A method for constructing target association tracking dataset based on self-reported position data

Abstract

The present invention provides a method for constructing a target association tracking dataset based on self-reported position data. Scenario parameters are set based on the sensor sources of the multi-source target tracks to be constructed. The desired tracks are extracted from publicly available civilian AIS data, and the target density and maneuverability within the spatial grid are statistically analyzed. This data is used as a basic database for track simulation and stored in CSV format as an AIS spatial coding index file. Based on the target density and maneuverability within the spatial grid, AIS track data for the desired grid is extracted from the AIS spatial coding index file. Based on the extracted real tracks, a multi-sensor target track dataset that meets the set scenario parameters is constructed through coordinate translation, random batching, systematic error generation, random error generation, and track association mapping. This method constructs a unified, standardized, large-scale, high-quality dataset based on publicly available data, providing a fundamental guarantee for training deep learning neural networks for multi-source track association.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of radar target association tracking and are applicable to solving the problem of constructing a radar target association tracking data training set. Background Art

[0002] The problem of multi-source track correlation is a common problem in radar data processing, including the problem of reconstructing interrupted tracks, and has a long history of research. With the rapid development of artificial intelligence and deep learning technologies, some scholars have proposed using a data-driven approach based on deep learning to accomplish track correlation. This approach uses neural networks to extract information such as track motion characteristics, interruption characteristics, and multi-source error characteristics. A loss function is then used to constrain track characteristics in high-dimensional space, enabling a mapping from track data to correlation results.

[0003] Deep learning approaches for track correlation require network training that relies on vast amounts of track data. In the field of deep learning, data is becoming as important as models and algorithms, and a unified, standardized, and large-scale high-quality dataset is essential for the advancement of relevant research. However, the field of multi-source track correlation lacks a unified, standardized, and large-scale track correlation dataset to support it, and the construction of such datasets remains a rarity both domestically and internationally.

[0004] To address the lack of track association datasets, this paper proposes a method for constructing a target association tracking dataset based on self-reported position data. This dataset is constructed from global AIS track data through gridding, automatic interruption, and noise addition. Each scenario sample consists of tracks based on publicly available data, covering a variety of motion patterns, target types, and durations. Summary of the Invention

[0005] The embodiment of the present invention provides a method for constructing a target association tracking dataset based on self-reported position data, and constructs a unified, standardized, large-scale, high-quality dataset based on public data, thereby providing basic guarantees for the training of multi-source track association deep learning neural networks.

[0006] The method for constructing a target association tracking dataset based on self-reported position data described in the present invention includes the following technical measures:

[0007] Step 1: Set the scene parameters;

[0008] Set the scene parameters according to the sensor source of the multi-source target track to be constructed, including scene coordinate parameters, sensor update cycle, target detection probability, track start time range, track end time range, minimum duration, interruption frequency, interruption time range, position system deviation, and track quality;

[0009] Step 2: extract the real track;

[0010] The public AIS track data is encoded according to the spatial grid, and the MMSI number, number of tracks, target density, mean heading variance, mean speed variance, and target maneuverability within each grid are counted and saved as an AIS spatial coding index file in CSV format;

[0011] Step 3: Generate source track;

[0012] Based on the real tracks extracted in step 2, a dataset of multiple sensor target tracks that meets the scenario parameters set in step 1 is constructed.

[0013] Furthermore, the method for extracting the real track described in step 2 is:

[0014] S21: Construction of AIS basic track library;

[0015] Extract the required tracks from public civil AIS data, calculate the density and maneuverability of targets within the spatial grid, and use this data as the basic database for track simulation. Save the data as an AIS spatial coding index file in CSV format.

[0016] S22: ground truth track extraction;

[0017] According to the target parameter settings in step 1, the AIS track data of the required grid is extracted from the AIS spatial coding index file based on the target density and maneuverability in the spatial grid.

[0018] Furthermore, the method for generating the source track described in step 3 is:

[0019] S31: According to the scene coordinate parameter settings in step 1, coordinate translation is performed on the AIS track in the grid extracted in step 2, and the true value batch number of the track is set;

[0020] S32: Based on the parameter settings of the source update cycle, discovery probability, track duration, interruption frequency, etc. in step 1, the AIS track is screened, interrupted, and interpolated to obtain multiple track segments and randomly batched.

[0021] S33: According to the source system deviation setting in step 1, a systematic error is obtained by random sampling from a uniformly distributed sequence, and then added to the track segment generated in S32 to obtain a simulated track with a systematic error;

[0022] S34: According to the source track quality setting in step 1, random errors are obtained by random sampling from a Gaussian distribution or Rayleigh distribution sequence, and then added to the track generated in S33 to obtain a simulated track with random errors;

[0023] S35: Generate a track association mapping table to record the correspondence between the simulated track and the original AIS track batch number;

[0024] S36: Mix the simulated tracks of the two sources, sort them by time, and save them as a source track CSV file;

[0025] S37: The association mapping tables of the two source tracks are mixed, sorted by start time, and saved as an association mapping table CSV file.

[0026] The embodiment of the present invention constructs a training sample set corresponding to the target object, including sample pairs of different types of images of the target object, and inputs the training sample set into a convolutional neural network for training. Then, the image to be identified of the target object is obtained, and the image to be identified of the target object is input into the trained convolutional neural network to obtain a multi-target detection result, and continues to solve the multi-target detection result to determine the multi-target association result based on the multi-target detection result, thereby solving the problem of low target association accuracy in the existing multi-target association method, and can reduce the ambiguity caused by the use of a single feature association, thereby improving the accuracy of multi-target association of the target object.

[0027] To address the issue of missing track association datasets, the present invention constructs a track association dataset by gridding, automatically interrupting, and adding noise to publicly available global AIS track data. This dataset consists of two major parts: a training set and a test set, containing over one million tracks. The training set contains 5,000 scene samples, and the test set contains 1,000 scene samples. Each scene sample consists of several to several hundred tracks, covering a variety of motion modes, target types, and durations of varying lengths. Furthermore, the constructed dataset is further visualized and analyzed, with detailed research on the characteristics of the tracks within each grid, demonstrating the richness, rationality, and effectiveness of the dataset. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flowchart of generating associated samples provided by an embodiment of the present invention;

[0029] Figure 2 is a target density heat map provided by an embodiment of the present invention;

[0030] Figure 3 is a target maneuverability heat map provided by an embodiment of the present invention;

[0031] Figure 4 This is a typical scene display diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention.

[0033] It should also be noted that, for ease of description, only the parts relevant to the present invention, not all of the contents, are shown in the accompanying drawings. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe each operation as sequential processes, many of the operations therein can be implemented in parallel, concurrently, or simultaneously. In addition, the order of each operation can be rearranged. When its operation is completed, the process can be terminated, but can also have additional steps not included in the accompanying drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0034] The terms "including," "comprising," and "having," and any variations thereof, in the description, claims, and drawings of the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements and may include steps or elements that are not listed.

[0035] Example conditions: The Automatic Identification System (AIS) is a broadcast shipborne transponder system that enables ships to continuously transmit their identity, location, heading, speed, and other data to nearby ships and shore monitoring departments over a public wireless channel. The user identity is usually represented by MMSI. Compared with radar track data, AIS track data has the advantages of wide distribution, low difficulty in acquisition, and good timeliness. Therefore, global AIS data is used here to construct a multi-source track correlation dataset. The specific steps of the present invention are shown in the attached figure. Figure 1 shown.

[0036] Step 1: Set scene parameters

[0037] Scene settings mainly set the grid accuracy and the longitude and latitude of the scene center; the grid accuracy α is used for global grid division and represents the minimum interval of global longitude and latitude division; the longitude and latitude of the scene center W o ,Subscript o represents the center latitude, which is used for the subsequent translation of the spatial grid;

[0038] S11: Goal Setting

[0039] Target setting mainly involves setting target density and target mobility; target density Reflects the number of targets in grid area j, with a value range of [0, 1]; the degree of target mobility Reflects the degree of target motion change within grid area j, with a value range of [0, 1]. Their calculation method is:

[0040]

[0041]

[0042] In the formula, min means to find the minimum value, max means to find the maximum value; n j Represents the number of targets in grid area j; It represents the average value of the standard deviation of target speed in grid area j, and the subscript v is the abbreviation of velocity. It represents the average value of the standard deviation of the target course in grid area j. The subscript c is the abbreviation of the English representation of course.

[0043] S12: Source Settings

[0044] The signal source settings mainly set the detection characteristics of signal source 1 and signal source 2; the main parameters include update cycle, target detection probability, track start time range, track end time range, minimum duration, interruption frequency, interruption time range, position system deviation, and track quality noise;

[0045] Step 2: Extract the real track based on the spatial grid

[0046] The ground-truth track extraction based on spatial grid includes two steps: building the AIS basic track library and extracting the ground-truth track.

[0047] S21a: Extract the single target track from the AIS data file according to the MMSI number and save it as a CSV file named MMSI number, where CSV is a file type in a value-separated format;

[0048] S21b: Preprocessing of a single target track, including splitting tracks that have not been updated for a long time, deleting tracks that are stationary or have too low a speed, deleting tracks with sample point jumps, and deleting tracks that are too short; splitting tracks that have not been updated for a long time, and truncating the track once every 600 seconds when the track update time is greater than until the track ends;

[0049] S21c: Encode the AIS data for one day according to the spatial grid, and divide the longitude and latitude according to the grid accuracy α, that is, every α longitude and α latitude constitute a grid;

[0050] S21d: Considering the longitude range [-180,180] and latitude [-90,90], set the grid division accuracy α = 0.5, then the number of grid divisions is: the number of longitudes Number of latitudes The grid tensor is constructed according to the number of divisions, and its dimension is [360, 720, *], where * indicates that the last dimension is indefinite; the corresponding relationship between the grid index (i, j) and the latitude and longitude of the upper left corner of the grid (a, b) is

[0051]

[0052] S21e: Traverse each sampling point of each track, solve i,j according to formula (3) and round down to obtain the grid index of the corresponding track sampling point After removing duplicate indexes, the MMSI of the track is stored in the corresponding grid index;

[0053] S21f: Count the MMSI number, number of tracks, number of targets, mean heading variance, mean speed variance, target density, and target maneuverability in each grid, and save them in CSV format as an AIS spatial coding index file, with one row for each spatial grid. The specific format is {spatial grid dimension index, spatial grid longitude index, number of tracks, number of targets, mean heading variance, mean speed variance, target density, target maneuverability, MMSI number sequence};

[0054] S22: There are two modes for extracting true track: random extraction and conditional extraction. Random extraction involves randomly extracting spatial codes and then obtaining all MMSI numbers in the grid based on the AIS spatial code index file to obtain the true track. Conditional extraction involves selecting the spatial grid that is most similar to the set target density and maneuverability, or extracting from multiple similar spatial grids.

[0055] Step 3: Source track generation

[0056] S31: First, take the extracted AIS track Z0 in the grid as the true value, translate the AIS track in the grid according to the latitude and longitude of the scene center W0 and the source parameters, and set the true value batch number of the track. The longitude and latitude after translation are Expressed as

[0057]

[0058] Among them, (a,b) represents the original longitude and latitude, (i,j) represents the longitude and latitude index of the current grid, Indicates the longitude and latitude of the center of the battlefield;

[0059] S32a: Find the minimum duration τ of the track in the current scene min , the starting time of the track is set to 0, and each track thereafter is evenly distributed Randomly select a time as the new starting time, and move each point in the track by the absolute value of the difference between the original starting time and the new starting time. When the new starting time is greater than the original starting time, the overall time increases by the difference; when the new starting time is less than the original starting time, the overall time decreases by the difference.

[0060] S32b: Track start and end time processing: randomly extract a random number Δt from U(0,60), and randomly set the start time and end time of each track in Z0 to obtain the start time T'0 and the end time T' of each track respectively. E ; The new start time is the original start time plus the random number Δt, and the new end time is the original end time minus the random number Δt:

[0061]

[0062] S32c: Target detection probability processing: Based on the set target detection probability, all AIS tracks in the grid are randomly sampled to obtain the detection track index I1 of the source. If the number of source tracks after sampling is 0, re-sample;

[0063] S32d: Track interpolation processing: Based on the AIS true value track Z0 in the grid and the detection track index I1 of the source, for each track in the index, except for the first and last time points, the duration of the track is calculated based on the update period T of the source. s Segment the breakpoints, add random errors at each time point, and then interpolate to obtain the detection track Z1 of the source;

[0064] S32e: Track interruption processing: According to the interruption frequency and interruption time range, random sampling is performed to obtain the number of interruptions and the interruption time. Under the condition of meeting the minimum duration, each track of Z1 is interrupted and decomposed into multiple track segments to obtain the source detection track Z2. When the track duration is greater than 300s, the interruption is performed. The interruption start time T b The distribution is U(0,T m ), interruption interval T I Distribution Remaining interruption time T r =T' E -(T b +T I ), if and only if

[0065]

[0066] Save the interrupted track, otherwise do not interrupt; record the correspondence between the source track and the true value track, and then randomly number all the tracks of the source to obtain their track batch number;

[0067] S33: Add system error: According to the set system deviation: e s1 ~e s2 ,,Using uniform distribution, add systematic errors to the longitude and latitude positions of each track;,Source 1 does not add systematic errors, and the systematic errors of Source 2 follow U(-0.03,-0.01) or U(0.01,0.03) with a probability of 50%, and the unit is degree;

[0068] S34: Add random error: According to the set track quality [1-15], add random error to each track longitude and latitude position according to Gaussian distribution;

[0069] S35a: Based on the longitude and latitude of each track, the speed and heading are calculated, and then the information Z3 of each track is obtained, including {batch number, source number, time, longitude, latitude, speed, heading}, and an association mapping table is generated, which is a table consisting of multiple columns of {start time - end time - true value batch number - source number - track batch number};

[0070] S36: The association mapping tables of the two sources are mixed, sorted by start time, a new track batch number is set, the batch is re-edited, and saved as an association mapping table CSV file;

[0071] S37: The track information of the two sources is mixed and sorted by time, and re-batch according to the association mapping table and saved as a source track CSV file;

[0072] Step 4: Dataset presentation and analysis

[0073] S41: Raster Visualization and Analysis

[0074] Dividing AIS tracks into a global grid is the prerequisite and foundation for the subsequent generation of interruption tracks and multi-source tracks. Since the tracks in the dataset are all extracted from grid tracks, the quality of the tracks in the grid will directly affect the quality of the generated dataset. This section visualizes the AIS data divided into a global grid, including visualization of the number of MMSIs, the number of targets, the density, and the maneuverability, and analyzes the comprehensiveness and effectiveness of the tracks within the grid.

[0075] S41: Target density visualization

[0076] As shown in formula (1), the target density of a grid reflects the proportion of the number of targets in a grid to the total number of grids, which is of great importance to the construction of the dataset. A heat map is drawn based on the target density of all grids in the global map to visualize the target density. The results are shown in the attached figure. Figure 2As shown in the figure, it can be seen that the density of targets near important port cities is higher, while the density of ocean-going targets is lower. Therefore, when testing and verifying the algorithm, you can choose dense grids or sparse grids to construct the test scene according to the requirements of the algorithm's usage scenario;

[0077] S42: Visualization of mobility

[0078] As shown in Formula 2, the target maneuverability of a grid reflects the proportion of the standard deviation of the target speed and heading in a grid to the total number of grids, which is of great importance to the construction of the dataset. A heat map is drawn based on the target maneuverability of all grids on the global map to visualize the target maneuverability. The results are shown in the attached figure. Figure 3 As shown in the figure, it can be seen that the target maneuverability is greater in complex channels and channel bends, and the target maneuverability is smaller in the straight area of ​​the channel, which can meet the construction requirements of large maneuverable target scenes;

[0079] S43: Typical scenario presentation and analysis

[0080] In order to illustrate the richness, rationality and effectiveness of the data, this section extracts a set of typical track data from the dataset for display to illustrate the effectiveness of "track coexistence time processing"; the typical scene track image is shown in the attached figure. Figure 4 As shown in the figure, the solid track is the track observed by source 1, source number is 9001; the dotted track is the track observed by source 2, source number is 9002. It can be seen that:

[0081] (1) Overall, the track motion types are rich, including various maneuvering states and various density scenarios. There are no stationary tracks, low-speed tracks, too short tracks, or jump tracks. The longitude and latitude of the center of all scenarios are (20, 30), which meets the expected setting requirements. Comparing the time-latitude image and the time-longitude image, it can be seen that the phenomenon of simultaneous and spatial track intersection exists in each scenario, which is consistent with the actual situation and proves the effectiveness of the "track coexistence time processing";

[0082] (2) In terms of interrupted tracks, there is at least one interrupted track in each scenario, and the interruption position, interruption time, interruption interval, and number of interrupted targets between the two sources are inconsistent, which proves the rationality of the track interruption setting and meets the actual requirements;

[0083] (3) In terms of multi-source tracks, by comparing the tracks of source 9001 (solid line) and source 9002 (dashed line) in the figure, it can be found that there is an obvious multi-source observation phenomenon; due to the setting of the target discovery probability, the number of tracks observed by the two sources is inconsistent, which meets the actual requirements; the inconsistency of the starting and ending points of the tracks observed by the two sources proves the effectiveness of the "track start and end time processing".

[0084] Note that the above are only preferred embodiments of the present invention and the technical principles used therein; those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that those skilled in the art can make various obvious changes, readjustments and substitutions without departing from the scope of protection of the present invention; therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for constructing a target association tracking dataset based on self-reported position data, characterized in that: The following steps are involved: Step 1: Set the scene parameters; Set the scene parameters according to the sensor source of the multi-source target track to be constructed, including scene coordinate parameters, sensor update cycle, target detection probability, track start time range, track end time range, minimum duration, interruption frequency, interruption time range, position system deviation, and track quality; Step 2 extracts the real track, including: S21: Construction of AIS basic track library; Extract the required tracks from the public AIS track data, and count the MMSI number, number of tracks, target density, mean heading variance, mean speed variance, and target maneuverability within the spatial grid. This data is used as the basic database for track simulation and saved as an AIS spatial coding index file in CSV format. S22: ground truth track extraction; There are two modes for ground truth track extraction: random extraction and conditional extraction. Random extraction involves randomly extracting spatial codes. According to the AIS spatial code index file, all MMSI numbers in the grid are obtained, and then the ground truth track is obtained. Conditional extraction involves selecting the spatial grid that is most similar to the set target density and maneuverability, or extracting from multiple similar spatial grids. Step 3: Generate source track; Based on the real tracks extracted in step 2, multiple sensor target track datasets that meet the scenario parameters set in step 1 are constructed, including: S31: According to the scene coordinate parameter settings in step 1, coordinate translation is performed on the AIS track in the grid extracted in step 2, and the true value batch number of the track is set; S32: Based on the source update cycle, target detection probability, track duration, and interruption frequency parameter settings in step 1, the AIS track is screened, interrupted, and interpolated to obtain multiple track segments, which are randomly batched. S33: According to the position system deviation setting in step 1, a systematic error is obtained by random sampling from a uniform distribution sequence, and then added to the track segment generated in S32 to obtain a simulated track with a systematic error; S34: According to the source track quality setting in step 1, random errors are obtained by random sampling from a Gaussian distribution or Rayleigh distribution sequence, and then added to the track generated in S33 to obtain a simulated track with random errors; S35: Generate a track association mapping table to record the correspondence between the simulated track and the original AIS track batch number; S36: Mix the simulated tracks of the two sources, sort them by time, and save them as a source track CSV file; S37: The association mapping tables of the two source tracks are mixed, sorted by start time, and saved as an association mapping table CSV file.

Images