Extended target tracking method based on long and short term memory network
Through the combination of long-term memory networks and Bayesian filtering framework, the problem of insufficient accuracy in traditional algorithms in extended target tracking is solved, and high-precision tracking of extended targets is achieved, combining the fitting ability of neural networks and the reliability of prior knowledge.
Patent Information
- Application Number
- CN202510667544.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-22
- Publication Date
- 2025-08-26
AI Technical Summary
When processing extended targets, traditional target tracking algorithms cannot effectively utilize rich point cloud data and ignore the long-term motion patterns of the target, resulting in insufficient tracking accuracy.
The movement characteristics of the long-term memory network learning targets are adopted, combined with the Bayesian filtering framework, memory, prediction and update links are constructed, and the long-term non-Markov and short-term variable speed characteristics of the target are captured by the long-term memory network, and extended target tracking is combined with prior knowledge.
High-precision tracking of the extension target is achieved, combining the fitting ability of the neural network and the reliability of prior knowledge, and improving the accuracy and interpretability of the tracking algorithm.
Smart Images
Figure CN120541431A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of extended target tracking, and in particular to an extended target tracking method based on a long short-term memory network. Background Art
[0002] Smart transportation and autonomous driving technologies have been hot topics in recent scientific and technological research. Accurately tracking dynamic targets, such as vehicles, is a key topic. Traditional target tracking algorithms, operating with limited target information, assume that the target can be viewed as a point mass, with its motion approximately following a first-order Markov model and a manually configured prior motion model. Based on this assumption, mathematical models are developed and algorithms are designed to achieve target tracking. However, with the continuous improvement of sensor resolution, target measurement information acquired through sensors has gradually shifted from single point data to a larger number of point clouds, making the assumption of treating the target as a point mass no longer reasonable. Furthermore, the actual motion of a target is constrained by physical laws. While explicit descriptions of motion models are difficult, extracting hidden patterns from the spatiotemporal data of the target's historical motion is becoming a more reliable approach. Manually configured prior motion models are too coarse in comparison. These shortcomings limit the performance of traditional point target tracking algorithms.
[0003] Many researchers consider the target as an extended target, describing it using its center of mass position and extended form, and establishing target tracking models and algorithms based on this. Koch et al. assumed that the extended form of the target is roughly elliptical (spherical) and, based on random matrix theory, derived a reliable extended target tracking algorithm under a large number of assumptions. Lan et al. expanded on this approach and studied tracking algorithms for non-ellipsoidal extended forms. However, these methods are still based on the assumptions of a first-order Markov model and a priori motion models, focusing only on the patterns of the target's short-term motion and ignoring the underlying patterns of the target's long-term motion. The discrepancy between these ideal models and the actual target is a key limitation that hinders further improvement in algorithm accuracy.
[0004] Neural network technology uses gradient descent methods to extract hidden information from massive amounts of data and approximate a complex real-world model using a large number of parameters, achieving significant breakthroughs in fields such as computer vision and natural language processing. The rise of this technology has provided a new approach to solving target tracking problems. Revach et al. combined the traditional point target tracking algorithm, the Kalman filter, with a neural network, using the neural network to learn the target's motion model, thereby improving the tracking accuracy of the Kalman filter algorithm. Yan et al. used a recurrent neural network to learn the non-Markov characteristics and model errors in the target's historical motion information, deriving a target tracking algorithm under multiple assumptions and achieving significant results in point target tracking scenarios. However, these methods are still based on the assumption of point targets and cannot fully utilize the rich measurement data of extended targets. Summary of the Invention
[0005] The present application provides an extended target tracking method based on a long short-term memory network to solve the problem of low accuracy in existing extended target tracking.
[0006] The technical solution of the present invention is: an extended target tracking method based on a long short-term memory network, comprising:
[0007] Step 1: Use the long short-term memory network to learn the motion characteristics of the target. The long-term memory is used to capture the non-Markov characteristics of the target's long-term motion, and the short-term memory is used to capture the speed variation characteristics of the target's short-term motion.
[0008] Step 2: Combining prior knowledge with the LSTM network, under reasonable assumptions, and based on the Bayesian filtering framework, derive the recursive form of the prediction and update steps of the extended target tracking method;
[0009] Step 3: Construct the memory link, prediction link and update link of the algorithm that combines the neural network and prior knowledge, and track the extended target by recursively executing the three links.
[0010] Furthermore, in step 1, the long short-term memory network is used to learn the motion characteristics of the target, the long-term memory is used to capture the non-Markov characteristics of the target's long-term motion, and the short-term memory is used to capture the speed-varying characteristics of the target's short-term motion, specifically including:
[0011] The target's motion state at multiple discrete moments is regarded as a string of time series data, and the target's motion state is recursively sent to the long short-term memory network. The long short-term memory network calculates the target's motion state according to its own parameters to obtain the target's motion characteristics at each moment.
[0012] By using the long short-term memory network, in the process of recursively receiving the target motion features, redundant information is selectively discarded, and the deeper motion feature information of the target is retained, so as to obtain the influence of the more distant historical motion features of the target on the current motion features, and then extract the non-Markov features of the target motion from them.
[0013] Using the long short-term memory network, the current motion features of the target are generated in the short-term memory, which is then interacted with the long-term memory to capture the random disturbances in the current motion of the target, and then the speed variation features of the target motion are extracted from it.
[0014] Furthermore, in step 2, the prior knowledge and the long short-term memory network are combined, specifically including:
[0015] Assume that the target center of mass motion state evolution model is
[0016]
[0017] where x k represents the motion state of the target center of mass, X k Indicates the extended form of the target, represents the actual target center of mass motion state evolution model, w k is the process noise, subscript k represents the time, and subscript 1:k-1 represents from time 1 to time k-1;
[0018] Since it is impossible to accurately obtain the actual target According to prior knowledge, an approximate target center of mass motion state evolution model f can be used k (·), such as the uniform linear motion model, this approximation process will introduce errors
[0019]
[0020] In the approximate target center of mass motion state evolution model f k (·) Target motion features obtained by superimposing the long short-term memory network Can reduce
[0021] Assume that the observation model of the target center of mass motion state is
[0022]
[0023] where z k represents the set of measurements of the target at time k, represents the actual target center of mass motion state observation model, v k is the observation noise;
[0024] Since it is impossible to accurately obtain the actual target According to prior knowledge, an approximate target center of mass motion state observation model h can be used k (·), this approximation process will introduce errors
[0025]
[0026] In the approximate target center of mass motion state observation model h k (·) The features related to target motion and observation obtained by superimposing the long short-term memory network Can reduce
[0027] In addition, due to model approximation, the covariance of variables such as target center of mass motion and extended morphology also differs from the actual covariance, and a neural network is also required to extract relevant information from the target motion features.
[0028] Furthermore, in step 2, under reasonable assumptions, the following are specifically included:
[0029] Considering the tracking process at time k, we make the following reasonable assumptions that are in line with reality:
[0030] Assume that the target center of mass moves x k and x k-1 and X k The probability density distribution of its evolution process is related to Approximately obey the expectation f k (x k-1 ), the variance is Q k Gaussian distribution of the target center of mass motion state, the posterior distribution p(x k |Z 1:k ,D) obey the expectation of x k|k , the variance is P k|k Gaussian distribution;
[0031] Assuming the target center of mass motion state evolution model error The probability density distribution of The variance is Gaussian distribution of the target center of mass motion state observation model error The probability density distribution of The variance is Gaussian distribution, assuming that the target extended morphological evolution model error The expectation of the probability density distribution is
[0032] Assuming the target expansion form X k and X k-1 The probability density distribution of its evolution process is related to The degree of freedom parameter is δ k , the matrix parameters are The Wishart distribution of the target expansion form p(X k |Z 1:k ,D) subject to the freedom parameter v k|k , the matrix parameter is X k|k The inverse Wishart distribution.
[0033] Hypothesized likelihood distribution of the measurement set For the expectation The variance is B k X k (B k ) T The product of Gaussian distributions of kRepresents the set of all measurements generated by the target at time k, with a total of n k indivual;
[0034] Assume that the short-term memory variable c k with x k-1 、X k-1 、c k-1 Related to D, the probability density distribution of its update link p(c k |x k-1 ,X k-1 ,c k-1 ,D) obey expectations The variance is Gaussian distribution of , where D represents the offline dataset for training and testing the neural network.
[0035] Furthermore, in step 2, the recursive form of the prediction phase of the extended target tracking method is derived based on the Bayesian filtering framework, specifically including:
[0036] The prediction model of the target center of mass motion state is:
[0037]
[0038] The recursive form of the target center of mass motion state covariance prediction is:
[0039]
[0040] Defining intermediate variables
[0041] λ k-1 =v k-1|k-1 -2d-2
[0042] The recursive form of target extended morphology uncertainty prediction is:
[0043]
[0044] The recursive form of target expansion morphology prediction is
[0045]
[0046] Furthermore, in step 2, the recursive form of the update link of the extended target tracking method is derived based on the Bayesian filtering framework, specifically including:
[0047] The observation model of the target center of mass motion state is:
[0048]
[0049] The cross-covariance between the target center of mass motion state and the target center of mass motion state observation is
[0050]
[0051] The covariance of the target center of mass motion state observation is
[0052]
[0053] The recursive form of the target center of mass motion state update is
[0054]
[0055] in Indicates measurement z k expectations;
[0056] The recursive form of the target center of mass motion state covariance update is:
[0057]
[0058] Target extended form innovation covariance
[0059]
[0060] The recursive form of the target extended morphology uncertainty update is
[0061] v k|k =v k|k-1 +n k
[0062] The recursive form of the target extended morphology update is derived as follows:
[0063]
[0064] in Indicates measurement z k The covariance of .
[0065] Furthermore, in step three, the memory link of the algorithm combining the neural network and prior knowledge is constructed, which specifically includes:
[0066] The updated result x of the motion state of the center of mass of the extended target at the last moment k-1|k-1 , the updated result P of the covariance of the extended target center of mass motion state at the last moment k-1|k-1 And the updated result X of the expansion form of the expansion target at the last moment k-1|k-1 , sent to a fully connected layer, then normalized, and sent to the long short-term memory network; the expectation of the short-term memory variable at the previous moment Covariance of the short-term memory variable at the previous moment Splicing, defined as the short-term memory c of the previous moment k-1 , sent to the long short-term memory network;
[0067] The long-term memory C of the previous momentk-1 Send it to the long short-term memory network;
[0068] Generate the current short-term memory variable c k and the current long-term memory variable C k
[0069] The short-term memory c of the current moment k Send it to two fully connected layers respectively to generate the expectation of the short-term memory variable at the current moment and the covariance of the short-term memory variable at the current moment
[0070] Furthermore, in step three, the prediction link of the algorithm combining neural network and prior knowledge is constructed, which specifically includes:
[0071] The expectation of the short-term memory variable at the current moment and the covariance of the short-term memory variable at the current moment After splicing, they are sent to three fully connected layers to generate the expected error of the target center of mass motion state evolution model at the current moment Covariance of the target center of mass motion state evolution model error at the current moment and the expected error of the target extended morphological evolution model at the current moment
[0072] The expectation x of the updated result of the target center of mass motion state at the last moment k-1|k-1 , the covariance P of the updated result of the target center of mass motion state at the last moment k-1|k-1 , the update result X of the target expansion form at the last moment k-1|k-1 , the expectation of the target center of mass motion state evolution model error at the current moment and the covariance of the target center of mass motion state evolution model error at the current moment Substitute the recursive form of the target center of mass motion state prediction link to obtain the expected prediction result x of the target center of mass motion state at the current moment k|k-1 , the covariance P of the prediction result of the target center of mass motion state at the current moment k|k-1 ;
[0073] The updated result X of the target expansion form at the last moment k-1|k-1 , the uncertainty v of the update result of the target expansion shape at the last moment k-1|k-1 and the expected error of the target extended morphological evolution model at the current moment Substitute the recursive form of the target expansion form prediction link to obtain the prediction result X of the target expansion form at the current moment k|k-1 , the uncertainty of the prediction result of the target expansion shape at the current moment v k|k-1 ;
[0074] The values of other variables in this process are set based on prior experience.
[0075] Furthermore, in step 3, the update process of the algorithm combining the neural network and prior knowledge is constructed, which specifically includes:
[0076] The expected result x of the prediction result of the target center of mass motion state at the current moment k|k-1 And the prediction result X of the target expansion shape at the current moment k|k-1 Splicing and standardization are performed, and then sent to two fully connected layers to generate the expected error of the target center of mass motion state observation model at the current moment. and the covariance of the target center of mass motion state observation model error at the current moment
[0077] The expected result x of the prediction result of the target center of mass motion state at the current moment k|k-1 , the covariance P of the prediction result of the target center of mass motion state at the current moment k|k-1 , the current target measurement set Z k , the expectation of the target center of mass motion state observation model error at the current moment and the covariance of the target center of mass motion state observation model error at the current moment Substitute the recursive form of the target center of mass motion state update link to obtain the expected update result x of the target center of mass motion state at the current moment k|k , the covariance P of the updated result of the target center of mass motion state at the current moment k|k ;
[0078] The prediction result X of the target expansion form at the current moment k|k-1 , the uncertainty of the prediction result of the target expansion shape at the current moment v k|k-1 , the current target measurement set Z k and the covariance of the target center of mass motion state observation model error at the current moment Substitute the recursive form of the target extension form update link to obtain the update result X of the target extension form at the current moment k|k , the uncertainty v of the update result of the target expansion shape at the last moment k|k ;
[0079] The values of other variables in this process are set based on prior experience.
[0080] Beneficial effects:
[0081] (1) The present invention uses a long short-term memory network to learn the motion characteristics of the target, uses long-term memory to capture the non-Markov characteristics of the target's long-term motion, and uses short-term memory to capture the speed-varying characteristics of the target's short-term motion. By comprehensively utilizing the long-term motion laws and short-term motion characteristics of the target's motion, the present invention achieves continuous high-precision tracking of the extended target over a long period of time.
[0082] (2) The present invention combines prior knowledge with long short-term memory networks and, under reasonable assumptions, derives the recursive form of the prediction and update formulas for the extended target tracking method based on a Bayesian filtering framework. Therefore, the algorithmic model of the present invention is closer to the physical model of the target in the real world than the prior model and has greater interpretability and reliability than the pure neural network model.
[0083] (3) The present invention constructs the memory link, prediction link and update link of the algorithm that combines neural network and prior knowledge, providing a reliable implementation solution for scientific research and technological development in related fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0084] Figure 1 , is an overall flow chart of the algorithm in the embodiment of the present invention;
[0085] Figure 2 , is a flow chart of the algorithm memory link in an embodiment of the present invention;
[0086] Figure 3 , is a flow chart of the algorithm prediction step in an embodiment of the present invention;
[0087] Figure 4 , is a flow chart of the algorithm update process in an embodiment of the present invention;
[0088] Figure 5 , is a diagram showing the tracking result of the straight-moving target by the algorithm in an embodiment of the present invention;
[0089] Figure 6 , is a diagram showing the tracking result of the turning target by the algorithm in an embodiment of the present invention. DETAILED DESCRIPTION
[0090] The present invention will be further described in detail below by means of specific embodiments in conjunction with the accompanying drawings. Similar variables in different embodiments are represented by associated similar variables. In the following embodiments, many detailed descriptions are provided to enable the present application to be better understood. However, those skilled in the art will readily appreciate that some of the features may be omitted in different circumstances, or may be replaced by other elements, materials, or methods. In some cases, some operations related to the present application are not shown or described in the specification. This is to avoid the core portion of the present application being overwhelmed by excessive descriptions. For those skilled in the art, it is not necessary to describe these related operations in detail. They will fully understand the related operations based on the description in the specification and the general technical knowledge in the art.
[0091] In addition, the features, operations, or characteristics described in the specification may be combined in any appropriate manner to form various embodiments. Furthermore, the steps or actions in the method description may be reordered or adjusted in a manner readily apparent to those skilled in the art. Therefore, the various sequences in the specification and drawings are provided solely for the purpose of clearly describing a particular embodiment and are not intended to be mandatory, unless otherwise specified.
[0092] The embodiment of the present invention provides an extended target tracking method based on long short-term memory network. The overall process is as follows: Figure 1 The details are as shown in Figure 2-4 The following is a detailed description.
[0093] Scenario setting: Assume that the target is an elliptical extended target on a two-dimensional plane, with a major axis length of 10 meters and a minor axis length of 2 meters. It moves at a speed of 10 meters per second for 140 seconds, and the sampling interval (seconds) △t = 1;
[0094] Establish mathematical model: Based on random matrix theory, use the target center of mass motion state x k , target center of mass motion state covariance P k 、Target expansion form X k and target extended morphological freedom v k Describe the goal, expressed as
[0095]
[0096] where x k and y k Respectively represent the coordinates of the target mass center at time k on the X-axis and Y-axis, and They represent the speed of the target mass center at time k in the X-axis direction and the Y-axis direction respectively, and They represent the acceleration of the target mass center at moment k in the X-axis direction and the Y-axis direction respectively.
[0097] P k is a 4×4 matrix;
[0098]
[0099] where θ k represents the target heading angle at time k, a k and b k Indicates the major and minor semi-axes of the target expansion at time k.
[0100] v k is a scalar;
[0101] Assume that the target center of mass motion state evolution model is
[0102]
[0103] in represents the actual target center of mass motion state evolution model, switching between the uniform velocity model and the uniform angular velocity model, switching once every 20 seconds; w k Subject to the mean μ w , the variance is P w Normal distribution;
[0104] μ w =[0,0,0,0] T
[0105]
[0106] Assume that the observation model of the target center of mass motion state is
[0107]
[0108] in represents the actual target center of mass motion state observation model, v k Subject to the mean μ v , the variance is P v Normal distribution;
[0109] μ v =[0,0] T
[0110]
[0111] Assume that the target measurement quantity follows a Poisson distribution with a mean of 20;
[0112] According to the above parameters, 600 sets of data are generated.
[0113] Step 1: Use the long short-term memory network to learn the motion characteristics of the target. The long-term memory is used to capture the non-Markov characteristics of the target's long-term motion, and the short-term memory is used to capture the speed variation characteristics of the target's short-term motion.
[0114] Introduce the short-term memory variable c at time k k and long-term memory variable C k , are defined as vectors of size 64×1, the number of hidden nodes in the linear layer is set to 64, and the variable dimensions of the input and output are set according to the input and output variables.
[0115] The updated result x of the motion state of the center of mass of the extended target at the last moment k-1|k-1 , the updated result P of the covariance of the extended target center of mass motion state at the last moment k-1|k-1 And the updated result X of the expansion form of the expansion target at the last moment k-1|k-1 , sent to a fully connected layer, then normalized, and sent to the long short-term memory network; the expectation of the short-term memory variable at the previous moment Covariance of the short-term memory variable at the previous moment Splicing, defined as the short-term memory c of the previous moment k-1 , sent to the long short-term memory network;
[0116] The long-term memory C of the previous moment k-1 Send it to the long short-term memory network;
[0117] Generate the current short-term memory variable c k and the current long-term memory variable C k
[0118] The short-term memory c of the current moment k Send it to two fully connected layers respectively to generate the expectation of the short-term memory variable at the current moment and the covariance of the short-term memory variable at the current moment
[0119] Step 2: Combining prior knowledge with the LSTM network, under reasonable assumptions, and based on the Bayesian filtering framework, derive the recursive form of the prediction and update steps of the extended target tracking method;
[0120] The algorithm adopts the prior model f k (·)for
[0121]
[0122] Prior model h k (·)for
[0123]
[0124] Substituting into the formula, we can get the formulas for center of mass state prediction, extended morphology prediction, center of mass state update and extended morphology update.
[0125] Step 3: Recursively perform the memory phase, prediction phase, and update phase to obtain the target tracking result;
[0126] The expectation of the short-term memory variable at the current moment and the covariance of the short-term memory variable at the current moment After splicing, they are sent to three fully connected layers to generate the expected error of the target center of mass motion state evolution model at the current moment Covariance of the target center of mass motion state evolution model error at the current moment and the expected error of the target extended morphological evolution model at the current moment
[0127] The expectation x of the updated result of the target center of mass motion state at the last moment k-1|k-1 , the covariance P of the updated result of the target center of mass motion state at the last moment k-1|k-1 , the update result X of the target expansion form at the last moment k-1|k-1 , the expectation of the target center of mass motion state evolution model error at the current moment and the covariance of the target center of mass motion state evolution model error at the current moment Substitute the recursive form of the target center of mass motion state prediction link to obtain the expected prediction result x of the target center of mass motion state at the current moment k|k-1 , the covariance P of the prediction result of the target center of mass motion state at the current moment k|k-1 ;
[0128] The updated result X of the target expansion form at the last moment k-1|k-1 , the uncertainty v of the update result of the target expansion shape at the last moment k-1|k-1 and the expected error of the target extended morphological evolution model at the current moment Substitute the recursive form of the target expansion form prediction link to obtain the prediction result X of the target expansion form at the current moment k|k-1 , the uncertainty of the prediction result of the target expansion shape at the current moment v k|k-1 ;
[0129] The expected result x of the prediction result of the target center of mass motion state at the current moment k|k-1 And the prediction result X of the target expansion shape at the current moment k|k-1 Splicing and standardization are performed, and then sent to two fully connected layers to generate the expected error of the target center of mass motion state observation model at the current moment. and the covariance of the target center of mass motion state observation model error at the current moment
[0130] The expected result x of the prediction result of the target center of mass motion state at the current moment k|k-1 , the covariance P of the prediction result of the target center of mass motion state at the current moment k|k-1 , the current target measurement set Z k , the expectation of the target center of mass motion state observation model error at the current moment and the covariance of the target center of mass motion state observation model error at the current moment Substitute the recursive form of the target center of mass motion state update link to obtain the expected update result x of the target center of mass motion state at the current moment k|k , the covariance P of the updated result of the target center of mass motion state at the current moment k|k ;
[0131] The prediction result X of the target expansion form at the current moment k|k-1 , the uncertainty of the prediction result of the target expansion shape at the current moment v k|k-1 , the current target measurement set Z k and the covariance of the target center of mass motion state observation model error at the current moment Substitute the recursive form of the target extension form update link to obtain the update result X of the target extension form at the current moment k|k , the uncertainty v of the update result of the target expansion shape at the last moment k|k ;
[0132] The values of other variables in this process are set based on prior experience.
[0133] Apply the extended target tracking algorithm to track the extended target in the scene and obtain the tracking result, such as Figure 5-6 shown.
[0134] Across the entire dataset, the present invention achieves a root mean square error of 0.398 meters during tracking. Existing methods based on first-order Markov models achieve a root mean square error of 1.227 meters, and existing methods using pure neural networks achieve a root mean square error of 2.704 meters. This demonstrates that the present invention considers non-Markov information and fully utilizes the neural network's ability to fit data, achieving a smaller tracking error.
[0135] The above examples are used to illustrate the present invention, which are only used to help understand the present invention and are not intended to limit the present invention. Those skilled in the art can make several simple deductions, modifications or substitutions based on the concept of the present invention.
Claims
1. An extended target tracking method based on long short-term memory network, characterized in that: The method comprises: Use the long short-term memory network to learn the motion characteristics of the target, use the long-term memory to capture the non-Markov characteristics of the target's long-term motion, and use the short-term memory to capture the speed-varying characteristics of the target's short-term motion; By combining prior knowledge with long short-term memory networks, and under reasonable assumptions, we derive the recursive form of the prediction and update steps of the extended target tracking method based on the Bayesian filtering framework. The memory link, prediction link and update link of the algorithm that combines neural network and prior knowledge are constructed, and the tracking of the extended target is achieved by recursively executing the three links.
2. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: Use the long short-term memory network to learn the target's motion characteristics, use long-term memory to capture the non-Markov characteristics of the target's long-term motion, and use short-term memory to capture the speed-varying characteristics of the target's short-term motion. Specifically, it includes: The target's motion state at multiple discrete moments is regarded as a string of time series data, and the target's motion state is recursively fed into the long short-term memory network. The long short-term memory network calculates the target's motion state based on its own parameters to obtain the target's motion characteristics at each moment. Using a long short-term memory network, while recursively receiving target motion features, it selectively discards redundant information and retains deeper target motion feature information, thereby obtaining the influence of more distant target motion features on current motion features, and then extracting non-Markov features of target motion from them; Using the long short-term memory network, the current motion features of the target are generated in the short-term memory, which is then interacted with the long-term memory to capture the random disturbances in the current motion of the target, and then the speed variation features of the target motion are extracted from it.
3. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: Combining prior knowledge with long short-term memory networks, specifically including: Assume that the target center of mass motion state evolution model is Where x represents the motion state of the target's center of mass, X represents the target's extended form, and f real (·) represents the actual target center of mass motion state evolution model, w is the process noise, subscript k represents the time, and subscript 1:k-1 represents the time from time 1 to time k-1; Since it is impossible to accurately obtain the actual target According to prior knowledge, an approximate target center of mass motion state evolution model f can be used k (·), such as the uniform linear motion model, this approximation process will introduce errors In the approximate target center of mass motion state evolution model f k (·) Target motion features obtained by superimposing the long short-term memory network Can reduce Assume that the observation model of the target center of mass motion state is where z k represents the set of measurements of the target at time k, represents the actual target center of mass motion state observation model, v k is the observation noise; Since it is impossible to accurately obtain the actual target According to prior knowledge, an approximate target center of mass motion state observation model h can be used k (·), this approximation process will introduce errors In the approximate target center of mass motion state observation model h k (·) The features related to target motion and observation obtained by superimposing the long short-term memory network Can reduce In addition, due to model approximation, the covariance of variables such as target center of mass motion and extended morphology also differs from the actual covariance, and a neural network is also required to extract relevant information from the target motion features.
4. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: Under reasonable assumptions, specifically including: Considering the tracking process at time k, we make the following reasonable assumptions that are in line with reality: Assume that the target center of mass moves x k and x k-1 and X k The probability density distribution of its evolution process is related to Approximately obey the expectation f k (x k-1 ), the variance is Q k Gaussian distribution of the target center of mass motion state, the posterior distribution p(x k |Z 1:k ,D) obey the expectation of x k|k , the variance is P k|k Gaussian distribution; Assuming the target center of mass motion state evolution model error The probability density distribution of The variance is Gaussian distribution of the target center of mass motion state observation model error The probability density distribution of The variance is Gaussian distribution of , assuming that the target extended morphological evolution model error The expectation of the probability density distribution is Assuming the target expansion form X k and X k-1 The probability density distribution of its evolution process is related to The degree of freedom parameter is δ k , the matrix parameters are The Wishart distribution of the target expansion form p(X k |Z 1:k ,D) subject to the freedom parameter v k|k , the matrix parameter is X k|k The inverse Wishart distribution of Hypothesized likelihood distribution of the measurement set For the expectation The variance is B k X k (B k ) T The product of Gaussian distributions of k Represents the set of all measurements generated by the target at time k, with a total of n k indivual; Assume that the short-term memory variable c k with x k-1 、X k-1 、c k-1 Related to D, the probability density distribution of its update link p(c k |x k-1 ,X k-1 ,c k-1 ,D) obey expectations The variance is Gaussian distribution of , where D represents the offline dataset for training and testing the neural network.
5. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: The recursive form of the prediction phase of the extended target tracking method is derived based on the Bayesian filtering framework, specifically including: The prediction model of the target center of mass motion state is: The recursive form of the target center of mass motion state covariance prediction is: Defining intermediate variables λ k-1 =v k-1|k-1 -2d-2 The recursive form of target extended morphology uncertainty prediction is: The recursive form of target expansion morphology prediction is 6. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: The recursive form of the update phase of the extended target tracking method is derived based on the Bayesian filtering framework, specifically including: The observation model of the target center of mass motion state is: The cross-covariance between the target center of mass motion state and the target center of mass motion state observation is The covariance of the target center of mass motion state observation is The recursive form of the target center of mass motion state update is: in Indicates measurement z k expectations; The recursive form of the target center of mass motion state covariance update is: Target extended form innovation covariance The recursive form of the target extended morphology uncertainty update is v k|k =v k|k-1 +n k The recursive form of the target extended morphology update is derived as follows: in Indicates measurement z k The covariance of .
7. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: Constructing the memory link of the algorithm that combines neural networks and prior knowledge, specifically including: The updated result x of the motion state of the center of mass of the extended target at the last moment k-1|k-1 , the updated result P of the covariance of the extended target center of mass motion state at the last moment k-1|k-1 And the updated result X of the expansion form of the expansion target at the last moment k-1|k-1 , sent to a fully connected layer, then normalized, and sent to the long short-term memory network; the expectation of the short-term memory variable at the previous moment Covariance of the short-term memory variable at the previous moment Splicing, defined as the short-term memory c of the previous moment k-1 , sent to the long short-term memory network; The long-term memory C of the previous moment k-1 Send it to the long short-term memory network; Generate the current short-term memory variable c k and the current long-term memory variable C k The short-term memory c of the current moment k Send it to two fully connected layers respectively to generate the expectation of the short-term memory variable at the current moment and the covariance of the short-term memory variable at the current moment 8. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: Constructing the prediction phase of the algorithm that combines neural networks and prior knowledge, specifically including: The expectation of the short-term memory variable at the current moment and the covariance of the short-term memory variable at the current moment After splicing, they are sent to three fully connected layers to generate the expected error of the target center of mass motion state evolution model at the current moment Covariance of the target center of mass motion state evolution model error at the current moment and the expected error of the target extended morphological evolution model at the current moment The expectation x of the updated result of the target center of mass motion state at the last moment k-1|k-1 , the covariance P of the updated result of the target center of mass motion state at the last moment k-1|k-1 , the update result X of the target expansion form at the last moment k-1|k-1 , the expectation of the target center of mass motion state evolution model error at the current moment and the covariance of the target center of mass motion state evolution model error at the current moment Substitute the recursive form of the target center of mass motion state prediction link to obtain the expected prediction result x of the target center of mass motion state at the current moment k|k-1 , the covariance P of the prediction result of the target center of mass motion state at the current moment k|k-1 ; The updated result X of the target expansion form at the last moment k-1|k-1 , the uncertainty v of the update result of the target expansion shape at the last moment k-1|k-1 and the expected error of the target extended morphological evolution model at the current moment Substitute the recursive form of the target expansion form prediction link to obtain the prediction result X of the target expansion form at the current moment k|k-1 , the uncertainty of the prediction result of the target expansion shape at the current moment v k|k-1 ; The values of other variables in this process are set based on prior experience.
9. The extended target tracking method based on long short-term memory network according to claim 1, characterized in that: The update process of building an algorithm that combines neural networks and prior knowledge includes: The expected result x of the prediction result of the target center of mass motion state at the current moment k|k-1 And the prediction result X of the target expansion shape at the current moment k|k-1 Splicing and standardization are performed, and then sent to two fully connected layers to generate the expected error of the target center of mass motion state observation model at the current moment. and the covariance of the target center of mass motion state observation model error at the current moment The expected result x of the prediction result of the target center of mass motion state at the current moment k|k-1 , the covariance P of the prediction result of the target center of mass motion state at the current moment k|k-1 , the current target measurement set Z k , the expectation of the target center of mass motion state observation model error at the current moment and the covariance of the target center of mass motion state observation model error at the current moment Substitute the recursive form of the target center of mass motion state update link to obtain the expected update result x of the target center of mass motion state at the current moment k|k , the covariance P of the updated result of the target center of mass motion state at the current moment k|k ; The prediction result X of the target expansion form at the current moment k|k-1 , the uncertainty of the prediction result of the target expansion shape at the current moment v k|k-1 , the current target measurement set Z k and the covariance of the target center of mass motion state observation model error at the current moment Substitute the recursive form of the target extension form update link to obtain the update result X of the target extension form at the current moment k|k , the uncertainty v of the update result of the target expansion shape at the last moment k|k ; The values of other variables in this process are set based on prior experience.