the location information covariance; If the current observation vector contains a velocity component, then extract the velocity observation. And based on the fusion of prior states Constructing the velocity observation matrix Calculate the velocity information and velocity information covariance; Calculate the Mahalanobis distance of the new information based on the new information and its corresponding covariance. : ； in, For the innovation vector, The new information covariance matrix; Based on the fusion of prior covariance matrix Computational system uncertainty and confidence level The details are as follows: ； ； in, Based on confidence level, The uncertainty scaling coefficient. This is the uncertainty scaling factor. To observe the noise covariance matrix, tr(·) represents the trace of the matrix, and clip(·) represents limiting the confidence level to a preset minimum value. and maximum value between; Based on confidence level Determine the basic detection threshold The acceptance threshold is obtained by adjusting the data based on historical Mahalanobis distance statistics. and rejection threshold ; According to Mahalanobis distance The relationship with the threshold is used to make a three-level decision of acceptance, deweighting, and rejection, as shown below: ； in, This indicates that the result is accepted. This indicates that the outlier detection result is "reject". This indicates that the outlier detection result is reduced in weight, and the reduction weight w is calculated based on the position of the Mahalanobis distance D of the new information between the acceptance threshold, buffer threshold, and rejection threshold.

8. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 7, characterized in that, Steps 3-6, which describe intelligent observation fusion based on outlier detection results and the execution of Kalman updates for each motion model, include: When the location observation detection result is REJECT, it is considered that the current location observation is unavailable, the location observation will not be used this time, the current observation update will be skipped, and the prior prediction results of each motion model will be directly retained. When the position observation results in ACCEPT and the velocity observation results in REJECT, or when the current observation does not contain a velocity component, only the position observation is used to construct the enhanced observation vector. Enhanced observation matrix and enhance the observation noise matrix ; When both position and velocity observations are accepted for ACCEPT, an enhanced observation vector is constructed by combining position and velocity observations. Enhanced observation matrix and enhance the observation noise matrix ; When the position observation is downweighted and the velocity observation is available, the position observation is still used, and its observation noise is adjusted according to the position downweighting weight; when the position observation is downweighted and the velocity observation is rejected or does not exist, only the downweighted position observation is used for updating; when the velocity observation is downweighted, the velocity observation is still used, and its observation noise is adjusted according to the velocity downweighting weight. For the j-th motion model, the system is based on the prior state obtained in steps 3-4. Prior covariance and the constructed enhanced observation vector Enhanced observation matrix and enhance the observation noise matrix Calculate new information New information covariance Kalman gain Updated model state estimation and model covariance They are represented as follows: ； ； ； ； ； in, Let J be the Kalman gain of the j-th motion model. It is the identity matrix. and These are the posterior state estimate and posterior covariance of the j-th motion model at the current time, respectively. If a valid observation component exists, the Kalman update is not performed, and the prior state and prior covariance of the j-th motion model are used as the output state of the model at the current time.

9. The IMM track filtering method based on DS evidence theory and multi-level adaptive outlier detection according to claim 8, characterized in that, Steps 3-7, which describe updating the motion model probabilities based on the innovation vector and innovation covariance and fusing the posterior states of each motion model, include: For the j-th motion model, calculate its likelihood function: ； in, Let be the likelihood function of the j-th motion model at the current time k. For the j-th motion model, the innovation vector is obtained based on enhanced observations. The corresponding new information covariance matrix, This means that the mean is the zero vector. The multivariate Gaussian probability density function of the covariance matrix in The value at; Update the model probability at the current time step based on the likelihood function of each motion model and the predicted model probability in step 3-3: ； in, Let be the predicted model probability of the j-th motion model at the current time. Let be the posterior model probability of the j-th motion model at the current time. The posterior state estimates of each motion model are weighted and fused according to the posterior model probability to obtain the fused posterior state at the current time. , means as follows: ； The posterior covariance after fusion is calculated and expressed as follows: ； in, and Let be the posterior state estimate and posterior covariance of the j-th motion model at the current time, respectively. and These are the posterior state estimate and posterior covariance after fusion at the current time, respectively.

10. An IMM track filtering system based on DS evidence theory and multi-level adaptive outlier detection, used to implement the method described in any one of claims 1-9, characterized in that, include: The track association module performs multi-priority track association. The IMM filtering module performs multi-model interaction and fusion. The DS evidence theory module estimates the type of the current dominant motion model; The outlier detection module performs multi-level adaptive outlier detection. The intelligent observation fusion module dynamically selects the observation type and weight based on the outlier detection results. The data completion module automatically completes missing track data.

Images