High-precision continuous tracking method for strong-maneuvering infrared weak and small target under space-based detection visual angle

Abstract

The invention discloses a high-precision continuous tracking method for a strong-maneuvering infrared weak and small target under a space-based detection visual angle, and the method comprises the steps: obtaining the uncertainty measurement of a target at a non-maximum suppression stage of a detection result, and enabling the uncertainty measurement to act on a target state updating and data association stage; therefore, the experience distribution of the detector is fully transmitted to the tracking process, and the tracking accuracy is improved. In order to cope with a complex target maneuvering state, an interactive multi-model technical route is adopted to design a tracking algorithm; in order to reduce the dependence on a prior motion model, a dynamic Markov transfer matrix construction method is designed, and the model transfer probability is updated in a mode of comprehensively modulating historical dynamic information and current static information; in a data association stage, targets with different scales are associated by integrating advantages of IoU and NWD, uncertainty is transmitted to a cost calculation process, and tracks and targets which are indefinitely matched are processed based on scale invariant and energy invariant hypotheses, so that high-precision continuous tracking of the targets is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared target detection and tracking, and relates to an infrared target tracking method, and in particular to a high-precision continuous tracking method for a strong maneuvering infrared dim small target under the perspective of space-based detection. Background Art

[0002] With the continuous development of new types of moving targets such as drones and aircraft, real-time tracking of targets to form continuous trajectories has become a major demand. Space-based detection, with its wide field of view and high timeliness, has become the main development direction for the detection of highly maneuverable targets in the future. Targets from the perspective of space-based detection are usually sparse, with most areas occupied by scene clutter. At the same time, the color and texture characteristics of the targets are similar to cloud clutter and scene structure clutter, making visible light imaging unsuitable for target tracking tasks. Infrared has stronger night imaging capabilities than visible light, so using space-based platforms to track infrared targets is a feasible technical means.

[0003] At present, the infrared small target tracking task under the perspective of space-based detection still has the following technical challenges: (1) Uncertainty of detection results: Due to the complexity of space-based infrared scenes, the energy of small targets is weak and the morphological texture features are missing, so there is a deviation between the detection results and the true value. It is difficult for the tracking algorithm to eliminate this error with the detection results as input. (2) Trajectory crosstalk: The spatial resolution under the perspective of space-based detection is low, and the target and clutter trajectories are coupled, making it difficult to effectively distinguish the target trajectory. (3) Difficulty in association: Infrared small targets only occupy a few to dozens of pixels in remote sensing images, and traditional IoU matching methods are not suitable for extremely small-scale targets. (4) Difficulty in continuous tracking: The motion state of infrared targets such as aircraft is controlled by humans. When encountering special circumstances such as ground control instructions, they show strong maneuverability and unpredictable characteristics, resulting in broken tracking trajectories.

[0004] Currently, infrared small target tracking methods from the perspective of space-based detection are primarily divided into model-driven and data-driven approaches. Model-driven approaches use detection results as input and combine data association algorithms to label targets with the same identity. Tracking algorithm performance is highly dependent on detector accuracy, prone to association errors in scenarios with target occlusion or low signal-to-noise ratio. The computational complexity increases exponentially, and the algorithm is not adaptable to nonlinear / non-Gaussian noise. It requires a large amount of historical trajectory data and relies heavily on prior motion models. Data-driven approaches primarily address three key aspects: target feature extraction, state prediction, and data association. Feature extraction leverages the unique morphology and structure of the target to extract its apparent features. State prediction utilizes the target's motion characteristics to construct a state transition model to obtain the target's position, velocity, and other state variables at future time steps. Data association primarily addresses the matching of trajectories with detections. Data-driven approaches rely on extensive data annotation. Currently, datasets for infrared small target tracking from space-based remote sensing satellites are limited. Furthermore, factors such as the satellite's orbit, attitude, and payload design parameters directly impact image quality, making it difficult to adapt to all scenarios through training.

[0005] Both traditional and deep learning methods focus on trajectory prediction and data association during target tracking, while ignoring the impact of detection uncertainty on the tracking process. In traditional methods, detection uncertainty directly impacts the accuracy of model observations, which in turn affects the target state update process. When the target performs strong maneuvers, the impact of observation error is amplified, easily leading to trajectory breakage. In deep learning methods, detection uncertainty is not reflected in the final tracking results, and observation error, which should be considered, is ignored, resulting in inaccurate association and positioning of maneuvering targets. Summary of the Invention

[0006] To address the problems of detection result uncertainty, strong target motion state maneuvering, and track loss caused by weak targets from the perspective of space-based detection, the present invention provides a high-precision continuous tracking method for strong and weak infrared targets from the perspective of space-based detection. This method obtains the target uncertainty measurement in the non-maximum suppression stage of the detection results and applies it to the target state update and data association stage, thereby fully transferring the detector's empirical distribution to the tracking process and improving tracking accuracy. To cope with complex target maneuvering states, the tracking algorithm is designed using the interactive multi-model (IMM) technology route. To reduce the dependence on the prior motion model, a dynamic Markov transfer matrix construction method is specifically designed to update the model transition probability by comprehensively modulating historical dynamic information and current static information. Finally, in the data association stage, the advantages of IoU and NWD are combined to associate targets of different scales, and the uncertainty is transferred to the cost calculation process. Based on the scale-invariant and energy-invariant assumptions, unclear matching trajectories and targets are processed to prevent track breakage caused by matching failures, thereby achieving high-precision continuous tracking of the target.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] A high-precision continuous tracking method for a strong maneuvering infrared dim small target under the perspective of space-based detection includes the following steps:

[0009] Step 1: Measure the variance of each target detection result relative to the retained box after non-maximum suppression, complete the uncertainty measurement of the detection result, propagate it to the tracker and quantify the observation noise of the IMM tracker;

[0010] Step 2: Obtain the initial state of the mixture and the corresponding covariance according to the conditional probability matrix of the model transfer, use the obtained initial state and covariance to complete the sigma point sampling, and then use the UKF to predict the state of the sampled sigma points to obtain the predicted state of the target, the corresponding covariance and the predicted observation mean;

[0011] Step 3: Taking the predicted observation mean and detection results obtained in step 2 as input, we use the scale-aware joint NWD-IOU metric to calculate the association cost of objects of different scales, and obtain the detection results and predicted observation covariance associated with the historical trajectory;

[0012] Step 4: Using the sampled sigma points, predict the observation mean and the associated detection results, and calculate the state mean and covariance of the model update as the input of the IMM tracking algorithm;

[0013] Step 5: The state mean and covariance obtained in step 4 are combined with the associated detection results to calculate the likelihood probability of each model, and the target state mean and covariance of the IMM comprehensive output are obtained to complete the model probability update and data fusion;

[0014] Step 6: Use the associated detection results and observation predictions to obtain the normalized prediction error of the model, and combine the likelihood probability calculation results and the model probability change rate at two consecutive historical moments to autonomously update the Markov transfer matrix.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] (1) Taking the non-maximum suppression result of the detection as the benchmark, the covariance of all regressed target bounding boxes is calculated with it as a quantitative description of the detection uncertainty, and the covariance is injected into the observation noise in the IMM to accurately express the actual distribution of the observation value, thereby improving the accuracy of the target state estimation.

[0017] (2) The uncertainty measurement of detection is introduced into the data association process. The weighted values of IoU and NWD are used as distance metrics. The single distance metric between trajectory and detection is expanded to a distribution range, so that different detection results with close distances to a certain trajectory have ambiguous matching relationships. On this basis, an auxiliary distance metric is constructed based on the target's temporal scale and energy invariance assumptions to eliminate the ambiguous matching relationship between trajectory and detection, thereby improving the robustness of the data association process and the accuracy of tracking.

[0018] (3) Based on the historical model probability change rate and the model likelihood probability and state prediction error at the current moment, a dynamic modulation method of the Markov transfer matrix is proposed, which enables the algorithm to autonomously adjust the weighted values of different motion models based on the changing state of the target, thereby adapting to complex maneuvering targets and ultimately improving the accuracy of target state prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the high-precision continuous tracking method for strong maneuvering infrared dim small targets from the perspective of space-based detection;

[0020] Figure 2 Schematic diagram of uncertainty measurement of test results;

[0021] Figure 3 Schematic diagram of uncertainty of IoU;

[0022] Figure 4 Schematic diagram of ambiguous matching of data association;

[0023] Figure 5 To track the results. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0025] The present invention provides a high-precision continuous tracking method for strong maneuvering infrared dim small targets from the perspective of space-based detection. The method measures the uncertainty of the detection results to obtain the empirical distribution of the detector in the reasoning process, injects the uncertainty into the multi-model observation noise based on the unscented Kalman filter, and thus transfers the distribution knowledge of the detection process to the tracking process; introduces a dynamic modulation mechanism in the Markov transfer process of the multi-model mixture, so that the tracking model can autonomously adapt to the target with changing maneuvering state; introduces the uncertainty of detection into the data association method, and constructs the distance cost matrix of trajectory and detection in combination with the scale and energy invariance assumptions to improve tracking accuracy. Figure 1 As shown, the specific steps include:

[0026] Step 1: Measure the variance of each target detection result relative to the retained box after non-maximum suppression, complete the uncertainty measurement of the detection result, propagate it to the tracker and quantify the observation noise of the IMM tracker. The specific steps are as follows:

[0027] Step 1: For a certain target, the target detection result is a series of bounding boxes and confidence levels. The detection result is described as (x n ,y n ,w n ,h n ,s n ), n=1,2,…,N tar , N tar Indicates the number of detection boxes of the target, (x n ,y n ) represents the center position of the nth detection box, (w n ,h n ) represents the width and height of the nth detection box, s n ∈[0,1] represents the confidence of the nth detection box.

[0028] Step 1 and 2: After the detection results are processed by non-maximum suppression (NMS), the optimal target box is retained.

[0029] Step 1-3: Model the detection results as a spatial distribution with a standard deviation of σ. Obtain σ by measuring the variance of each target detection result relative to the retained box after non-maximum suppression, and propagate it to the tracker to quantify the observation noise of the IMM tracker and characterize the distribution range of the observation value. Figure 2 As shown, the covariance of the detection results at the kth moment after modeling is σx , σ y Respectively represent the standard deviation of the horizontal and vertical coordinates of the center position of the detection frame, σ w , σ h denote the standard deviation of the width and height of the detection box, respectively, and diag(·) denotes a diagonal matrix.

[0030] Step 2: Obtain the initial state of the mixture and the corresponding covariance based on the conditional probability matrix of the model transfer. The Unscented Kalman Filter (UKF) captures the statistical characteristics of the Gaussian distribution through a set of carefully selected points (i.e., sigma points), thereby replacing the linearization process of the traditional Kalman filter. The obtained initial state and covariance are used to complete the sigma point sampling, and then the UKF is used to predict the state of the sampled sigma points to obtain the predicted state of the target, the corresponding covariance, and the predicted observation mean. The specific steps are as follows:

[0031] Step 21: Define the target state quantity of the IMM filter output at the kth moment as s k =[x k ,y k ,w k ,h k ] Τ , where x k ,y k Represents the center position of the detection frame at the kth moment, w k , h k Represent the width and height of the detection frame at the kth moment respectively. Considering the uniform acceleration straight line (CV) model, uniform acceleration straight line (CA) model and uniform turning (CT) model, the state space of the target is obtained as follows: in and Respectively represent the target's movement speed in the row direction and column direction at the kth moment, θ k represents the rotation angle of the target at the kth moment.

[0032] Step 22: Define the conditional probability matrix for model transfer in represents the probability of transferring from model i to model j, Represents different motion models at k moments, t=1,2,…N, N is the number of models, state quantity and covariance They represent the state quantity and covariance output by model i at time k-1 respectively, and the mixed initial state and the corresponding covariance are calculated according to the model transition probability π:

[0033]

[0034] Where, and Respectively represent the initial state and covariance of the j-th model after mixing, π ij is calculated as follows:

[0035]

[0036] Where, is the Markov transition probability from model i to model j at the k-1th moment, is the probability of model i at time k-1, is an intermediate variable, which is calculated for normalization. Adaptive dynamic adjustment based on tracking situation to meet the precise tracking requirements of complex maneuvering targets.

[0037] Steps 2 and 3: For each model j, use UKF to predict the state. UKF replaces the complex linearization process by sampling sigma points and statistically predicting the distribution of data through nonlinear propagation. The sigma point sampling is as follows:

[0038]

[0039] Where, is the sigma point matrix, L represents the dimension of the state, and λ is the parameter of UKF, which is λ=α 2 (L+κ)-L, where α is the adjustment parameter and κ is 3-L.

[0040] Use the state transition function f for each sigma point j (·) Perform state prediction:

[0041]

[0042] Step 24: Get the prediction status and the corresponding covariance

[0043]

[0044] Where, and Represent the weights of the UKF state mean and covariance calculation process, Q j represents the process noise covariance of the j-th model.

[0045] Step 25: Propagate the predicted sigma point through the observation function h(·):

[0046]

[0047] Step 26: Predict the mean of the observations For subsequent data association, the association process is described in step three:

[0048]

[0049] Step 27: Predict observation covariance The likelihood probability calculation process used in step 4 is as follows:

[0050]

[0051] Where R k is the covariance of the detection result at the kth moment, obtained in step 1.

[0052] Step 3: Take the predicted observation mean and detection results obtained in step 2 as input, use the scale-aware joint NWD-IOU metric to complete the association cost calculation of targets of different scales, and obtain the detection results and predicted observation covariance associated with the historical trajectory. The specific steps are as follows:

[0053] Step 31: To calculate the associated cost of targets of different scales, for trajectory T p (Derived from the predicted observed mean ) and detection d q , using scale-aware joint NWD-IOU to measure the cost C between target boxes pq :

[0054]

[0055] Where A is the area of the detection box, C pq is the cost value, C is the hyperparameter of NWD cost, IoU and NWD are their respective metric values, and since the detection results are uncertain, T p and d q The IoU and NWD are extended from a single value to a range, and the corresponding cost value is Represents the mean cost between target boxes, ΔC pq represents the standard deviation of the cost value between target boxes. According to equation (12), we can know that:

[0056]

[0057] In the formula, ΔIoU and ΔNWD represent the standard deviation of IoU and NWD respectively. Since the uncertainty value A has little perturbation on the weighting coefficient of the above formula, in order to simplify the calculation, A is regarded as a constant and calculated from the detection results after NMS. Therefore, the factors that ultimately affect the cost value range are ΔIoU and ΔNWD. In the estimation process of ΔIoU, since the uncertainty of the intersection area is more sensitive to the calculation result of IoU, the uncertainty of the intersection part is used to represent ΔIoU, that is, ΔIoU / IoU=ΔA∩ / A ∩ ,like Figure 3 As shown, for the target boxes (x1, y1, h1, w1) and (x2, y2, h2, w2), the inaccuracies are (Δx1, Δy1, Δh1, Δw1) and (Δx2, Δy2, Δh2, Δw2) respectively. According to the principle of linear error propagation, the relationship between ΔIoU and IoU can be obtained as follows:

[0058]

[0059] Where A ∩ Represents the intersection between target frames, ΔA ∩ Represents the standard deviation of the intersection between target boxes. The uncertainty value Δr represents the standard deviation of the corresponding measurement value, r∈{x1,y1,w1,h1,x2,y2,w2,h2}, which is obtained by measuring the standard deviation of each detection result in step 1. Through the above calculation, the range of IoU can be obtained as Represents the average IoU metric between the actual trajectory and the predicted trajectory of the target, ΔIoU pq Indicates the standard deviation of the IoU metric between the target detection box and the predicted box. In the calculation process of ΔNWD, due to the central point Wasserstein distance d c and scaled Wasserstein distance d s The calculations are all nonlinear processes. To simplify the calculations, the first-order Taylor expansion is used to approximate the derivation process, and the high-order terms are discarded. The relationship between ΔNWD and NWD is as follows:

[0060]

[0061] The method of obtaining the uncertainty value Δr is consistent with the ΔIoU calculation process. Through the above calculation, the range of NWD values can be obtained as follows: Finally, we get the trajectory T p and detection d q The cost value range is

[0062] Step 3.2: In the association process, the cost matrix needs to be solved To obtain the optimal distribution between trajectories and detections, where P represents the number of trajectories at the current moment and Q represents the number of detections at the current moment. However, after introducing uncertain measurement values, each element in the cost matrix is expanded from a single value to an interval range, and the single matching relationship between trajectories and detections is destroyed (e.g. Figure 4As shown), an ambiguous matching situation is introduced. In order to correctly associate the trajectory with the detection d q , introducing an additional cost matrix (where Q s represents the number of detections with ambiguous matches), replaces the ambiguous items in the original cost matrix C, and finally uses the Hungarian algorithm to obtain the association results between trajectories and detections.

[0063] Step 3. Design an additional cost matrix U based on the following two prior knowledge of space-based infrared small target imaging:

[0064] (1) The scale of the target in two consecutive frames is basically the same, that is, the matching degree of the target with a large scale difference should be reduced;

[0065] (2) The energy of the target in two consecutive frames is basically the same, that is, the matching degree of the target with too large energy difference should be reduced.

[0066] Based on the above two assumptions, U consists of two parts: the shape constraint matrix U s And the energy constraint matrix U e , U s is calculated as follows:

[0067]

[0068] Where AT p 、 and σ T denote the bounding box area of track p, the average area of track p, and the mean square error of track area, respectively. is the area of the normalized bounding box with a mean of 0 and a mean square error of 1. Similarly, is the normalized result of the detection of ambiguous matches, Is the distance metric function between the two, which is normalized to [0, 1]. The closer the two are, the The closer to 1.

[0069] U e The calculation method is the same as U s Slightly different, since the target in the infrared scene may move in different backgrounds, the energy of the target point can be regarded as the cumulative value of the target energy and the background energy. Therefore, directly using the grayscale of the target point for calculation will introduce background interference factors. In order to solve this problem, the signal-to-clutter ratio (SCR) measurement method of the target and background is used to remove the influence of background energy. The gray value of the target point is The background grayscale mean is The background mean square error is Signal-to-noise ratio The calculation method is:

[0070]

[0071] Get U e The calculation method is:

[0072]

[0073] Where, ST p 、 and σ ST They represent the signal-to-noise ratio of trajectory p, the mean of the signal-to-noise ratio of trajectory, and the mean square error of the signal-to-noise ratio of trajectory. is the normalized signal-to-noise ratio of trajectory p. Similarly, Is to detect q s The standardized signal-to-noise ratio, Measure the distance between the two and normalize it to [0, 1]. Therefore, the additional cost matrix U is:

[0074]

[0075] Where η is the coefficient that balances shape constraints and energy constraints.

[0076] Step 34: Observed predicted values in IMM During the matching process with the detection, each model will produce an independent association result, and the detection result after association is defined as Used for subsequent model state updates.

[0077] Step 4: Use the sampled sigma points to predict the observation mean and the associated detection results, calculate the state mean and covariance of the model update, and use them as the input of the IMM tracking algorithm. The specific steps are as follows:

[0078] Step 41: Using the predicted sigma point (Equation (6)), predicted state (Equation (7)), the associated detection results and the predicted observed mean (Equation (10)) is calculated to get and The cross-covariance matrix is:

[0079]

[0080] Step 42: Kalman gain is calculated by the cross-covariance matrix and the observation covariance. According to step 1, the measured observation covariance R can be obtained. k , so the Kalman gain for:

[0081]

[0082] Among them, Cov is the covariance matrix.

[0083] Step 43: Update the state mean and covariance for:

[0084]

[0085] Get the updated state mean of each model and covariance After that, it is used as the input of the IMM tracking algorithm in the next time step.

[0086] Step 5: Based on the state mean and covariance obtained in step 4, combined with the associated detection results, calculate the likelihood probability of each model, obtain the target state mean and covariance of the IMM comprehensive output, and complete the model probability update and data fusion. The specific steps are as follows:

[0087] Step 51: After obtaining the updated state mean and covariance, combine them with the observed values Calculate the likelihood probability of each model to characterize the degree of match between the model's predicted observations and the actual observations. The model likelihood probability is calculated using Bayesian theory. The specific calculation method is:

[0088]

[0089] Where, represents the predicted observation covariance (Equation (11)), is the predicted observation mean (Equation (10)).

[0090] Step 52: Probability of model j By likelihood probability and Markov transition probability The calculation method is as follows:

[0091]

[0092] Where N is the number of models.

[0093] Step 53: According to Combine the state estimation results of all models and The target state mean and covariance of the IMM comprehensive output are obtained as follows:

[0094]

[0095]

[0096] Step 6: Use the associated detection results and observation predictions to obtain the normalized prediction error of the model, and combine the likelihood probability calculation results and the model probability change rate at two consecutive historical moments to autonomously update the Markov transfer matrix.

[0097] Since the final output fusion target state is affected by the Markov transition matrix The traditional IMM method uses a priori settings to determine the state transition matrix. Consequently, complex target motion patterns often involve numerous state transitions, making the priori approach inadequate for frequently changing maneuvering states. Therefore, a dynamic Markov transition matrix is designed that comprehensively considers model prediction errors, likelihood probability calculations, and model probability changes between two consecutive historical moments to autonomously update the transition matrix, thereby improving the accuracy of the IMM algorithm. The specific steps are as follows:

[0098] Step 61: The normalized state prediction error of model j at time k-1 is The normalized likelihood probability at time k is The model probabilities at time k-1 and k are and Then the model probability change rate at time k is Observational prediction and detection The closer, the The closer it is to 0, the better the model matches the real situation, and the higher the model transfer probability should be. The larger the value, the higher the likelihood probability of model j, and the higher the model transition probability should be. Indicates the change in the degree of matching between the model and the real motion at two consecutive moments. This indicates that as the degree of matching between model j and the target motion increases, the model transfer probability should increase.

[0099] Step 62: Based on the above analysis, the estimated Markov transition probability at time k is for:

[0100]

[0101] Step 63: Through Normalize so that the sum of each row element of the Markov transfer matrix is 1, then:

[0102]

[0103] The dynamic adjustment of the Markov transfer matrix occurs at the output of each time step and serves as the input of the next time step, replacing the inherently unchanging prior values in the traditional IMM. This allows the algorithm to dynamically adjust the transition probability between models according to the actual state and adapt to the goal of complex maneuvers.

[0104] The comparison results of the proposed method with classic methods such as VB-EOT-SN, PMB-EOT-BP, TrPMBM, TPMBM, Gaussian CD-PMBM and MEM-EKF are shown in Table 1.

[0105] Table 1 Performance comparison results of different tracking methods

[0106]

[0107] The proposed method achieves optimal performance. Compared with TPMBM, the continuous tracking rate, tracking success rate, tracking accuracy and high-order tracking accuracy indicators are improved by 3.71%, 2.32%, 0.29% and 0.54%, respectively. The position estimation accuracy and identity switching times are reduced by 0.74 and 22, respectively. This is mainly due to the multi-strategy fusion method such as detection distribution measurement, interactive multi-model, and dynamic Markov transfer matrix, which enables the algorithm to adapt to highly maneuverable targets, thereby ensuring the continuity and robustness of target trajectory tracking.

[0108] The 10 sequences shown in Table 2 are selected to compare the tracking results of different methods, such as Figure 5 As shown in FIG, the method proposed in the present invention can adapt to various complex environments and target motion states and identity switching. Figure 5 In (a), the VB-EOT-SN and MEM-EKF methods produce more false matches, while the proposed method produces fewer false trajectories. Figure 5 In (b), due to the large number of false detections in the striped background during the detection phase, other methods produced more error tracks, while the method proposed in this invention controlled the error tracks to a lower level. Figure 5 In (c), the movement of the broken clouds and the temporal brightness changes cause methods such as TrPMBM to mistakenly associate the cloud movement with the target, and a large number of trajectory switches occur, while the method proposed in this paper still robustly associates the target trajectory. Figure 5 In (d), strong noise interferes with the detection results, resulting in missed detections. Therefore, methods such as MEM-EKF produce more track switching. However, the method proposed in this invention has the best tracking effect due to the configuration of multiple models. Figure 5 In (e), methods such as TPMBM form continuous error track segments, while the unique correlation strategy of the method proposed in this invention suppresses the tracks generated by clutter to the greatest extent. Figure 5In (f), due to the target trajectory interaction and insufficient detection capability, the target trajectory associated by the VB-EOT-SN method is wrong, while the method proposed in this invention can adapt well to this situation and achieve the best association effect. Figure 5 In (g), the land-sea dividing line generates a large number of erroneous tracks due to the high contrast between light and dark coupled with satellite jitter. The proposed method has the best ability to suppress clutter tracks. Due to the presence of a strong maneuvering target, the Gaussian CD-PMBM method fails to track the target during the first maneuver, resulting in track switching. However, the proposed method can still robustly associate the target. Figure 5 In (h), the motion state of each target is different. The method using a single motion model results in continuous trajectory switching. The method proposed in this invention can continuously associate targets over a long time span. Figure 5 In (i), the intersection between the trajectories causes MEM-EKF to associate the two targets into one target. The method proposed in this paper can make good use of the prior estimation of the target state of each target and avoid trajectory crosstalk. Figure 5 In (j), since the target is submerged in the strong noise background, the MEM-EKF method has difficulty in accurately matching the target, resulting in incorrect trajectory association and identity switching. The method proposed in this paper can continuously and stably associate the target and maintain fewer incorrect trajectories.

[0109] Table 2 Different tracking image sequences for performance comparison

[0110]

[0111]

Claims

1. A high-precision continuous tracking method for strong maneuvering infrared dim small targets under the perspective of space-based detection, characterized by The method comprises the following steps: Step 1: Measure the variance of each target detection result relative to the retained box after non-maximum suppression, complete the uncertainty measurement of the detection result, propagate it to the tracker and quantify the observation noise of the IMM tracker; Step 2: Obtain the initial state of the mixture and the corresponding covariance according to the conditional probability matrix of the model transfer, use the obtained initial state and covariance to complete the sigma point sampling, and then use the UKF to predict the state of the sampled sigma points to obtain the predicted state of the target, the corresponding covariance and the predicted observation mean; Step 3: Taking the predicted observation mean and detection results obtained in step 2 as input, we use the scale-aware joint NWD-IOU metric to calculate the association cost of objects of different scales, and obtain the detection results and predicted observation covariance associated with the historical trajectory; Step 4: Using the sampled sigma points, predict the observation mean and the associated detection results, and calculate the state mean and covariance of the model update as the input of the IMM tracking algorithm; Step 5: The state mean and covariance obtained in step 4 are combined with the associated detection results to calculate the likelihood probability of each model, and the target state mean and covariance of the IMM comprehensive output are obtained to complete the model probability update and data fusion; Step 6: Use the associated detection results and observation predictions to obtain the normalized prediction error of the model, and combine the likelihood probability calculation results and the model probability change rate at two consecutive historical moments to autonomously update the Markov transfer matrix.

2. The high-precision continuous tracking method for a strong maneuvering infrared small target under the space-based detection perspective according to claim 1 is characterized in that The specific steps of step one are as follows: Step 1: For a certain target, the target detection result is a series of bounding boxes and confidence levels. The detection result is described as (x n ,y n ,w n ,h n ,s n ), n=1,2,…,N tar , N tar Indicates the number of detection boxes of the target, (x n ,y n ) represents the center position of the nth detection box, (w n ,h n ) represents the width and height of the nth detection box, s n ∈[0,1] represents the confidence of the nth detection box; Step 1 and 2: After the detection results are subjected to non-maximum suppression, the optimal target box is retained; Step 1 and 3: Model the detection results as a spatial distribution with a standard deviation of σ. Obtain σ by measuring the variance of each target detection result relative to the retained box after non-maximum suppression, propagate it to the tracker, quantify the observation noise of the IMM tracker, characterize the distribution range of the observation value, and model the covariance of the detection result at the kth moment. σ x , σ y Respectively represent the standard deviation of the horizontal and vertical coordinates of the center position of the detection frame, σ w , σ h denote the standard deviation of the width and height of the detection box, respectively, and diag(·) denotes a diagonal matrix.

3. The high-precision continuous tracking method for strong maneuvering infrared dim small targets under the space-based detection perspective according to claim 1 is characterized in that The specific steps of step 2 are as follows: Step 21: Define the target state quantity of the IMM filter output at the kth moment as s k =[x k ,y k ,w k ,h k ] Τ , where x k ,y k Represents the center position of the detection frame at the kth moment, w k , h k Represent the width and height of the detection frame at the kth moment respectively. Considering the uniform acceleration straight line CV model, uniform speed change straight line CA model and uniform speed turning CT model, the state space of the target is obtained as in and Respectively represent the target's movement speed in the row direction and column direction at the kth moment, θ k represents the rotation angle of the target at the kth moment; Step 22: Define the conditional probability matrix for model transfer in represents the probability of transferring from model i to model j, Represents different motion models at k moments, t=1,2,…N, N is the number of models, state quantity and covariance They represent the state quantity and covariance output by model i at time k-1 respectively, and the mixed initial state and the corresponding covariance are calculated according to the model transition probability π: Where, and Respectively represent the initial state and covariance of the j-th model after mixing; Steps 2 and 3: For each model j, use UKF to predict the state. UKF replaces the complex linearization process by sampling sigma points and statistically predicting the distribution of data through nonlinear propagation. The sigma point sampling is as follows: Where, is the sigma point matrix, L represents the dimension of the state, and λ is the parameter of UKF, which is λ=α 2 (L+κ)-L, where α is the adjustment parameter and κ is 3-L; Use the state transition function f for each sigma point j (·) Perform state prediction: Step 24: Get the prediction status and the corresponding covariance Where, and Represent the weights of the UKF state mean and covariance calculation process, Q j represents the process noise covariance of the j-th model; Step 25: Propagate the predicted sigma point through the observation function h(·): Step 26: Predict the mean of the observations Step 27: Predict observation covariance Where R k is the covariance of the detection result at the kth moment.

4. The high-precision continuous tracking method for a strong maneuvering infrared small target under the space-based detection perspective according to claim 3 is characterized in that The π ij is calculated as follows: Where, is the Markov transition probability from model i to model j at the k-1th moment, is the probability of model i at time k-1, is an intermediate variable.

5. The high-precision continuous tracking method for a strong maneuvering infrared small target under the space-based detection perspective according to claim 1 is characterized in that The specific steps of step three are as follows: Step 31: For trajectory T p and detection d q , using scale-aware joint NWD-IOU to measure the cost C between target boxes pq : Where A is the area of the detection box, C pq is the cost value, C is the hyperparameter of NWD cost, IoU and NWD are their respective metric values, and since the detection results are uncertain, T p and d q The IoU and NWD are extended from a single value to a range, and the corresponding cost value is Represents the mean cost between target boxes, ΔC pq represents the standard deviation of the cost value between target boxes, according to equation (12): Where ΔIoU and ΔNWD represent the standard deviations of IoU and NWD metrics, respectively; For the target boxes (x1, y1, h1, w1) and (x2, y2, h2, w2), the inaccuracies are (Δx1, Δy1, Δh1, Δw1) and (Δx2, Δy2, Δh2, Δw2) respectively. According to the principle of linear error propagation, the relationship between ΔIoU and IoU is as follows: Where A ∩ Represents the intersection between target frames, ΔA∩ represents the standard deviation of the intersection between target frames, and the uncertainty value Δr represents the standard deviation of the corresponding measurement value, r∈{x1,y1,w1,h1,x2,y2,w2,h2}. The range of IoU is Represents the average IoU metric between the actual trajectory and the predicted trajectory of the target, ΔIoU pq Indicates the standard deviation of the IoU metric between the detection box and the prediction box of the target; The relationship between ΔNWD and NWD is as follows: It represents the average NWD metric value between the target detection box and the prediction box, and ΔNWD represents the standard deviation of the NWD metric value between the target detection box and the prediction box. The value range of NWD is Finally, we get the trajectory T p and detection d q The cost value range is Step 3.2: In the association process, solve the cost matrix To obtain the optimal distribution between trajectories and detections, where P represents the number of trajectories at the current moment and Q represents the number of detections at the current moment, in order to correctly associate trajectories with detections d q , introducing an additional cost matrix Replace the ambiguous items in the original cost matrix C, Q s It represents the number of detections with unclear matches. Finally, the Hungarian algorithm is used to obtain the association result between the trajectory and the detection; Step 3. Design an additional cost matrix U based on the following two prior knowledge of space-based infrared small target imaging: (1) The scale of the target in two consecutive frames is basically the same, that is, the matching degree of the target with a large scale difference should be reduced; (2) The energy of the target in two consecutive frames is basically the same, that is, the matching degree of the target with too large energy difference should be reduced; Based on the above two assumptions, U consists of two parts: the shape constraint matrix U s And the energy constraint matrix U e , U s is calculated as follows: Where AT p 、 and σ T They represent the bounding box area of track p, the average area of the track, and the mean square error of the track area, respectively. is the normalized bounding box area with a mean of 0 and a mean square error of 1, is the normalized result of the detection of ambiguous matches, is the distance metric function between the two; U e is calculated as follows: For detection The gray value of the target point is The background grayscale mean is The background mean square error is Signal-to-noise ratio The calculation method is: Get U e The calculation method is: Where, ST p 、 and σ ST They represent the signal-to-noise ratio of trajectory p, the mean of the signal-to-noise ratio of trajectory, and the mean square error of the signal-to-noise ratio of trajectory. is the normalized signal-to-noise ratio of trajectory p, Is to detect q s The standardized signal-to-noise ratio, Measure the distance between the two and normalize it to [0, 1]; Therefore, the additional cost matrix U is: Where η is the coefficient of balancing shape constraint and energy constraint; Step 34: Observed predicted values in IMM During the matching process with the detection, each model will produce an independent association result, and the detection result after association is defined as Used for subsequent model state updates.

6. The high-precision continuous tracking method for a strong maneuvering infrared small target under the space-based detection perspective according to claim 3 is characterized in that The specific steps of step 4 are as follows: Step 41: Using the predicted sigma point Prediction Status Associated test results and the predicted observed mean Calculated and The cross-covariance matrix is: Step 42: Kalman Gain for: Among them, Cov is the covariance matrix; Step 43: Update the state mean and covariance for:

7. The high-precision continuous tracking method for a strong maneuvering infrared small target under the space-based detection perspective according to claim 6 is characterized in that The specific steps of step five are as follows: Step 51: After obtaining the updated state mean and covariance, combine them with the observed values Calculate the likelihood probability of each model to characterize the degree of match between the model's predicted observations and the actual observations. The model likelihood probability is calculated using Bayesian theory. The specific calculation method is: Step 52: Probability of model j By likelihood probability and Markov transition probability The calculation method is as follows: Step 53: According to Combine the state estimation results of all models and The target state mean and covariance of the IMM comprehensive output are obtained as follows:

8. The high-precision continuous tracking method for a strong maneuvering infrared small target under the space-based detection perspective according to claim 7 is characterized in that The specific steps of step six are as follows: Step 61: The normalized state prediction error of model j at time k-1 is The normalized likelihood probability at time k is The model probabilities at time k-1 and k are and Then the model probability change rate at time k is Indicates the change in the degree of matching between the model and the real motion at two consecutive moments; Step 62: Markov transition probability estimate at time k for: Step 63: Through Normalize so that the sum of each row element of the Markov transfer matrix is 1, then:

Images