Multi-target azimuth course extracting and tracking method and system based on LSTM (Long Short Term Memory)
By constructing a target orientation array and using LSTM to predict the orientation, the multi-objective tracking problem in a water acoustic environment is solved, and the accurate tracking and enhancement effect of multi-objective trajectory is achieved.
Patent Information
- Application Number
- CN202311639033.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-30
- Publication Date
- 2025-08-05
AI Technical Summary
In complex water acoustic environments, it is difficult for the prior art to effectively distinguish and continuously track the orientation of multiple targets, resulting in target loss or orientation extraction errors.
Using an LSTM-based method, multi-objective orientation process extraction and tracking is achieved by constructing target orientation arrays and using long and short-term memory networks to predict target orientations, combining the normalization of sound energy and preset change thresholds.
Effective classification and tracking of multi-target trajectories is achieved, the extraction effect of intermittent and unstable trajectories is improved, the impact of energy level differences is eliminated, and more accurate target orientation and trajectory is obtained.
Smart Images
Figure CN120429696A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of moving target tracking, and in particular to a method and system for extracting and tracking the position history of multiple targets based on LSTM. Background Art
[0002] Beamforming is the most commonly used method for underwater target detection. The bearing history map is its most important output, representing the amount of acoustic energy received in different directions at different times. Typically, the direction with the highest acoustic energy corresponds to the target. By extracting the bearing history map based on this criterion, we can determine the target's position at different times. However, when there are multiple targets, this criterion no longer applies. Furthermore, due to the distribution characteristics of the ocean acoustic field, the energy in the direction of the target may not always be highest. Therefore, extracting the bearing history map using a simple method of extracting the maximum energy value can lead to target loss or erroneous bearing extraction. Summary of the Invention
[0003] In view of this, the present invention provides a multi-target position history extraction and tracking method and system based on LSTM, which can distinguish multiple targets in a complex underwater acoustic environment and continuously track the target position.
[0004] The specific technical solutions adopted in the present invention are as follows:
[0005] A multi-target orientation history extraction and tracking method based on LSTM comprises: constructing a target orientation array according to an orientation history graph containing multiple target trajectories, wherein the rows of the target orientation array represent the orientations of different targets at the same moment, and the columns of the target orientation array represent the orientations of the same target at different moments; predicting the orientation of each target at the next moment using a long short-term memory (LSTM) network according to the target orientation array, and obtaining the orientation of each target at the next moment in the orientation history graph according to the predicted orientations to obtain a target orientation array at a complete moment; and obtaining the trajectory of each target at the complete moment according to the target orientation array at the complete moment.
[0006] Furthermore, the target orientation array is constructed based on the orientation history diagram containing multiple target trajectories, including: obtaining the maximum value and the maximum value of the sound energy at the initial moment as suspected values based on the orientation history diagram containing multiple target trajectories; obtaining the initial suspected orientations corresponding to the first N suspected values of the suspected values sorted from large to small, and constructing the target orientation array at the initial moment, wherein N is a positive integer; continuously collecting suspected values at K moments, and comparing the suspected orientation at each moment with the suspected orientation at the previous moment, and storing the suspected orientation that meets the preset change threshold into the target orientation array to complete the construction of the target orientation array, wherein K is a positive integer.
[0007] Furthermore, storing the suspected directions satisfying the preset change threshold into the target direction array includes: placing the suspected directions satisfying the preset change threshold in the same column of the target direction array according to time sequence.
[0008] Furthermore, before obtaining the maximum value and the maximum value of the acoustic energy at the initial moment according to the orientation history graph containing multiple target trajectories, the method further includes: normalizing the acoustic energy at each moment of the orientation history graph.
[0009] Furthermore, based on the target orientation array, a long short-term memory network (LSTM) is used to predict the orientation of each target at the next moment, including: inputting each column orientation of the target orientation array into the LSTM in sequence to predict the orientation of each target at the next moment.
[0010] Furthermore, obtaining the orientation of each target at the next moment in the orientation history graph based on the predicted orientation includes: at the next moment, traversing the suspected value corresponding to the predicted orientation of each target in the orientation history graph based on the predicted orientation of each target; if the traversal is successful, storing the orientation corresponding to the suspected value in the orientation column of the corresponding target as the trajectory of the target; if the traversal fails, starting the orientation prediction and traversal at subsequent moments, and deleting the trajectory of the target in the target orientation array when the preset time is reached and the traversal is still unsuccessful.
[0011] Furthermore, it also includes: in the case of traversal failure, starting the direction prediction and traversal at subsequent moments, and in the case of successful traversal at subsequent moments, fitting the suspected values corresponding to the directions in the same column of the target direction array to obtain the suspected values corresponding to the moment of traversal failure, so as to obtain the target direction array at the complete moment.
[0012] Furthermore, according to the target orientation array of the complete moment, the complete moment trajectory of each target is obtained, including: each column of orientation of the target orientation array of the complete moment is sequentially corresponded as the complete moment trajectory of each target.
[0013] A multi-target orientation history extraction and tracking system based on LSTM comprises: an orientation acquisition module for constructing a target orientation array based on an orientation history graph containing multiple target trajectories, wherein the rows of the target orientation array represent the orientations of different targets at the same moment, and the columns of the target orientation array represent the orientations of the same target at different moments; an orientation prediction module for predicting the orientation of each target at the next moment using a long short-term memory (LSTM) network based on the target orientation array, and obtaining the orientation of each target at the next moment in the orientation history graph based on the predicted orientation, so as to obtain a target orientation array at a complete moment; and a trajectory acquisition module for obtaining the trajectory of each target at a complete moment based on the target orientation array at a complete moment.
[0014] Furthermore, the orientation acquisition module includes: a suspected value acquisition unit, which is used to obtain the maximum value and the maximum value of the sound energy at the initial moment as the suspected value based on the orientation history diagram containing multiple target trajectories; an orientation acquisition unit, which is used to obtain the initial suspected orientations corresponding to the first N suspected values of the suspected values sorted from large to small, and construct a target orientation array at the initial moment, wherein N is a positive integer; an array acquisition unit, which is used to continuously collect suspected values at K moments, and compare the suspected orientation at each moment with the suspected orientation at the previous moment, and store the suspected orientation that meets the preset change threshold into the target orientation array to complete the construction of the target orientation array, wherein K is a positive integer.
[0015] Beneficial effects:
[0016] (1) A multi-target orientation history extraction and tracking method based on LSTM is constructed by constructing a target orientation array based on an orientation history graph containing multiple target trajectories, wherein the rows of the target orientation array represent the orientations of different targets at the same time, and the columns of the target orientation array represent the orientations of the same target at different times; based on the target orientation array, the orientation of each target at the next time is predicted using LSTM, and the orientation of each target at the next time is obtained from the orientation history graph based on the predicted orientation to obtain the target orientation array at the complete time; based on the target orientation array at the complete time, the trajectory of each target at the complete time is obtained. It can realize the classification and tracking of multiple trajectories and has a good extraction and enhancement effect on intermittent unstable trajectories.
[0017] (2) Normalizing the sound energy at each moment of the azimuth history diagram can eliminate the impact of the overall level difference of energy at different moments.
[0018] (3) In the case of traversal failure, the azimuth prediction and traversal at subsequent moments are started. In the case of traversal success at subsequent moments, the suspected values corresponding to the azimuths in the same column of the target azimuth array are fitted to obtain the suspected values corresponding to the moment of traversal failure, so as to obtain the target azimuth array at the complete moment and obtain better multi-target tracking effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 Flowchart of a method for extracting and tracking the position history of multiple targets based on LSTM according to an embodiment of the present invention;
[0020] Figure 2 It is a schematic diagram of the direction extraction result provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0021] The present invention is described in detail below with reference to the accompanying drawings and embodiments.
[0022] The embodiment of the present invention provides a multi-target position history extraction and tracking method based on a long short-term memory network (LSTM). Figure 1 : is a flowchart of a multi-target position history extraction and tracking method based on LSTM according to an embodiment of the present invention, such as Figure 1 Shown, including:
[0023] Step S101: constructing a target orientation array based on an orientation history graph containing multiple target trajectories, wherein the rows of the target orientation array represent the orientations of different targets at the same time, and the columns of the target orientation array represent the orientations of the same target at different times;
[0024] In an exemplary embodiment, before obtaining the maximum value and the maximum value of the acoustic energy at the initial moment according to the orientation history graph containing multiple target trajectories, the method further includes: normalizing the acoustic energy at each moment of the orientation history graph.
[0025] In actual implementation, the azimuth map containing multiple target tracks is preprocessed by normalizing the acoustic energy at each moment in the azimuth map. This normalization eliminates the effects of overall energy level differences at different moments. The azimuth map extraction is initialized by setting the maximum number of targets to be tracked, N.
[0026] In an exemplary embodiment, a target orientation array is constructed based on an orientation history diagram containing multiple target trajectories, including: obtaining the maximum value and the maximum value of the sound energy at the initial moment as suspected values based on the orientation history diagram containing multiple target trajectories; obtaining the initial suspected orientations corresponding to the first N suspected values sorted from large to small, and constructing a target orientation array at the initial moment, wherein N is a positive integer; continuously collecting suspected values at K moments, and comparing the suspected orientation at each moment with the suspected orientation at the previous moment, and storing the suspected orientation that meets the preset change threshold in the target orientation array to complete the construction of the target orientation array, wherein K is a positive integer.
[0027] In actual implementation, the initial moment orientation history is extracted: the maximum and maximum acoustic energy values at the initial moment are extracted as suspected values and sorted from largest to smallest. The orientations corresponding to the first N suspected values are taken as the initial moment suspected orientation group, i.e., the target orientation array mentioned above, and stored in the first row of the suspected orientation array. The descending sorting order is used to sort the acoustic energy values of multiple targets; each extreme or maximum value is likely a target.
[0028] After that, background accumulation: For the next K moments, each suspected value is extracted and compared with the suspected bearing at the previous moment. Bearings with a bearing difference less than the bearing change threshold are stored in the same column as the suspected bearing. This bearing change threshold, also known as the preset change threshold, is the maximum possible bearing change between two moments and is selected based on prior knowledge such as the target type and the time difference between the two moments.
[0029] In an exemplary embodiment, storing the suspected positions that meet the preset change threshold into the target position array includes placing the suspected positions that meet the preset change threshold in the same column of the target position array according to time sequence.
[0030] Step S102: Based on the target orientation array, the LSTM is used to predict the orientation of each target at the next moment, and the orientation of each target at the next moment is obtained from the orientation history graph based on the predicted orientation to obtain the target orientation array at the complete moment;
[0031] In an exemplary embodiment, based on the target orientation array, LSTM is used to predict the orientation of each target at the next moment, including: inputting each column orientation of the target orientation array into LSTM in sequence, and predicting the orientation of each target at the next moment.
[0032] In actual implementation, direction prediction is performed using an array of suspected direction angles accumulated over K moments as background data, where each column represents the same trajectory. An LSTM is used to predict the direction at the next moment and then find the maximum value near that direction at the next moment. The LSTM prediction method first trains a long short-term memory network using a large amount of historical data on targets of the same type. Each column of the suspected direction array is then passed through the LSTM network as input to obtain the next predicted direction value for that column. Repeating the prediction process for each column of the suspected direction array yields a predicted direction value for each trajectory. A maximum energy value corresponds to a direction. Because a direction history graph plots direction on the horizontal axis and time on the vertical axis, the energy maximum at a given moment corresponds to a direction.
[0033] In an exemplary embodiment, the orientation of each target at the next moment is obtained in the orientation history graph based on the predicted orientation, including: at the next moment, based on the predicted orientation of each target, traversing the suspected value corresponding to the predicted orientation of each target in the orientation history graph; if the traversal is successful, storing the orientation corresponding to the suspected value in the orientation column of the corresponding target as the trajectory of the target; if the traversal fails, starting the orientation prediction and traversal at subsequent moments, and if the preset time is reached and the traversal is still not successful, deleting the trajectory of the target in the target orientation array.
[0034] In an exemplary embodiment, it also includes: in the case of traversal failure, starting the azimuth prediction and traversal at subsequent moments, and in the case of successful traversal at subsequent moments, fitting the suspected values corresponding to the azimuths in the same column of the target azimuth array to obtain the suspected values corresponding to the moment of traversal failure, so as to obtain the target azimuth array at the complete moment.
[0035] In actual implementation, track maintenance is performed: suspected values are extracted for the next moment. If a suspected position meeting the requirements is found near the predicted position, it is included in the track. If no satisfactory position is found, the track is temporarily maintained, and the position at the next moment is predicted based on the historical values. The search continues until a suspected value near the predicted value is found or the hold time expires. If suspected values near the predicted value are found continuously, the historical values are fitted with the found values to fill in the missing values and obtain a complete track. If no suspected position is found after the hold time expires, the track is considered terminated and tracking is stopped.
[0036] Step S103: Obtain the complete moment trajectory of each target according to the complete moment target orientation array.
[0037] In an exemplary embodiment, obtaining the complete moment trajectory of each target according to the complete moment target orientation array includes: sequentially corresponding each column of the complete moment target orientation array to the complete moment trajectory of each target.
[0038] In practice, after all time steps are processed, an array with N columns is obtained, each representing a target's trajectory. Based on the completeness of the trajectory, the first N / 4 complete columns are used as the final N / 4 suspected target trajectory. Figure 2 is a schematic diagram of the direction extraction result provided by an embodiment of the present invention, such as Figure 2 As shown, the orientation extraction and tracking of multiple targets are demonstrated.
[0039] An embodiment of the present invention also provides a multi-target orientation history extraction and tracking system based on LSTM, including: an orientation acquisition module, used to construct a target orientation array based on an orientation history graph containing multiple target trajectories, wherein the rows of the target orientation array represent the orientations of different targets at the same moment, and the columns of the target orientation array represent the orientations of the same target at different moments; an orientation prediction module, used to use LSTM to predict the orientation of each target at the next moment based on the target orientation array, and obtain the orientation of each target at the next moment in the orientation history graph based on the predicted orientation to obtain the target orientation array at the complete moment; and a trajectory acquisition module, used to obtain the trajectory of each target at the complete moment based on the target orientation array at the complete moment.
[0040] In an exemplary embodiment, the orientation acquisition module includes: a suspected value acquisition unit, which is used to obtain the maximum value and the maximum value of the sound energy at the initial moment as the suspected value based on the orientation history diagram containing multiple target trajectories; an orientation acquisition unit, which is used to obtain the initial suspected orientations corresponding to the first N suspected values sorted from large to small, and construct a target orientation array at the initial moment, wherein N is a positive integer; an array acquisition unit, which is used to continuously collect suspected values at K moments, and compare the suspected orientation at each moment with the suspected orientation at the previous moment, and store the suspected orientation that meets the preset change threshold into the target orientation array to complete the construction of the target orientation array, wherein K is a positive integer.
[0041] In the actual implementation process, the above-mentioned location history extraction and tracking system is used to implement the specific steps of the location history extraction and tracking method. The functions of each module in the location history extraction and tracking system correspond to the location history extraction and tracking method, and will not be described one by one here.
[0042] In summary, the embodiments of the present invention provide a multi-target azimuth extraction and tracking method and system based on LSTM. Based on the energy levels at different azimuths at the initial moment on the azimuth history graph, the maximum and maximum energy values at the initial moment are sequentially found, along with their corresponding azimuths. These maximum and maximum values are called suspected values, and the corresponding azimuths are called suspected target azimuths, which are then saved. At the next moment, the same method is used to find suspected values and suspected azimuths. Each suspected azimuth found is then compared with all suspected azimuths from the previous moment. A threshold for azimuth change is used to determine whether the suspected azimuth belongs to the same trajectory as the suspected azimuth from the previous moment. Simultaneously, the LSTM is used to predict the azimuth of each trajectory at the next moment. At the next moment, a search is conducted near the predicted azimuth. If no suspected value is found, the trajectory is temporarily retained and prediction continues. If no suspected value is found after a period of prediction, the trajectory is abandoned. Otherwise, the predicted and suspected values are used to fit the trajectory. This method can classify and track multiple trajectories and has excellent extraction and enhancement effects for intermittent and unstable trajectories.
[0043] The above specific embodiments merely illustrate the design principles of the present invention. The shapes and names of the components described herein may vary and are not limiting. Therefore, those skilled in the art may modify or substitute equivalents for the technical solutions described in the above embodiments. Such modifications and substitutions, without departing from the inventive spirit and technical solutions of the present invention, shall fall within the scope of protection of the present invention.
Claims
1. A multi-target position history extraction and tracking method based on LSTM, characterized in that: include: Constructing a target orientation array according to an orientation history graph containing multiple target trajectories, wherein the rows of the target orientation array represent the orientations of different targets at the same time, and the columns of the target orientation array represent the orientations of the same target at different times; According to the target orientation array, a long short-term memory network (LSTM) is used to predict the orientation of each target at the next moment, and the orientation of each target at the next moment is obtained in the orientation history graph according to the predicted orientation, so as to obtain the target orientation array at the complete moment; The complete moment trajectory of each target is obtained according to the complete moment target orientation array.
2. The method according to claim 1, wherein The target orientation array is constructed according to the orientation history graph containing multiple target trajectories, including: According to the azimuth history diagram containing multiple target trajectories, the maximum and maximum values of the acoustic energy at the initial moment are obtained as suspected values; Obtaining the initial suspected directions corresponding to the first N suspected values of the suspected values sorted from largest to smallest, and constructing a target direction array at the initial moment, where N is a positive integer; Continuously collect suspected values at K moments, compare the suspected orientation at each moment with the suspected orientation at the previous moment, and store the suspected orientation that meets the preset change threshold into the target orientation array to complete the construction of the target orientation array, where K is a positive integer.
3. The method according to claim 2, wherein The step of storing the suspected orientation that meets the preset change threshold into the target orientation array includes: The suspected directions satisfying the preset change threshold are placed in the same column of the target direction group according to the time sequence.
4. The method according to claim 2, wherein Before obtaining the maximum value and the maximum value of the acoustic energy at the initial moment according to the azimuth history diagram containing multiple target trajectories, the method further includes: Normalization processing is performed on the acoustic energy at each moment of the azimuth history diagram.
5. The method according to claim 1, wherein Based on the target orientation array, a long short-term memory network (LSTM) is used to predict the orientation of each target at the next moment, including: Each column of the target orientation array is input into the LSTM in sequence to predict the orientation of each target at the next moment.
6. The method according to claim 1, wherein The step of obtaining the position of each target at a next moment in the position history graph according to the predicted position includes: At a next moment, according to the predicted orientation of each target, traversing the suspected value corresponding to the predicted orientation of each target in the orientation history graph; If the traversal is successful, the direction corresponding to the suspected value is stored in the direction column of the corresponding target as the trajectory of the target; if the traversal fails, the direction prediction and traversal at subsequent moments are started. If the preset time is reached and the traversal is still not successful, the trajectory of the target in the target direction array is deleted.
7. The method according to claim 6, wherein Also includes: In the case of traversal failure, the azimuth prediction and traversal at subsequent moments are started. In the case of successful traversal at subsequent moments, the suspected values corresponding to the azimuths in the same column of the target azimuth array are fitted to obtain the suspected values corresponding to the moment of traversal failure, so as to obtain the target azimuth array at the complete moment.
8. The method according to claim 1, wherein According to the target orientation array at the complete moment, the complete moment trajectory of each target is obtained, including: Each column of the target orientation array at the complete moment corresponds in sequence to the complete moment trajectory of each target.
9. A multi-target position history extraction and tracking system based on LSTM, characterized in that: include: A direction acquisition module is used to construct a target direction array based on a direction history graph containing multiple target trajectories, wherein the rows of the target direction array represent the directions of different targets at the same time, and the columns of the target direction array represent the directions of the same target at different times; A direction prediction module is used to predict the direction of each target at the next moment based on the target direction array using a long short-term memory network (LSTM), and obtain the direction of each target at the next moment in the direction history graph based on the predicted direction to obtain the target direction array at the complete moment; The trajectory acquisition module is used to acquire the trajectory of each target at the complete moment according to the target orientation array at the complete moment.
10. The system according to claim 1, wherein: The position acquisition module includes: A suspected value obtaining unit, configured to obtain, according to a azimuth history diagram containing multiple target trajectories, a maximum value and a maximum value of the acoustic energy at an initial moment as suspected values; A direction acquisition unit, configured to acquire initial suspected directions corresponding to the first N suspected values sorted from largest to smallest, and construct a target direction array at an initial moment, where N is a positive integer; An array acquisition unit is used to continuously collect suspected values at K moments, compare the suspected orientation at each moment with the suspected orientation at the previous moment, and store the suspected orientation that meets the preset change threshold into the target orientation array to complete the construction of the target orientation array, where K is a positive integer.