A motion-feature-assisted active sonar method for real-time tracking of small targets

Abstract

This invention relates to a motion feature-assisted real-time tracking method for small targets using active sonar. First, motion features are extracted from the active sonar echo sequence to obtain target position measurement information and target motion features. Then, the target's velocity measurement information is estimated using the target position measurement information and target motion features. Next, the target state is predicted (if it is the tracking start time or a newly generated trajectory, state initialization is required; if it is an existing trajectory, the state estimated at the previous moment is used for prediction). Next, measurement correlation is performed. Finally, the state is updated, and it is determined whether the termination time has been reached. If it has, the tracking trajectory is output; otherwise, the state prediction for the next moment continues. This invention requires less computation and can meet the requirements of real-time moving target detection. Even in complex clutter backgrounds and when the target is maneuvering, it can establish an effective correspondence between measurements and the target, effectively preventing reduced tracking accuracy and target loss.

Description

Technical Field

[0001] This invention belongs to the field of sonar information processing and relates to a motion feature-assisted real-time tracking method for small underwater targets detected by an active sonar system. Background Technology

[0002] Tracking small moving targets in complex underwater environments using active sonar systems is a significant research topic with important applications in both military and civilian fields. Existing methods typically rely on the classic Kalman filter, optimizing the target's true state and updating its measurements to track the target's state in real time, thus achieving automatic underwater target tracking. However, for maneuvering targets, their speed and direction of motion can change significantly in a short period, and even small target detection errors can lead to filter divergence and target loss. Furthermore, factors such as clutter can cause sonar systems to generate multiple measured tracks for a single underwater target, resulting in misidentification of a single target as multiple targets. Therefore, underwater target tracking algorithms need to establish an effective correspondence between measurements and targets, i.e., multi-target data association is required to prevent reduced tracking accuracy or even target loss.

[0003] Among data association methods, Bayesian methods are the most important. These include Nearest Neighbor (NN) data association methods, Global Nearest Neighbor (GNN) data association methods, Probabilistic Data Association (PDA) methods, Joint Probabilistic Data Association (JPDA) methods, and Multiple Hypothesis Tracking (MHT) methods. A common problem with these methods is that when there are many underwater targets and a large amount of underwater clutter and false alarms, the association process can lead to combinatorial explosion and a sharp increase in computational cost.

[0004] Therefore, although the target tracking methods based on Kalman filters and Bayesian data association have been widely used in radar, video tracking and other fields, they are still subject to challenges from the interference of complex underwater clutter and the maneuverability of small underwater targets such as frogmen and UUVs. Directly applying these technologies to underwater small target tracking still has problems such as high false alarm rate and poor tracking accuracy. Summary of the Invention

[0005] Technical problems to be solved

[0006] To avoid the shortcomings of existing technologies, this invention proposes a motion feature-assisted active sonar real-time tracking method for small targets.

[0007] Technical solution

[0008] A motion-feature-assisted active sonar method for real-time tracking of small targets, characterized by the following steps:

[0009] Step 1: At time t, extract motion features using the echo sequence obtained by the active sonar system:

[0010] M t ={m t (x,y):-B≤x≤B;0≤y≤L};

[0011] Where, m t (x,y) represents the motion characteristics of each spatial resolution unit of the active sonar system, and B and L represent the beam opening angle and echo distance limit, respectively.

[0012] For M t Threshold detection is performed to obtain the motion characteristics of the potential moving target k at time t. and location measurement information

[0013] Step 2, Target velocity estimation:

[0014] Using the extracted motion features M t ={m t (x,y):-B≤x≤B;0≤y≤L} Estimate the target's velocity The velocity measurement information of the moving target k at time t

[0015] Step 3, State Initialization:

[0016] For the moving target k, if state initialization has already been performed, proceed directly to step 4. If state initialization has not been performed, the target's state is initialized using position measurement information. The average value of the initial state of the moving target k is... Covariance Matrix The covariance matrix Based on the sensor accuracy settings; after state initialization is complete, increment t by 1 and return to step 1;

[0017] Step 4, State Prediction: Predict the current state using the state from the previous time step:

[0018]

[0019]

[0020] Where F is the state transition matrix:

[0021]

[0022] in: and Let be the mean and covariance matrix of the state at the previous time step. If the previous time step is the initial time step, then... and Q is the initial value; Q is the noise covariance.

[0023] Step 5, Measurement Association: The predicted state in Step 4 is associated with the target position measurement information and velocity measurement information obtained in Steps 1 and 2.

[0024] First association: using the motion feature M obtained in step 1 t Several spatial connected regions are constructed using the connected component analysis method. For target k, the Euclidean distance between its predicted state and the centroid measurements of all connected regions is calculated, and the connected region with the closest centroid measurement to the predicted state is associated with it.

[0025] Second association: Calculate the Euclidean distance between all measurements in the connected region associated in the first association and the predicted state of the target k, and then associate the nearest measurement with the predicted state of the target k.

[0026] After two associations, there will still be unassociated measurements that generate new trajectories;

[0027] Step 6, Status Update: Used to correct the error in Step 4;

[0028] The target state is updated using the state predicted in step 4 and the measurements associated in step 5, and then the updated state is calculated using the following formula:

[0029]

[0030]

[0031]

[0032] Where H is the measurement matrix and R is the measurement error covariance, the value of which is related to the actual environment; These are measured values;

[0033] If the tracking has not ended, increment t by 1 and return to step 4; if it is the termination time, output the target trajectory.

[0034] Q is the noise covariance, obtained based on the actual system noise.

[0035] The threshold in step 1 is obtained using the method proposed in the literature "A threshold selection method from gray-level histograms".

[0036] Beneficial effects

[0037] This invention proposes a motion feature-assisted real-time tracking method for small targets using active sonar. First, motion features are extracted from the active sonar echo sequence to obtain target position measurement information and target motion features. Then, the target's velocity measurement information is estimated using the target position measurement information and target motion features. Next, the target's state is predicted (if it is the starting moment of tracking or a newly generated trajectory, state initialization is required; if it is an existing trajectory, the state estimated at the previous moment is used for prediction). Next, measurement correlation is performed. Finally, the state is updated, and it is determined whether the termination moment has been reached. If it has, the tracking trajectory is output; otherwise, the state prediction for the next moment continues.

[0038] The beneficial effects of this invention are:

[0039] 1. The underwater moving small target real-time tracking method designed by the method disclosed in this invention has a lower false alarm rate than traditional data association-based methods, including MHT-based tracking methods, and requires less computation, thus meeting the requirements for real-time moving target detection.

[0040] 2. Compared with the classic methods based on Kalman filters and generalized Kalman filters, the method disclosed in this invention can establish an effective correspondence between measurements and targets in complex clutter backgrounds and when the target is maneuvering, effectively preventing the reduction of tracking accuracy and the loss of targets. Attached Figure Description

[0041] Figure 1 This is a system diagram of the real-time tracking method for moving small targets according to the present invention.

[0042] Figure 2 The image is a sonar echo generated from the sea trial data used in this embodiment of the invention.

[0043] Figure 3 This is a comparison diagram of the tracking method used in this invention and the manually annotated real target trajectory, wherein... Figure 3 (a) shows the actual target trajectory manually annotated. Figure 3 (b) is the result of the tracking method used in this invention. Detailed Implementation

[0044] The present invention will now be further described in conjunction with the embodiments and accompanying drawings:

[0045] Challenged by the maneuverability of small targets and the complex clutter in port environments, traditional methods based on classical and generalized Kalman filters suffer from reduced tracking accuracy and even target loss. Bayesian data association methods, when faced with numerous underwater targets and significant underwater clutter and false alarms, suffer from combinatorial explosion and a sharp increase in computational load. Therefore, this invention proposes a novel motion feature-assisted active sonar real-time tracking method for small targets. To mitigate the impact of target maneuverability on tracking accuracy, a motion feature extraction module and a motion velocity estimation module are introduced to estimate the target's real-time motion information. Addressing the high computational load of Bayesian data association methods, a measurement association sub-module based on two data association steps is introduced into the tracker module to reduce unnecessary measurement associations.

[0046] like Figure 1 As shown, this invention provides a motion feature-assisted active sonar real-time target tracking method, comprising the following steps:

[0047] Step 1: Motion Feature Extraction

[0048] At time t, using the echo sequence obtained by the active sonar system, the motion features are extracted using the flux tensor method proposed in the paper "Flux TensorConstrained Geodesic Active Contours with Sensor Fusion for Persistent Object Tracking": M t ={m t (x,y):-B≤x≤B;0≤y≤L};

[0049] Where, m t (x,y) represents the motion characteristics of each spatial resolution unit of the active sonar system, and B and L represent the beam opening angle and echo distance limit, respectively.

[0050] The threshold is obtained using the method proposed in the paper "A threshold selection method from gray-level histograms", and this threshold is used to select M. t Threshold detection is performed to obtain the motion characteristics of the potential moving target k at time t. and location measurement information

[0051] Step 2, Target velocity estimation:

[0052] Using the extracted motion features M t ={m tThe target's velocity is estimated using the optical flow method proposed in the paper "Aniterative image registration technique with an application to stereo vision" (x,y):-B≤x≤B;0≤y≤L}. The velocity measurement information of the moving target k at time t

[0053] Step 3, State Initialization:

[0054] For the moving target k, if state initialization has already been performed, proceed directly to step 4. If state initialization has not been performed, the target's state is initialized using position measurement information. The average value of the initial state of the moving target k is... Covariance Matrix The covariance matrix Based on the sensor accuracy settings. After state initialization is complete, increment t by 1 and return to step 1;

[0055] Step 4: State Prediction

[0056] A state prediction module is constructed to predict the state at the current moment based on the state at the previous moment.

[0057] Predict the current state using equation (4):

[0058] Step 4, State Prediction: Predict the current state using the state from the previous time step:

[0059]

[0060]

[0061] Where F is the state transition matrix:

[0062]

[0063] in: and Let be the mean and covariance matrix of the state at the previous time step. If the previous time step is the initial time step, then... and Q is the initial value; Q is the noise covariance.

[0064] Step 5, Measurement Association: The predicted state in Step 4 is associated with the target position measurement information and velocity measurement information obtained in Steps 1 and 2.

[0065] First association: using the motion feature M obtained in step 1 tSeveral spatial connected regions are constructed using the connected component analysis method. For target k, the Euclidean distance between its predicted state and the centroid measurements of all connected regions is calculated, and the connected region with the closest centroid measurement to the predicted state is associated with it.

[0066] Second association: Calculate the Euclidean distance between all measurements in the connected region associated in the first association and the predicted state of the target k, and then associate the nearest measurement with the predicted state of the target k.

[0067] After two associations, there will still be unassociated measurements that generate new trajectories;

[0068] Step 6, Status Update:

[0069] The target state is updated using the state predicted in step 4 and the measurements associated in step 5, and then the updated state is calculated using the following formula:

[0070]

[0071] Where H is the measurement matrix and R is the measurement error covariance, the value of which is related to the actual environment; These are measured values;

[0072] If the tracking has not ended, increment t by 1 and return to step 4; if it is the termination time, output the target trajectory. Specific implementation examples:

[0074] like Figure 2 As shown, in the sonar echo image generated using the acoustic echo signal of the active sonar system, strong interference sources and target echoes coexist within the monitored spatiotemporal range. At this time, there is a large amount of clutter around the target (an arc-shaped bright band with a direction range of 50-150° and a distance range of 200-500° and a rectangular bright area with a direction range of 180-220° and a distance range of 200-500°).

[0075] like Figure 3 As shown in (a), small underwater targets exhibit maneuvering during their movement, including changes in direction and speed. Figure 3 As shown in (b), in the tracking trajectory diagram obtained by the tracking method used in this invention, most of the interference is suppressed and the target trajectory is clearly visible.

[0076] Furthermore, we compared the performance metrics of tracking methods based on the classical generalized Kalman filter framework, tracking methods based on the MHT framework, and the motion feature-assisted tracking method used in this invention. The results are shown in Table 1. It can be seen that the tracking method used in this invention significantly outperforms traditional methods in terms of performance metrics.

[0077] Table 1. Performance comparison between the tracking method used in this invention and traditional methods.

[0078] method TPD↑ TFAR↓ TLE↓ T-FRAG↓ ER↓ Generalized Kalman Filter 1 3.79 2.79 0.06 115.47 MHT 0.97 0.28 2.79 0.008 7.28 Motion feature assistance 1.0054 0.0163 1.51 0.0054 0.0128

[0079] TPD: Track Detection Probability; TFAR: Tracking False Alarm Rate; TLE: Track Positioning Error; T-FRAG: Track Fragmentation Rate;

[0080] ER: Execution speed. If the ER value is less than 1, the tracker performs real-time processing.

Claims

1. A motion-feature-assisted active sonar real-time target tracking method, characterized in that... The steps are as follows: Step 1: In t At any given moment, motion features are extracted from the echo sequence obtained by the active sonar system: ； in, Motion characteristics of each spatial resolution unit in an active sonar system. and These represent the beam opening angle and echo distance limits, respectively. right Threshold detection is performed to obtain potential moving targets. k exist t Motion characteristics at any moment and location measurement information ; Step 2, Target velocity estimation: Utilizing the extracted motion features Estimate the target's speed As t Constant movement target k Speed ​​measurement information ; Step 3, State Initialization: For the moving target k If state initialization has already been performed, proceed directly to step 4. If state initialization has not been performed, initialize the target's state using position measurement information. k The mean of the initial state is Covariance Matrix The covariance matrix Based on sensor accuracy settings; after state initialization is complete, set as follows: t Add 1, return to step 1; Step 4, State Prediction: Predict the current state using the state from the previous time step: in, F Here is the state transition matrix: in: and Let be the mean and covariance matrix of the state at the previous time step. If the previous time step is the initial time step, then... and Initialized value; Q For noise covariance; Step 5, Measurement Association: The predicted state in Step 4 is associated with the target position measurement information and velocity measurement information obtained in Steps 1 and 2. First association: using the motion features obtained in step 1 Several spatially connected regions are constructed using connected component analysis; for the target k Calculate the Euclidean distance between its predicted state and the centroid measurements of all connected regions, and associate the connected region whose centroid measurement is closest to the predicted state with it. Second association: Calculate all measurements and targets in the connected regions associated in the first association. k Predict the Euclidean distance between states, then compare the nearest measurement with the target. k The predicted state is associated; After two associations, there will still be unassociated measurements that generate new trajectories; Step 6, Status Update: The target state is updated using the state predicted in step 4 and the measurements associated in step 5, and then the updated state is calculated using the following formula: in, For the measurement matrix, The value of the measured error covariance is related to the actual environment; These are measured values; If the tracking has not ended, t Add 1 and return to step 4; if it is the termination time, output the target trajectory.

2. The motion feature-assisted active sonar real-time tracking method for small targets according to claim 1, characterized in that: The Q The noise covariance is obtained based on the actual system noise.

Images