LSTM-KF hybrid tracking method based on parameter adaptation

Abstract

The invention relates to the technical field of target tracking, and discloses an LSTM-KF hybrid tracking method based on parameter adaptation, which comprises the following steps: acquiring target position measurement data of a moving target, and extracting dual-channel characteristics of the target position measurement data through a sliding window; inputting the dual-channel features into a pre-trained dual-channel LSTM network to obtain a target position prediction result and a target speed prediction result; analyzing target motion characteristics based on a prediction result, selecting a corresponding dynamic model according to a motion mode, and performing target state estimation on the dynamic model through a Kalman filter; and finally, optimizing noise parameters of the Kalman filter by adopting an expectation maximization algorithm, carrying out weighted fusion on a prediction result of the dual-channel LSTM network and a target state estimation result, and outputting a final target state. According to the invention, high-precision and high-robustness target tracking in a complex environment is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target tracking, and particularly relates to a LSTM-KF hybrid tracking method based on parameter adaptation. BACKGROUND

[0002] In the field of target tracking, especially in key application scenarios such as military air defense, air traffic control and unmanned aerial vehicle monitoring, the demand for accurate tracking of targets is increasingly urgent.

[0003] Traditional target tracking methods, such as Kalman filter (KF), extended Kalman filter (EKF) and particle filter (PF), can provide reliable state estimation when the target motion mode is relatively stable and the system noise characteristics are determined. However, in actual applications, the target motion mode is complex and variable, and the noise characteristics are difficult to accurately obtain in advance, which leads to a significant decrease in tracking accuracy. In recent years, although the long short-term memory network (LSTM) based on deep learning can learn the target motion mode from historical data, the data-driven characteristics make it unstable in the presence of large noise interference or sparse data scenes, and it lacks the constraint ability of the physical model. For example, when electromagnetic interference causes measurement data to be abnormal, the long short-term memory network (LSTM) is prone to prediction deviation due to overfitting of noise, and cannot guarantee the reliability of the tracking results.

[0004] It can be seen that the traditional target tracking method has the problems of insufficient tracking accuracy and poor adaptability in complex environments. SUMMARY

[0005] Therefore, the present application provides a LSTM-KF hybrid tracking method based on parameter adaptation to solve the problem of insufficient tracking accuracy and poor adaptability of the traditional target tracking method in complex environments.

[0006] In a first aspect, the present application provides a LSTM-KF hybrid tracking method based on parameter adaptation, which comprises:

[0007] obtaining target position measurement data of a moving target;

[0008] extracting double-channel features of the target position measurement data through a sliding window;

[0009] inputting the double-channel features into a pre-trained double-channel LSTM network to obtain target position prediction results and target speed prediction results;

[0010] analyzing target motion characteristics based on the target position prediction results and the target speed prediction results to obtain a motion mode of the moving target;

[0011] According to the motion mode, a corresponding dynamic model is selected, and a target state of the dynamic model is estimated by a Kalman filter;

[0012] A noise parameter of the Kalman filter is optimized by using an expectation maximization algorithm;

[0013] The prediction result of the dual-channel LSTM network and the result of the target state estimation are weighted and fused to output a final target state.

[0014] In an optional implementation, before the dual-channel features of the target position measurement data are extracted by the sliding window, the method further comprises:

[0015] The target position measurement data is preprocessed by removing outliers and standardizing formats to obtain a target position measurement sequence.

[0016] In an optional implementation, the dual channel includes a coordinate feature channel and a speed feature channel;

[0017] The coordinate feature channel includes a position sequence, a displacement sequence, an average speed sequence, and an instantaneous speed sequence;

[0018] The speed feature channel includes a speed evolution sequence of consecutive time steps.

[0019] In an optional implementation, the dual-channel LSTM network includes a coordinate feature processing channel, a speed feature processing channel, and a feature fusion layer;

[0020] The coordinate feature processing channel is a two-layer cascaded LSTM unit, the speed feature processing channel is a single-layer LSTM unit, and the feature fusion layer is configured to integrate the output of the coordinate feature processing channel and the output of the speed feature processing channel, and map the integrated features to a target position prediction result and a target speed prediction result.

[0021] In an optional implementation, the motion mode of the moving target is determined based on the target position prediction result and the target speed prediction result, and includes:

[0022] The acceleration and the change rate of the heading of the moving target between adjacent time frames are calculated based on the target position prediction result and the target speed prediction result.

[0023] Based on a preset acceleration threshold and a preset angular velocity threshold, the acceleration and the change rate of the heading are classified into a motion mode to determine the motion mode of the moving target; the motion mode includes a uniform motion mode, a uniform acceleration motion mode, or a turning motion mode.

[0024] In an alternative embodiment, the step of optimizing the noise parameters of the Kalman filter using the expectation-maximization algorithm comprises iteratively performing the following steps until the current noise parameter variation is less than a threshold or a maximum number of iterations is reached:

[0025] performing Kalman filtering and backward smoothing on the target position measurement data within the sliding window based on the current noise parameters to obtain smoothed state estimates;

[0026] updating the process noise covariance matrix and the measurement noise covariance matrix of the Kalman filter based on the smoothed state estimates.

[0027] In an alternative embodiment, the step of fusing the prediction results of the dual-channel LSTM network and the results of the target state estimation with a weighting to output a final target state comprises:

[0028] calculating a norm of a Kalman filter innovation sequence based on the results of the target state estimation; the Kalman filter innovation sequence being a difference between a measurement value at a current time and a prior state estimate of the Kalman filter at the current time;

[0029] determining a weight of the prediction results of the dual-channel LSTM network based on the norm of the Kalman filter innovation sequence and a preset weight adjustment strategy;

[0030] fusing the prediction results of the dual-channel LSTM network and the results of the target state estimation with the weighting according to the weight to output a final target state.

[0031] In a second aspect, the present application provides a parameter-adaptive LSTM-KF hybrid tracking device, the device comprising:

[0032] a target position measurement data acquisition module configured to acquire target position measurement data of a moving target;

[0033] a dual-channel feature extraction module configured to extract dual-channel features of the target position measurement data through a sliding window;

[0034] a position and velocity prediction module configured to input the dual-channel features into a pre-trained dual-channel LSTM network to obtain target position prediction results and target velocity prediction results;

[0035] a motion pattern acquisition module configured to analyze target motion characteristics based on the target position prediction results and the target velocity prediction results to obtain a motion pattern of the moving target;

[0036] a target state estimation module configured to select a corresponding dynamic model according to the motion pattern and perform target state estimation on the dynamic model through a Kalman filter.

[0037] a noise parameter optimization module configured to optimize noise parameters of the Kalman filter using an expectation maximization algorithm;

[0038] a final target state output module configured to perform weighted fusion on a prediction result of the dual-channel LSTM network and a result of the target state estimation, and output a final target state.

[0039] In a third aspect, the present application provides a computer device, comprising a memory and a processor, the memory and the processor are connected with each other in communication, the memory stores computer instructions, and the processor executes the computer instructions to perform the parameter adaptive based LSTM-KF hybrid tracking method of the first aspect or any of the corresponding embodiments.

[0040] In a fourth aspect, the present application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are used to make a computer execute the parameter adaptive based LSTM-KF hybrid tracking method of the first aspect or any of the corresponding embodiments.

[0041] In a fifth aspect, the present application provides a computer program product, which comprises computer instructions, and the computer instructions are used to make a computer execute the parameter adaptive based LSTM-KF hybrid tracking method of the first aspect or any of the corresponding embodiments.

[0042] The technical solution provided by the present application can include the following beneficial effects:

[0043] Firstly, by acquiring the position measurement data of the moving target and extracting the dual-channel features using a sliding window, more comprehensive and accurate input information is provided for subsequent target state prediction. The dual-channel features can fully capture the position and speed changes of the target, and compared with single feature input, can better reflect the motion state of the target and improve the prediction accuracy. Secondly, the dual-channel features are input into the pre-trained dual-channel LSTM network to obtain the prediction results of the target position and speed. The dual-channel LSTM network structure is adopted to process the coordinate features and speed features respectively, fully utilize the multi-dimensional information of the target, and improve the accuracy of trajectory prediction. Then, the target motion characteristics are analyzed based on the target position measurement data and the target speed prediction results to obtain the motion mode, and the corresponding motion model is selected according to the motion mode, which can automatically adapt to different types of target motion and enhance the universality of the system. The target state estimation is realized through the Kalman filter, which realizes the adaptive processing of different motion modes and can accurately estimate the state according to the characteristics of different motion states, improving the flexibility and adaptability of the tracking system. At the same time, the expectation maximization algorithm is used to optimize the noise parameters of the Kalman filter, reducing the dependence on prior knowledge and improving the adaptive ability of the system. Finally, through the adaptive fusion mechanism, the prediction results of the dual-channel LSTM network and the target state estimation results are weighted and fused to output the final target state, combining the advantages of the two methods to improve the tracking stability. The present application effectively solves the problems of insufficient tracking accuracy and poor adaptability of traditional target tracking methods in complex environments, and realizes high-precision and high-robustness target tracking in complex environments. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the specific embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0045] Figure 1 is a structural schematic diagram of a parameter adaptive LSTM-KF hybrid tracking system according to an embodiment of the present application;

[0046] Figure 2 is a flowchart of a parameter adaptive LSTM-KF hybrid tracking method according to an embodiment of the present application;

[0047] Figure 3 is a flowchart of another parameter adaptive LSTM-KF hybrid tracking method according to an embodiment of the present application;

[0048] Figure 4is a double-channel feature extraction schematic diagram according to an embodiment of the present application;

[0049] Figure 5 is a structure schematic diagram of a double-channel LSTM network according to an embodiment of the present application;

[0050] Figure 6 is an optimization flow schematic diagram of an expectation maximization algorithm according to an embodiment of the present application;

[0051] Figure 7 is a positioning error comparison schematic diagram of different tracking methods in various motion scenes according to an embodiment of the present application;

[0052] Figure 8 is a structure block diagram of a parameter adaptive based LSTM-KF hybrid tracking device according to an embodiment of the present application.

[0053] Figure 9 is a hardware structure schematic diagram of a computer device according to an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0055] The present application relates to the technical field of target tracking, and in particular to state estimation and trajectory prediction of aerial targets (such as unmanned aerial vehicles, birds, etc.). In modern air defense early warning systems, accurate tracking of maneuvering targets is directly related to situation awareness and command decision-making. The present application aims to provide a parameter adaptive based LSTM-KF hybrid tracking method, which combines the advantages of deep learning and traditional filtering technology, to achieve higher precision and stronger robustness of target tracking performance, and to achieve high-precision and high-robustness target tracking in complex environments.

[0056] According to an embodiment of the present application, a parameter adaptive based LSTM-KF hybrid tracking system embodiment is provided, Figure 1 is a structure schematic diagram of a parameter adaptive based LSTM-KF hybrid tracking system according to an embodiment of the present application, as Figure 1 shown, the system comprises a data receiving module 101, a feature extraction module 102, a double-channel LSTM model 103, a motion pattern recognition module 104, an adaptive Kalman filter 105, an EM parameter optimization module 106 and an adaptive fusion module 107.

[0057] The data receiving module 101 is responsible for receiving target position measurement data collected by the radar, performing outlier detection and data standardization. The feature extraction module 102 extracts double-channel features from the measurement sequence corresponding to the target position measurement data. The double-channel LSTM model 103 performs prediction based on the double-channel features extracted by the feature extraction module 102 to obtain target position prediction results and target speed prediction results; the motion pattern recognition module 104 analyzes target motion characteristics based on the target position prediction results and the target speed prediction results to obtain a motion pattern; the adaptive Kalman filter 105 performs target state estimation according to the motion pattern recognized by the motion pattern recognition module 104; the EM parameter optimization module 106 optimizes the noise parameters of the Kalman filter; and the adaptive fusion module 107 intelligently fuses the results of each module to improve tracking performance through mutual cooperation.

[0058] According to the embodiment of the present application, a parameter adaptive based LSTM-KF hybrid tracking method is provided. It should be noted that the steps shown in the flowchart can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0059] In this embodiment, a parameter adaptive based LSTM-KF hybrid tracking method is provided, Figure 2 is a flowchart of a parameter adaptive based LSTM-KF hybrid tracking method according to the embodiment of the present application, as Figure 2 shown, the flow includes the following steps:

[0060] Step S201, obtaining target position measurement data of a moving target.

[0061] Further, step S201 is completed by a radar sensor, which periodically scans the monitoring area and captures the position information of the target. The obtained target position measurement data may contain certain errors and noise, such as deviations caused by environmental interference or precision limitations of the sensor itself. These raw data are continuously collected at certain time intervals to form time series data, providing basic input for subsequent target tracking processing.

[0062] Step S202, extracting double-channel features of the target position measurement data through a sliding window.

[0063] Further, the sliding window is a data processing technique that moves over time series data, covering a certain number of consecutive data points each time. The dual-channel feature extraction means that two different sets of features are obtained from the data. For example, the coordinate feature channel can include position coordinates, relative displacement, and other information that can depict the trajectory of the position change of the target in space; the speed feature channel focuses on describing the speed change of the target, such as the size and direction change of the speed. In this way, step S202 can more comprehensively depict the motion characteristics of the target.

[0064] The embodiment directly processes the target position measurement data obtained in step S201 and extracts time sequence features using a sliding window mechanism. For example, the sliding window length is set to 8 time steps (corresponding to 0.8 seconds of physical time, matching the 10 Hz scanning frequency of the radar sensor), and each time it is moved forward by 1 time step. For the position sequence in each window, two types of complementary features are extracted: the coordinate feature channel includes position coordinates, relative displacement (relative to the first point of the window), average speed, and instantaneous speed; the speed feature channel processes the speed evolution sequence of the consecutive 7 time steps. This dual-channel design simultaneously captures the position and speed change characteristics of the target, and is particularly suitable for describing the complex trajectory characteristics of high-maneuvering targets.

[0065] Step S203 inputs the dual-channel features into the pre-trained dual-channel LSTM network to obtain the target position prediction result and the target speed prediction result.

[0066] Further, the dual-channel LSTM network is a recurrent neural network that can process time series data and capture long-term dependencies therein. The pre-trained dual-channel LSTM network has learned the general rules of moving targets through a large amount of historical data. The extracted dual-channel features are input into the network, and the network will predict the future target position and speed based on the knowledge learned before.

[0067] The two-channel features extracted in step S203 are respectively input into a specially designed two-channel LSTM network. For example, the coordinate feature channel adopts two-layer cascaded LSTM units (the hidden layer dimension is 64, which matches the input feature dimension 8), and layer normalization is applied to enhance the training stability; the speed feature channel adopts a single-layer LSTM structure (the hidden layer dimension is 32). Compared with the traditional LSTM structure, the embodiment introduces three key improvements: (1) a forgetting gate bias initialization technique, which sets the bias value to 1.0 to alleviate the gradient disappearance problem in long sequence training; (2) an orthogonal weight initialization strategy (using the Glorot orthogonal initialization method), which significantly improves the network convergence speed and stability; (3) a residual connection structure, which adds a skip connection in the feature fusion layer to ensure that the deep network can effectively use the shallow features. The outputs of the two branches are integrated through a feature fusion layer (containing a fully connected layer of 256 neurons), and finally mapped to a 4-dimensional output space to represent the target future position and speed prediction values. When processing the trajectory of a highly maneuverable target, the prediction accuracy of the network is improved by about 25.5% compared with the standard LSTM structure, especially in complex motion modes such as target turning and acceleration.

[0068] In step S204, the target motion characteristics are analyzed based on the target position prediction result and the target speed prediction result, and the motion mode of the moving target is obtained.

[0069] Further, step S204 analyzes the target position prediction result and the target speed prediction result obtained by prediction to determine the current motion mode of the target. For example, if the speed prediction result is relatively stable and the direction change is small, it can be determined that the target is in a uniform straight-line motion mode; if the speed is gradually increasing, it can be in a uniform acceleration motion mode; and when the speed direction changes significantly, it can be in a turning motion mode.

[0070] The embodiment calculates the speed vector change rate and the heading angle change rate of the moving target between adjacent time frames based on the measurement sequence in the sliding window. Specifically, the acceleration of the target can be calculated by the central difference method, and then the current motion mode is classified into uniform speed, uniform acceleration or turning mode based on the acceleration threshold and the angular velocity threshold.

[0071] In step S205, a corresponding dynamic model is selected according to the motion mode, and a Kalman filter is used to estimate the target state of the dynamic model.

[0072] Further, the dynamic model describes the change relationship between the target motion states, for example, the uniform motion model assumes that the target keeps the speed unchanged in a short time, and the uniform acceleration motion model considers the influence of acceleration. After selecting the appropriate dynamic model according to the previously identified motion mode, the Kalman filter is used to estimate the target state. The Kalman filter is a recursive filtering algorithm based on the prediction-correction principle, which can combine the prediction results of the dynamic model and the actual measurement data to optimally estimate the state of the target (such as position, speed, etc.), and filter out the noise in the measurement data, thereby improving the estimation accuracy.

[0073] In step S206, the expectation maximization algorithm is used to optimize the noise parameters of the Kalman filter.

[0074] Further, in the Kalman filtering process, the noise parameters (including the process noise covariance matrix and the measurement noise covariance matrix) have an important influence on the filtering effect. The expectation maximization (EM) algorithm is an iterative optimization algorithm used to estimate the parameters of a statistical model in the presence of hidden variables or incomplete data. Through the expectation maximization (EM) algorithm, the noise parameters of the Kalman filter can be adaptively adjusted according to the historical data and the current estimated state, so as to make them more consistent with the actual measurement and motion, thereby further improving the state estimation accuracy of the Kalman filter.

[0075] The embodiment is based on the historical data stored in the sliding window, and applies the expectation maximization algorithm to iteratively optimize the noise parameters of the Kalman filter. In the E step, the system performs Kalman smoothing using the current noise parameters to obtain the posterior distribution of the state; in the M step, the process noise covariance matrix and the measurement noise covariance matrix are updated based on the smoothed state; this step is closely related to the dynamic model, and the EM algorithm can adaptively adjust the noise parameters of the corresponding dynamic model according to the actual motion characteristics of the target, so that the Kalman filter can maintain optimal performance under different motion modes, solving the problem that the noise parameters in the traditional Kalman filter are difficult to accurately set, and significantly improving the tracking accuracy of the system in the target maneuvering process.

[0076] In step S207, the prediction results of the dual-channel LSTM network are weighted and fused with the estimated results of the target state, and the final target state is output.

[0077] Further, since the prediction results of the dual-channel LSTM network and the state estimation results of the Kalman filter each have certain advantages and limitations, for example, the dual-channel LSTM network may perform better in processing complex nonlinear motion patterns, while the Kalman filter may be more advantageous in processing linear motion and noise filtering, therefore, by means of weighted fusion, the results of the two methods are integrated. According to the judgment of the current target motion situation and the data reliability, appropriate weights are given to each result, and then they are weighted averaged or other fusion operations, so that more accurate and reliable final target state estimation results can be obtained.

[0078] In different scenarios, the embodiment will automatically adjust the weight distribution. When the target is in a stable motion state and the measurement quality is good, the weight of the prediction result of the dual-channel LSTM network is reduced to maintain stability; when the target performs complex maneuvers or the measurement condition deteriorates, the weight of the prediction result of the dual-channel LSTM network is increased to improve the prediction performance, so that the system has both stability and flexibility. Especially in the scene where the target frequently changes the motion mode, the tracking performance is better than that of the pure LSTM and KF methods.

[0079] In summary, compared with the traditional Kalman filtering method, the LSTM-KF hybrid tracking method of the embodiment reduces the position estimation RMSE by about 13% on average, and improves the speed estimation accuracy by about 40% on average, especially in scenes where the target motion mode changes frequently. First, the embodiment obtains the position measurement data of the moving target and extracts the dual-channel features using a sliding window, providing more comprehensive and accurate input information for subsequent target state prediction. The dual-channel features can fully capture the position and speed changes of the target, and compared with single feature input, they can better reflect the motion state of the target and improve the prediction accuracy. Second, the dual-channel features are input into the pre-trained dual-channel LSTM network to obtain the prediction results of the target position and speed. The dual-channel LSTM network structure processes the coordinate features and speed features respectively, fully utilizes the multi-dimensional information of the target, and improves the accuracy of trajectory prediction. Then, the target motion characteristics are analyzed based on the target position measurement data and the target speed prediction results to obtain the motion mode, and the corresponding motion model is selected according to the motion mode, which can automatically adapt to different types of target motion and enhance the universality of the system. The Kalman filter is used for target state estimation to realize adaptive processing of different motion modes, which can accurately estimate the state according to the characteristics of different motion states and improve the flexibility and adaptability of the tracking system. At the same time, the expectation maximization algorithm is used to optimize the noise parameters of the Kalman filter, reducing the dependence on prior knowledge and improving the adaptive ability of the system. Finally, the prediction results of the dual-channel LSTM network and the target state estimation results are weighted and fused through the adaptive fusion mechanism to output the final target state, combining the advantages of the two methods to improve the tracking stability. The present application effectively solves the problems of insufficient tracking accuracy and poor adaptability of traditional target tracking methods in complex environments, and realizes high-precision and high-robustness target tracking in complex environments.

[0080] In the present embodiment, another LSTM-KF hybrid tracking method based on parameter adaptation is provided, Figure 3 is a flowchart of another LSTM-KF hybrid tracking method based on parameter adaptation according to an embodiment of the present application, as Figure 3 shown, the flow includes the following steps:

[0081] Step S301, obtaining target position measurement data of a moving target.

[0082] Further, the embodiment uses radar sensors to collect target position measurement data about moving targets. The target position measurement data typically contains coordinate information of moving targets, such as longitude, latitude and altitude, and possibly other dynamic attributes such as speed and direction. The target position measurement data is the basic input of the target tracking system for subsequent processing and analysis. Moving targets refer to objects that need to be monitored, identified and tracked in the target tracking system, which can be various types of moving objects depending on the application scenario. For example, in the field of civil aviation, moving targets refer to aircrafts for air traffic control and route monitoring to ensure flight safety. In the intelligent traffic monitoring system, moving targets refer to cars, trucks and other ground vehicles for traffic flow analysis and accident prevention. In the maritime monitoring system, moving targets refer to ships and vessels for maritime safety monitoring and navigation.

[0083] In step S302, the target position measurement data is pre-processed by outlier filtering and format standardization to obtain a target position measurement sequence.

[0084] Further, the embodiment cleans and normalizes the original measurement data (i.e. target position measurement data) to improve data quality. Outlier filtering refers to identifying and excluding abnormal data points caused by radar sensor errors or external interference. Format standardization converts the original measurement data (i.e. target position measurement data) into a unified format or dimension, ensuring data consistency and comparability, laying a foundation for subsequent feature extraction and analysis.

[0085] The radar sensor of the embodiment can be a frequency-modulated continuous wave radar sensor, which can be fixedly installed at a monitoring area high point with a scanning frequency of 10 Hz. After the system receives the target echo signal, the polar coordinate measurement data (distance, azimuth angle, elevation angle) of the target position measurement data is converted into position data in the Cartesian coordinate system, and pre-processing is also performed, including outlier filtering and format standardization. Outlier filtering uses a multiple testing mechanism to calculate the Mahalanobis distance of each measurement point from the predicted position, and if it exceeds a pre-set threshold (typical value is 2.5), it is marked as an outlier; at the same time, a speed constraint test is performed, and if the speed calculated from adjacent measurement points exceeds the physically feasible range (such as 50 m / s), it is considered to be subject to electronic interference. Format standardization includes converting all measurement data into a unified time series format to adapt to subsequent processing units, significantly improving the anti-interference ability of the system in complex electromagnetic environments.

[0086] The target position measurement data after format standardization can be obtained by the following formula:

[0087]

[0088] where X is the original data, which can be position coordinates, velocity, acceleration, etc. target position measurement data, X norm is the normalized target position measurement data, ranging from [-1, 1], X min and X max are the minimum and maximum values of the original data, respectively.

[0089] Step S303: Extract the dual-channel features of the target position measurement data through a sliding window.

[0090] In an optional embodiment, the dual-channel includes a coordinate feature channel and a velocity feature channel;

[0091] The coordinate feature channel includes a position sequence, a displacement sequence, an average velocity sequence, and an instantaneous velocity sequence.

[0092] The velocity feature channel includes a velocity evolution sequence of consecutive time steps.

[0093] Further, the embodiment extracts dual-channel features from the preprocessed target position measurement sequence using a sliding window technique. The sliding window is a time series analysis method that slides over the data sequence to extract consecutive data segments of fixed length. Dual-channel features refer to two groups of features extracted from these data segments: coordinate features and velocity features. The coordinate features can include position coordinates, relative displacement, average velocity, and instantaneous velocity, while the velocity features can include a velocity change sequence of consecutive time steps.

[0094] The embodiment extracts time series features based on the preprocessed target position measurement sequence using a sliding window mechanism to achieve the extraction of dual-channel features (including a coordinate feature channel and a velocity feature channel), please refer to Figure 4The illustrated double-channel feature extraction schematic diagram first performs data receiving 401 and preprocessing 402. The radar sensor collects polar coordinate measurement data (distance, azimuth angle, pitch angle), which is converted into position coordinates (i.e. target position measurement data) in the Cartesian coordinate system. After that, multiple tests are performed to remove abnormal points (Mahalanobis distance threshold 2.5, speed constraint 50 m / s), and the data is standardized by minimum-maximum to form a time series format in the interval [-1, 1], forming a preprocessed target position measurement sequence. Coordinate feature parallel extraction 403 extracts the position sequence in each 8-time-step sliding window, and sequentially performs: extracting the three-dimensional position coordinates of each time step 403a, calculating the relative displacement relative to the first point of the window 403b, calculating the average speed by the mean value of the position difference in the window 403c, calculating the instantaneous speed by the difference between adjacent time steps 403d, and combining to form an 8x8-dimensional feature matrix 403e, which integrates the spatiotemporal features of position evolution. Then, velocity feature parallel extraction 404 calculates the velocity vector sequence of 7 consecutive time steps by the difference between adjacent positions 404a, and forms a 7x2-dimensional feature matrix 404b after dimension standardization, focusing on the dynamic characteristics of speed changes. Next, feature standardization 405 and output 406 perform layer normalization on the two features respectively to eliminate the dimension effect, and finally output the double-channel features containing position trajectory and speed pattern. Among them, the sliding window slides with a step size of 0.1 seconds (corresponding to a radar scanning frequency of 10 Hz), ensuring the time sequence continuity of feature extraction. This design decouples the position evolution (coordinate feature channel) and the speed dynamics (velocity feature channel), constructing an 8x8+7x2=78-dimensional feature space, which not only preserves the absolute reference information of the position coordinates, but also captures the acceleration, heading change and other maneuvering features through the velocity sequence. Compared with a single feature channel, the double-channel design reduces the trajectory modeling error of high-maneuvering targets (such as turning and rapid acceleration) by 22.3%, especially when the target motion mode changes, it can more accurately capture the changes in dynamic characteristics.

[0095] In step S304, the double-channel features are input into a pre-trained double-channel LSTM network to obtain target position prediction results and target velocity prediction results.

[0096] In an optional embodiment, the double-channel LSTM network includes a coordinate feature processing channel, a velocity feature processing channel, and a feature fusion layer.

[0097] The coordinate feature processing channel is a two-layer cascaded LSTM unit, the velocity feature processing channel is a single-layer LSTM unit, and the feature fusion layer is used to integrate the output of the coordinate feature processing channel and the output of the velocity feature processing channel, and map the integrated features to target position prediction results and target velocity prediction results.

[0098] Further, the LSTM network is a special kind of recurrent neural network (RNN) that can learn and predict time series data. The dual-channel LSTM network has two parallel input channels, which process the coordinate features and velocity features, respectively. Each channel consists of one or more LSTM layers that can capture long-term dependencies in the data. The feature fusion layer combines the outputs of the two channels to generate the future position and velocity predictions of the target.

[0099] This embodiment inputs the dual-channel features into a specially designed dual-channel LSTM network to generate the future state predictions of the target. Please refer to Figure 5 The structure diagram of the dual-channel LSTM network is shown in the figure. The dual-channel LSTM network adopts a dual-channel parallel structure. The dual-channel LSTM network structure includes a coordinate feature processing channel 501 and a velocity feature processing channel 502, which are two parallel branches. The coordinate feature processing channel 501 is composed of the following components: an input layer 501a for receiving a coordinate feature matrix with a dimension of 8x8; a first LSTM layer 501b with 64 hidden units; a first Layernorm layer 501c for normalizing the output of the first LSTM layer 501b; a second LSTM layer 501d with 64 hidden units; a second Layernorm layer 501e for normalizing the output of the second LSTM layer 501d; and a Dropout (0.2) 501f for regularizing the output of the second Layernorm layer 501e, where 0.2 represents randomly discarding 20% of the neurons. The velocity feature processing channel 502 is composed of the following components: an input layer 502a for receiving a velocity feature matrix with a dimension of 7x2; and a third LSTM layer 502b with 32 hidden units. The feature fusion layer 503 is composed of the following components: a feature concatenation unit 503a for concatenating the outputs of the two channels into a 96-dimensional vector; a fully connected layer 503b for mapping the 96-dimensional features to a 256-dimensional feature space; a ResidualBlock 503c, which is a residual block layer, containing two fully connected layers and a skip connection; and an output layer 504 for mapping the fused features to a four-dimensional output representing the predicted position and velocity.

[0100] The coordinate feature channel is processed by the following formula:

[0101]

[0102] where LSTM coord represents the LSTM layer of the coordinate feature channel, represents the hidden state output of the coordinate feature channel at time t, represents the memory cell state of the coordinate feature channel at time t, represents the hidden state output of the coordinate feature channel at time t-1, denotes the memory cell state of the coordinate feature channel at time t-1, denotes the coordinate feature input at time t, t and t-1 represent two adjacent times respectively, θ coord denotes the network parameter set of the coordinate LSTM.

[0103]

[0104] wherein, denotes the coordinate LSTM final hidden state after layer normalization processing, LayerNorm denotes the layer normalization processing of the coordinate feature channel, denotes the final hidden state after the LSTM layer.

[0105] The velocity feature channel is processed by the following formula:

[0106]

[0107] wherein, LSTM vel denotes the LSTM layer of the velocity feature channel, denotes the velocity feature input at time t, denotes the hidden state output of the velocity feature channel at time t, denotes the memory cell state of the velocity feature channel at time t, denotes the hidden state output of the velocity feature channel at time t-1, denotes the memory cell state of the velocity feature channel at time t-1, t and t-1 represent two adjacent times respectively, θ vel denotes the network parameter set of the velocity feature channel.

[0108] The coordinate feature channel and the velocity feature channel are processed by the following formula:

[0109]

[0110] z = W fusion h fusion +b fusion ;

[0111] wherein, denotes the final hidden state of the velocity feature channel, W fusion denotes the feature fusion layer weight matrix, b fusion denotes the feature fusion layer bias vector; h fusion denotes the spliced fusion feature, Concat denotes the splicing operation, which splices two feature vectors according to the dimension; Z denotes the output after linear transformation of the fusion feature.

[0112] The residual connection processing is performed by the following formula:

[0113] z res = z + ReLU (W2ReLU (W1z + b1) + b2) ;

[0114] wherein, Z res represents the residual connection processing, b1 and b2 are bias vectors, and W1 and W2 represent weight matrices of the first layer and the second layer linear transformation respectively.

[0115] The final output prediction result is obtained by the following formula:

[0116]

[0117] wherein, W out represents the weight matrix of the output layer, b out represents the bias vector of the output layer, represents the final output prediction result.

[0118] In this embodiment, the forgetting gate bias is set to 1.0 to enhance the long sequence history information retention capability and solve the long-time dependence memory decay problem in trajectory prediction; the network weight is initialized by the Glorot orthogonal method to make the gradient propagation amplitude of each layer consistent and improve the training convergence speed; the skip connection is added to the feature fusion layer to ensure that the deep network can directly reuse the shallow features and improve the prediction accuracy in the target high maneuvering scene.

[0119] In step S305, the target motion feature is analyzed based on the target position prediction result and the target speed prediction result to obtain the motion mode of the moving target.

[0120] In an optional implementation, step S305 includes:

[0121] The target motion feature is analyzed based on the target position prediction result and the target speed prediction result to obtain the motion mode of the moving target, including:

[0122] The acceleration and the heading rate of change of the moving target between adjacent time frames are calculated based on the target position prediction result and the target speed prediction result.

[0123] Based on a preset acceleration threshold and a preset angular velocity threshold, the motion mode of the moving target is classified based on the acceleration and the heading rate of change to determine the motion mode of the moving target; the motion mode includes a uniform motion mode, a uniform acceleration motion mode, or a turning motion mode.

[0124] Further, the embodiment utilizes the prediction results provided by the LSTM network to calculate the acceleration and the heading rate of change between adjacent time frames. Then, according to a preset threshold, the calculation results are classified into uniform motion, uniform acceleration motion or turning motion mode. These modes help to understand the behavior of the target and provide the basis for selecting an appropriate tracking strategy. The embodiment identifies the current motion mode by analyzing the target position prediction result and the target speed prediction result. The system calculates the acceleration of the target based on the speed sequence in the sliding window, that is, the acceleration is obtained by the following formula:

[0125]

[0126] The heading rate of change is obtained by the following formula:

[0127] Δθ = |arctan2(v y,t ,v x,t )-arctan2(v y,t-1 ,v x,t-1 )| / Δt;

[0128] wherein, Δt represents the time interval; v y,t represents the speed of the y-axis at time t; v x,t represents the speed of the x-axis at time t; v y,t-1 represents the speed of the y-axis at time t-1; v x,t-1 represents the speed of the x-axis at time t-1; and Δθ represents the heading rate of change.

[0129] According to the set acceleration threshold α threshold (typically 1.5 m / s 2 ) and the heading change threshold θ threshold (typically 5° / s), the target motion mode is classified into uniform motion (CV), uniform acceleration motion (CA) or turning motion (CT).

[0130] To avoid frequent mode switching caused by noise, the embodiment introduces a mode confidence recursive updating mechanism:

[0131] c mode,k = α·c mode,k-1 +(1-α)·I mode (a k ,Δθ k );

[0132] wherein, α is a smoothing factor (typically 0.7); I mode represents a mode indication function, and mode switching is triggered only when the confidence exceeds the set threshold, which significantly reduces the false mode switching, ensures the continuity and stability of tracking, and C mode,k represents the confidence of the motion mode at time k (0-1).mode,k-1 denotes the mode confidence at time k-1, Δθ k denotes the heading rate at time k, a k denotes the acceleration at time k.

[0133] At step S306, a corresponding dynamic model is selected according to the motion mode, and a Kalman filter is used to estimate the target state based on the dynamic model.

[0134] Further, the embodiment selects a suitable dynamic model according to the motion mode of the target, such as a constant velocity model, a constant acceleration model, or a turning model. Then, a Kalman filter is used to estimate the state of the target based on the model and the measurement data. The Kalman filter is a recursive filtering algorithm that can estimate the state of a dynamic system and predict its future state in the presence of noise.

[0135] The constant velocity motion mode (CV) is obtained by the following formula:

[0136]

[0137] The constant acceleration motion mode (CA) is obtained by the following formula:

[0138]

[0139] The turning motion mode (CT) is obtained by the following formula: For the turning motion mode, the state transition matrix is a nonlinear function, and the following approximate linearization form can be used:

[0140]

[0141] where Δt is the time step, taking a value of 0.1 seconds; F CV , F CA , F CT denote the state transition matrices for the constant velocity motion mode, the constant acceleration motion mode, and the turning motion mode, respectively; ω is the angular velocity parameter in the turning motion mode, with a unit of rad / s.

[0142] At step S307, the following steps are iteratively performed until the current noise parameter variation is less than a threshold value or the maximum number of iterations is reached: Kalman filtering and backward smoothing processing are performed on the target position measurement data in the sliding window based on the current noise parameter, to obtain smoothed state estimation; the process noise covariance matrix and the measurement noise covariance matrix of the Kalman filter are updated based on the smoothed state estimation.

[0143] Further, the embodiment uses the expectation maximization (EM) algorithm to iteratively optimize the noise parameters of the Kalman filter, including performing Kalman filtering and backward smoothing on the data within the sliding window to obtain smoothed state estimates, and updating the process noise covariance matrix and the measurement noise covariance matrix based on the estimates, which helps to improve the adaptability of the Kalman filter to the actual measurement data. The embodiment uses historical data (including target position measurement data, target state estimation results, and state estimation covariance matrices thereof) to iteratively optimize the noise parameters of the Kalman filter by the expectation maximization algorithm. Please refer to Figure 6 The optimization flowchart of the expectation maximization algorithm is shown. Unlike the traditional EM algorithm, the embodiment designs a sliding window online EM algorithm, which maintains a data buffer with a fixed length T (a typical value is 20):

[0144]

[0145] wherein D represents the data buffer, z i represents the target position measurement data, represents the target state estimation results, P i represents the state estimation covariance matrix, i represents the time index, the range is t-T+1 to t, t represents the current time, and T represents the fixed window size.

[0146] As Figure 6 shown, the expectation maximization algorithm includes the following iterative process: initializing parameters 601, setting the initial process noise covariance matrix Q and the measurement noise covariance matrix R. Data collection 602, collecting data of the data buffer D from the sliding window. E-step: forward filtering 603, performing Kalman filtering on the data within the sliding window to obtain the prior estimate and the posterior estimate of each time step. E-step: backward smoothing 604, performing backward smoothing on the filtering results to obtain the smoothed state estimate and the smoothed covariance of each time step. M-step: updating the Q matrix 605, updating the process noise covariance matrix Q based on the smoothed state estimate; updating the R matrix 606, updating the measurement noise covariance matrix R based on the smoothed state estimate. Convergence check 607, checking whether the change of the noise parameters is less than a preset threshold or whether the maximum number of iterations is reached. If the conditions are met, terminate the iteration; otherwise, continue the optimization. Finally, output the optimized parameters 608 for output.

[0147] The embodiment updates the process noise covariance matrix Q and the measurement noise covariance matrix R based on the smoothed state estimate, and updates the process noise covariance matrix Q by the following formula:

[0148]

[0149] The measurement noise covariance matrix R is updated by the following formula:

[0150]

[0151] where Q new is the updated measurement noise covariance matrix; T is the length of the sliding window, set to 20 time steps; R new represents the updated measurement noise covariance matrix; k represents the time index within the window; P k+1|T is the smoothed state cross-covariance matrix at time k; P k,k+1 / T represents the smoothed state cross-covariance matrix between time k and k+1; P k+1,k / T represents the smoothed state cross-covariance matrix between time k+1 and k; is the smoothed state estimate at time k+1; represents the transpose of ; is the smoothed state estimate at time k; represents the transpose of k|T ; P k is the smoothed state covariance matrix at time k+1 based on all observations within the window; H k is the transpose of F k ; F k represents the state transition matrix at time k; represents the transpose of z k ; z k is the measurement at time k, with dimension 2; is the transpose of H k ; H t represents the measurement matrix at time k.

[0152] The embodiment introduces a time decay factor, so that the algorithm responds more quickly to environmental changes, and gives higher weight to recent data. The semi-positive definiteness and condition number of the noise covariance matrix are constrained to prevent parameter estimation divergence and enhance algorithm stability. The number of iterations is dynamically adjusted according to the change in noise parameters, balancing calculation efficiency and optimization effect. These improvements can adaptively adjust noise parameters online, and maintain optimal tracking performance when the radar interference environment or target maneuvering characteristics change.

[0153] Step S308, the prediction results of the dual-channel LSTM network are weighted and fused with the results of the target state estimation, and the final target state is output.

[0154] In an optional implementation, the step S308 includes:

[0155] Based on the result of the target state estimation, a norm of a Kalman filter innovation sequence is calculated, the Kalman filter innovation sequence being a difference between a measurement value at a current time and a prior state estimation of the Kalman filter at the current time;

[0156] Based on the norm of the Kalman filter innovation sequence and a preset weight adjustment strategy, a weight of a prediction result of the dual-channel LSTM network is determined;

[0157] According to the weight, the prediction result of the dual-channel LSTM network and the result of the target state estimation are weighted and fused to output a final target state.

[0158] Further, the embodiment fuses the prediction result of the LSTM network and the state estimation result of the Kalman filter by weighting, including calculating the norm of the Kalman filter innovation sequence, i.e. the difference between the measurement value and the prior state estimation, and determining the weight of the LSTM network prediction result based on the norm and the preset weight adjustment strategy. Then, the two results are weighted and fused according to the weight to output the final target state estimation. The fusion method combines the prediction ability of the deep learning model and the stability of the traditional filtering algorithm, and improves the accuracy and robustness of target tracking.

[0159] The embodiment adaptively fuses the prediction result of the dual-channel LSTM network and the result of the target state estimation to generate a final state estimation. First, the norm of the Kalman filter innovation sequence can be calculated by the following formula:

[0160]

[0161] wherein ||e|| is the Euclidean norm of the Kalman filter innovation sequence, z t is a measurement value at a current time, H is a measurement matrix, is a state estimation from a previous time to the current time.

[0162] The weight of the prediction result of the dual-channel LSTM network is calculated by the following formula:

[0163] w lstm = min(w max , max(w min , ae -β||e|| ))· y mode · y vel ;

[0164] wherein w min is a minimum weight (set as 0.05), w max is a maximum weight (set as 0.8), a and b are parameters for controlling the influence of the innovation sequence on the weight (0.5 and 0.3 respectively), and y modeAdjustment factor for current motion mode, CV mode is 0.7, CA mode is 1.2, CT mode is 1.3, gamma vel Adjustment factor for current speed, decrease when change sharply, increase when change gently.

[0165] The final target state is obtained by the following formula:

[0166]

[0167] Wherein, x lstm The prediction result of the double-channel LSTM network is x kf The estimation result of the Kalman filter is x The final state estimation after fusion.

[0168] Further, the present application combines and improves the three technologies of LSTM, KF and EM. For the LSTM network, the present example adopts a double-channel parallel architecture, which is different from the traditional single-input LSTM. Two independent channels of coordinate features and speed features are designed to model the target position evolution and speed change characteristics respectively. This design enables the network to capture the regularity of uniform motion and the nonlinear characteristics of maneuvering turns at the same time. Moreover, the double-channel LSTM network adopts a residual connection structure, which introduces residual connection in the feature fusion layer, effectively alleviating the gradient vanishing problem in deep network training and preserving the original feature information. This enables the network to learn the complex motion patterns of the target more effectively. The double-channel LSTM network also adopts a forgetting gate bias initialization. For the LSTM long sequence memory mechanism, the forgetting gate bias is specially initialized (set to 1.0), enhancing the network's ability to retain historical trajectory information, especially suitable for tracking target trajectories with long-term correlation.

[0169] For the Kalman filter, the present embodiment adopts a dynamic model switching mechanism based on mode confidence recursive update and smooth transition design, which solves the state jump problem caused by mode switching in the traditional IMM algorithm. During the mode switching process, the system uses linear interpolation smooth transition to reduce state estimation fluctuations. Moreover, the speed estimation optimization adopts a speed estimation method based on central difference and exponential weighted average, replacing the traditional forward difference method, which improves the numerical stability and accuracy while maintaining real-time performance.

[0170] For the EM algorithm, the sliding window online implementation breaks through the offline batch processing limitation of the traditional EM algorithm, an online EM algorithm based on a fixed-size sliding window is designed, the real-time adaptive optimization of the noise parameter is realized, and the computational complexity is reduced. In the parameter update, a time decay factor is introduced to give higher weight to recent data and improve the response speed of the algorithm to environmental changes. In addition, the semi-definiteness and condition number constraints of the noise covariance matrix are designed to prevent parameter divergence in the EM iteration process and enhance the stability of the algorithm in complex electromagnetic environments.

[0171] The adaptive fusion of the three technologies of LSTM, KF and EM is based on the LSTM weight adaptive adjustment mechanism of the Kalman filter innovation sequence, which realizes the intelligent fusion of data-driven and model-driven methods. When the innovation sequence is small (filter prediction is accurate), the LSTM weight is reduced, and when the innovation sequence is large (filter prediction deviation is large), the LSTM weight is increased. Moreover, the LSTM has the ability to extract deep features of the target historical trajectory, which provides auxiliary information for motion pattern recognition and improves the accuracy and stability of pattern recognition. A closed-loop feedback mechanism is constructed from EM optimized parameters, KF state estimation, LSTM training, pattern recognition to KF model selection, so that the components of the system promote each other and optimize together to form an adaptive learning system. Through the adaptive fusion of the three technologies of LSTM, KF and EM, the accuracy of position and velocity estimation of the application is significantly improved in the high-maneuvering target tracking scene, especially in the target turning, acceleration and deceleration processes. In complex electromagnetic environments, it can maintain high tracking performance and has stronger resistance to interference and abnormal measurement. It has faster response speed and more accurate recognition ability to changes in target motion patterns, and can quickly adapt to changes in target motion characteristics. While maintaining high precision, it meets the real-time processing requirements and is suitable for deployment in resource-constrained radar systems.

[0172] Further, the embodiment can use a multi-dimensional evaluation index system to evaluate the performance of the final output target state, including position root mean square error, velocity root mean square error, innovation consistency index and calculation efficiency index.

[0173] The position root mean square error is obtained by the following formula:

[0174]

[0175] Where n is the number of samples, indicating the number of time steps evaluated; is the estimated position at time i; is the true position at time i.

[0176] The velocity root mean square error is obtained by the following formula:

[0177]

[0178] wherein, is the estimated velocity vector [vx, vy] at time i; is the true velocity [vx, vy] at time i.

[0179] The innovation consistency index is obtained by the following formula:

[0180]

[0181] wherein, NIS k is the normalized innovation square at the kth time step; e k is the innovation sequence at the kth time step, i.e., the difference between the measurement and the prediction; n is the total number of samples. This index is used to evaluate the consistency of the Kalman filter. Ideally, NIS should obey the chi-square distribution with the degree of freedom equal to the measurement dimension. If the NIS value is too large, it indicates that the filter is too confident; if the NIS value is too small, it indicates that the filter is too conservative, represents the transpose matrix of e k , represents the inverse matrix of S k , S k represents the covariance matrix of the innovation sequence, represents the average value of NIS k at the last n time steps.

[0182] The computational efficiency index is obtained by the following formula:

[0183]

[0184] wherein, t proc is the average processing time, which measures the computational efficiency of the algorithm; n is the total number of samples processed; t end,k is the end timestamp of processing the kth sample; t start,k is the start timestamp of processing the kth sample.

[0185] Please refer to Figure 7The positioning error comparison diagram of different tracking methods under various motion scenes is shown. The smaller the value of the vertical axis (RMSE), the more accurate the target position or state estimation, and the better the tracking effect. The horizontal axis covers 5 typical motion scenes (uniform motion scene, uniform acceleration motion scene, turning motion scene, mixed mode motion scene, and high noise motion scene), which simulate the complexity of target motion and environment in actual tracking. Different colored curves in the figure represent the mean square error value (RMSE) of different methods from the true position. It can be seen that the error of the method of the embodiment (blue in the figure) is relatively small, and the predicted trajectory is closest to the true trajectory, especially in the target turning. For the first three scenes of uniform motion scene, uniform acceleration motion scene and turning motion scene, the traditional Kalman filter method (CV / CA model, corresponding to brown / orange) relies on a fixed motion model (uniform CV, uniform acceleration CA), and when the model does not match the scene, the error will increase significantly (the brown column is very high in the turning scene). The pure LSTM method (corresponding to purple) only uses neural network prediction, although it can fit complex motion, but lacks the model constraint of Kalman filter, and is prone to drift in high noise or long sequence tracking. The fixed weight LSTM-KF method (corresponding to green) simply fuses LSTM and Kalman filter, but the weight is fixed and cannot dynamically adapt to scene changes. The EM optimization method proposed in the embodiment (corresponding to blue) uses the EM algorithm to optimize the noise parameters of Kalman filter online, so that the model can adapt to the motion mode in real time, so the error is low in various scenes. For the fourth type of mixed mode motion scene, the target contains complex motions such as uniform motion, acceleration, turning, etc. The traditional method (pure Kalman, fixed fusion) has a single model or fixed weight, resulting in a significant increase in error (the highest brown column in the traditional CV model); The method proposed in the embodiment (blue, adaptive fusion red) relies on the combination of motion mode recognition, dynamic fusion and EM optimization, which can flexibly switch models and adjust parameters, and the error is much lower than other schemes. For the fifth type of high noise motion scene, when the radar sensor is disturbed, the measurement noise increases, and the traditional Kalman filter has a fixed noise parameter, and the error increases. Overall, the EM optimization method proposed in the embodiment has a lower RMSE value in various test scenes, especially in mixed mode and high noise environments, and its performance advantage is more obvious.

[0186] In summary, first, the embodiment obtains position measurement data of a moving target and extracts double-channel features using a sliding window, providing more comprehensive and accurate input information for subsequent target state prediction. Double-channel features can fully capture the position and speed changes of the target, better reflect the motion state of the target than single feature input, and improve the accuracy of prediction. Second, the double-channel features are input into a pre-trained double-channel LSTM network to obtain the prediction results of the target position and speed. The double-channel LSTM network structure processes coordinate features and speed features respectively, fully utilizes the multi-dimensional information of the target, and improves the accuracy of trajectory prediction. Then, the target motion characteristics are analyzed based on the target position measurement data and the target speed prediction results to obtain the motion mode, and the corresponding motion model is selected according to the motion mode, which can automatically adapt to different types of target motion and enhance the universality of the system. The target state is estimated through the Kalman filter to realize adaptive processing of different motion modes, which can accurately estimate the state according to the characteristics of different motion states, and improve the flexibility and adaptability of the tracking system. At the same time, the expectation maximization algorithm is used to optimize the noise parameters of the Kalman filter, reducing the dependence on prior knowledge and improving the adaptive ability of the system. Finally, the prediction results of the double-channel LSTM network and the target state estimation results are weighted and fused through an adaptive fusion mechanism to output the final target state, combining the advantages of the two methods to improve the tracking stability. The present application effectively solves the problems of insufficient tracking accuracy and poor adaptability of traditional target tracking methods in complex environments, and realizes high-precision and high-robustness target tracking in complex environments.

[0187] In the embodiment, a parameter adaptive based LSTM-KF hybrid tracking device is also provided, which is used to implement the above embodiments and preferred embodiments, and will not be described again. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware or a combination of software and hardware is also possible and contemplated.

[0188] The embodiment provides a parameter adaptive based LSTM-KF hybrid tracking device, as shown in Figure 8 , comprising:

[0189] A target position measurement data acquisition module 801 is configured to acquire target position measurement data of a moving target.

[0190] A double-channel feature extraction module 802 is configured to extract double-channel features of the target position measurement data through a sliding window.

[0191] The position and speed prediction module 803 is configured to input the two-channel feature into a pre-trained two-channel LSTM network to obtain a target position prediction result and a target speed prediction result.

[0192] The motion pattern acquisition module 804 is configured to analyze target motion characteristics based on the target position prediction result and the target speed prediction result to obtain a motion pattern of the moving target.

[0193] The target state estimation module 805 is configured to select a corresponding dynamic model according to the motion pattern and perform target state estimation on the dynamic model through a Kalman filter.

[0194] The noise parameter optimization module 806 is configured to optimize noise parameters of the Kalman filter by using an expectation maximization algorithm.

[0195] The final target state output module 807 is configured to perform weighted fusion on the prediction result of the two-channel LSTM network and the result of the target state estimation to output a final target state.

[0196] In some optional embodiments, the apparatus is further configured to:

[0197] The target position measurement data is preprocessed by performing outlier filtering and format standardization before the two-channel feature of the target position measurement data is extracted through the sliding window to obtain a target position measurement sequence.

[0198] In some optional embodiments, the two channels include a coordinate feature channel and a speed feature channel.

[0199] The coordinate feature channel includes a position sequence, a displacement sequence, an average speed sequence, and an instantaneous speed sequence.

[0200] The speed feature channel includes a speed evolution sequence of consecutive time steps.

[0201] In some optional embodiments, the two-channel LSTM network includes a coordinate feature processing channel, a speed feature processing channel, and a feature fusion layer.

[0202] The coordinate feature processing channel is two layers of cascaded LSTM units, the speed feature processing channel is a single-layer LSTM unit, and the feature fusion layer is configured to integrate the output of the coordinate feature processing channel and the output of the speed feature processing channel, and map the integrated feature to the target position prediction result and the target speed prediction result.

[0203] In some optional embodiments, the motion pattern acquisition module 804 is further configured to:

[0204] calculate an acceleration and a heading rate of change of the moving target between adjacent time frames based on the target position prediction result and the target speed prediction result;

[0205] classify the acceleration and the heading rate of change into a motion pattern based on a preset acceleration threshold and a preset angular velocity threshold, and determine a motion pattern of the moving target; the motion pattern includes a uniform motion pattern, a uniform acceleration motion pattern, or a turning motion pattern.

[0206] In some optional embodiments, the noise parameter optimization module 806 is further configured to:

[0207] includes iteratively performing the following steps until a current noise parameter variation is less than a threshold or a maximum number of iterations is reached:

[0208] perform Kalman filtering and backward smoothing processing on the target position measurement data in the sliding window based on a current noise parameter, to obtain a smoothed state estimation;

[0209] update a process noise covariance matrix and a measurement noise covariance matrix of the Kalman filter based on the smoothed state estimation.

[0210] In some optional embodiments, the final target state output module 807 is further configured to:

[0211] based on the result of the target state estimation, calculate a norm of a Kalman filtering innovation sequence; the Kalman filtering innovation sequence is a difference between a measurement value at a current time and a prior state estimation of the Kalman filter at the current time;

[0212] based on the norm of the Kalman filtering innovation sequence and a preset weight adjustment strategy, determine a weight of a prediction result of the dual-channel LSTM network;

[0213] weight the prediction result of the dual-channel LSTM network and the result of the target state estimation according to the weight, and output a final target state.

[0214] Further function descriptions of the above-mentioned various modules and units are the same as those of the above-mentioned corresponding embodiments, and will not be repeated here.

[0215] Please refer to Figure 9 , Figure 9 is a structural schematic diagram of a computer device provided in an optional embodiment of the present application, as shown in Figure 9As shown, the computer device includes one or more processors 910, memory 920, and interfaces 950 for external devices such as a keyboard and a mouse and a display device. The one or more processors 910 can be a single processing unit or a number of processing units, all of which can be on a single motherboard or in other relationships to each other. The memory 920 can be a single storage device or a number of storage devices, all of which can be on a single motherboard or in other relationships to each other. The interfaces 950 can be a number of interfaces, all of which can be on a single motherboard or in other relationships to each other. The computer device can be a number of these kinds of devices connected by a network. Figure 9 The processor 910 is used as an example in the embodiments.

[0216] The processor 910 can be a central processing unit, a network processor, or a combination thereof. The processor 910 can further include a hardware chip. The hardware chip can be an application specific integrated circuit, a programmable logic device, or a combination thereof. The programmable logic device can be a complex programmable logic device, a field programmable logic device, a general array logic, or any combination thereof.

[0217] The memory 920 stores instructions that can be executed by the at least one processor 910, so that the at least one processor 910 implements the method shown in the above embodiments.

[0218] The memory 920 can include a program region and a data region. The program region can store an operating system and an application program required by at least one function. The data region can store data created by use of the computer device, and the like. In addition, the memory 920 can include a high-speed random access memory, and can further include a non-transitory memory such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid state memory device. In some optional embodiments, the memory 920 can optionally include a memory disposed remotely with respect to the processor 910, and these remote memories can be connected to the computer device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0219] The memory 920 can include a volatile memory such as a random access memory, and can also include a non-volatile memory such as a flash memory, a hard disk, or a solid state disk. The memory 920 can further include a combination of the above-mentioned kinds of memories.

[0220] The computer device further includes a communication interface 930 for communication of the computer device with other devices or communication networks.

[0221] The embodiments of the present application further provide a computer readable storage medium, and the method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or stored in a remote storage medium or a non-transitory machine readable storage medium and downloaded to a local storage medium through network, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special hardware. The storage medium can be a magnetic disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned memories. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the method shown in the above embodiments is implemented.

[0222] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, the method and / or technical solutions according to the present application can be invoked or provided. Those skilled in the art should understand that the form of computer program instructions in a computer readable medium includes but is not limited to source files, executable files, installation package files, etc. Correspondingly, the way of executing computer program instructions by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0223] Although the embodiments of the present application are described in conjunction with the accompanying drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined.

Claims

1. A parameter adaptive based LSTM-KF hybrid tracking method, characterized in that, The method comprises: acquiring target position measurement data of a moving target; extracting double-channel features of the target position measurement data through a sliding window; inputting the double-channel features into a pre-trained double-channel LSTM network to obtain target position prediction results and target speed prediction results; analyzing target motion features based on the target position prediction results and the target speed prediction results to obtain a motion mode of the moving target; selecting a corresponding dynamic model according to the motion mode and performing target state estimation on the dynamic model through a Kalman filter; optimizing noise parameters of the Kalman filter by using an expectation maximization algorithm; performing weighted fusion of the prediction results of the double-channel LSTM network and the results of the target state estimation to output a final target state.

2. The method of claim 1, wherein, Before the step of extracting the double-channel features of the target position measurement data through the sliding window, the method further comprises: performing preprocessing of outlier filtering and format standardization on the target position measurement data to obtain a target position measurement sequence.

3. The method of claim 1, wherein, The double channel comprises a coordinate feature channel and a speed feature channel; The coordinate feature channel comprises a position sequence, a displacement sequence, an average speed sequence, and an instantaneous speed sequence; The speed feature channel comprises a speed evolution sequence of consecutive time steps.

4. The method of claim 1, wherein, The double-channel LSTM network comprises a coordinate feature processing channel, a speed feature processing channel, and a feature fusion layer; The coordinate feature processing channel is a two-layer cascaded LSTM unit, the speed feature processing channel is a single-layer LSTM unit, and the feature fusion layer is configured to integrate the output of the coordinate feature processing channel and the output of the speed feature processing channel, and map the integrated features to target position prediction results and target speed prediction results.

5. The method of claim 1, wherein, The step of analyzing target motion features based on the target position prediction results and the target speed prediction results to obtain a motion mode of the moving target comprises: calculating acceleration and heading rate change of the moving target between adjacent time frames based on the target position prediction results and the target speed prediction results; performing motion mode classification on the acceleration and the heading rate change based on preset acceleration threshold and preset angular velocity threshold to determine the motion mode of the moving target; the motion mode comprises a uniform motion mode, a uniform acceleration motion mode, or a turning motion mode.

6. The method of claim 1, wherein, The step of optimizing noise parameters of the Kalman filter by using an expectation maximization algorithm comprises iteratively performing the following steps until the current noise parameter change amount is less than a threshold or the maximum number of iterations is reached: performing Kalman filtering and backward smoothing processing on the target position measurement data in the sliding window based on the current noise parameters to obtain smoothed state estimation; updating the process noise covariance matrix and the measurement noise covariance matrix of the Kalman filter based on the smoothed state estimation.

7. The method according to any one of claims 1 to 6, characterized in that, The step of performing weighted fusion of the prediction results of the double-channel LSTM network and the results of the target state estimation to output a final target state comprises: Based on the result of the target state estimation, a norm of a Kalman filter innovation sequence is calculated; the Kalman filter innovation sequence is a difference between a measurement value at a current time and a prior state estimation of the Kalman filter at the current time; Based on the norm of the Kalman filter innovation sequence and a preset weight adjustment strategy, a weight of a prediction result of the dual-channel LSTM network is determined; According to the weight, the prediction result of the dual-channel LSTM network and the result of the target state estimation are weighted and fused to output a final target state.

8. A parameter-adaptive based LSTM-KF hybrid tracking apparatus, characterized in that, The device comprises: a target position measurement data acquisition module configured to acquire target position measurement data of a moving target; a dual-channel feature extraction module configured to extract dual-channel features of the target position measurement data through a sliding window; a position and speed prediction module configured to input the dual-channel features into a pre-trained dual-channel LSTM network to obtain a target position prediction result and a target speed prediction result; a motion pattern acquisition module configured to analyze target motion characteristics based on the target position prediction result and the target speed prediction result to obtain a motion pattern of the moving target; a target state estimation module configured to select a corresponding dynamic model according to the motion pattern and perform target state estimation on the dynamic model through a Kalman filter; a noise parameter optimization module configured to optimize noise parameters of the Kalman filter using an expectation maximization algorithm; a final target state output module configured to weightedly fuse the prediction result of the dual-channel LSTM network and the result of the target state estimation to output a final target state.

9. A computer device, comprising: comprise: a memory and a processor, which are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the method in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer instructions, and the computer instructions are used to make the computer execute the method in any one of claims 1 to 7.

Images