Track association algorithm based on spatio-temporal characteristics

Abstract

The invention relates to the technical field of multi-target tracking data processing, and discloses a track association algorithm based on spatial-temporal characteristics, which comprises the following steps of: acquiring track data, constructing an Euclidean distance matrix between targets, screening the targets smaller than a coarse threshold, and constructing a target neighborhood topological structure on the basis; the distance measurement between sets is applied to the neighborhood topological structure, and original distance matrix elements are replaced; calculating the corrected average Hausdorff distance of the track history set; and weighting the neighborhood topological structure distance and the Hausdorff distance to obtain a correlation cost matrix, and realizing track correlation through a linear distribution algorithm. According to the method, a multi-target track association task can be effectively completed, and a good association effect and robust performance are achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of multi-target tracking data processing, and in particular relates to a track association algorithm based on spatiotemporal features. Background Art

[0002] Track association is one of the key issues in solving the situation estimation of maneuvering targets and is also the premise of multi-radar track data fusion. Its purpose is to solve the problem of whether the two tracks of two sensors represent the same target. The result of the association will directly affect the performance of the entire fusion system. However, in a battlefield environment with dense targets and complex target maneuvers, the sensors are affected by many factors such as errors, noise, and obstacles in the process of collecting target information, resulting in a significant reduction in target tracking accuracy. How to discover the hidden connections between the data from the large amount of track data collected by multiple sensors, so as to reflect the target status more quickly, more accurately, and more comprehensively is the key to the research.

[0003] In recent years, scholars at home and abroad have conducted in-depth analysis and research on track association algorithms. Among them, the most famous and effective methods in the existing literature mainly include:

[0004] 1. Track association algorithm based on time series covariance: 2018, TOKTA A, HOCAOGLU A KA fasttrack to track association algorithm by sequence processing of target state[C] / / Proceedings of 21st International Conference on InformationFusion.Cambridge: IEEE, 2018: 1280-1284. This method calculates the mean and covariance of the target in the time window and realizes track association based on Mahalanobis distance. As the number of associated tracks increases, the calculation time only increases linearly, while the association performance decreases slightly.

[0005] 2. Track association method based on distance distribution histogram: 2019, Xu Yasheng, Ding Chibiao, Ren Wenjuan, etc. Multi-feature combination track association based on histogram statistical characteristics [J]. Journal of Radar, 2019, 8(01): 25-35. This method extracts the speed difference distribution histogram features and sensor source features between tracks, effectively utilizes the overall characteristics between tracks, and has good anti-noise performance and association accuracy.

[0006] 3. Multi-target association algorithm for UAVs based on triangular topological sequence: 2022, Li X, Wu L, Niu Y, et al. Multi-target association for UAVs based on triangular topological sequence[J]. Drones, 2022, 6(5): 119. This method constructs a target adjacent triangle sequence based on the relative position of the target. In addition, a topological sequence similarity calculation method with similarity transformation invariance and a two-step optimal association method considering global target association consistency are proposed, which improves the accuracy of target tracking and track association.

[0007] 4. Asynchronous track association algorithm based on neighborhood topological structure characteristics: 2022, HH, A K.Asynchronous track-to-track association algorithm based on referencetopology feature[J].Signal,Image and Video Processing,2022:1-8. This method uses a one-step, memoryless track propagation scheme for the synchronization of sensor tracks, handles both asynchronous and synchronous sensors, and improves performance by using a generalized sub-pattern assignment metric to calculate the association cost between two neighborhood topologies.

[0008] 5. Asynchronous Anti-Bias Track-to-Track Association Algorithm Based on Nearest Neighbor Interval Average Distance for Multi-Sensor Tracking Systems[J]. Electronics, 2023, 12(11): 2413. The algorithm proposes a system error interval processing method for tracking coordinates, and then defines the nearest neighbor interval average distance between the interval coordinate data set and the interval coordinate point, and then uses the gray theory to calculate the correlation between the trajectories. Finally, combined with the Jonker-Volgenant algorithm, the classical allocation method is used for track association judgment. The algorithm maintains good performance under different target numbers and processing cycles, and has good superiority and robustness.

[0009] Therefore, based on the technical defects of the current algorithm such as not effectively utilizing spatiotemporal features and weak generalization ability, the present invention proposes a track association algorithm based on spatiotemporal features. Summary of the invention

[0010] The purpose of the present invention is to provide an algorithm for distinguishing different targets by using the spatiotemporal characteristics of tracks and improving the accuracy of track association.

[0011] The technical solution of the present invention is achieved in this way:

[0012] A track association algorithm based on spatiotemporal features comprises the following steps:

[0013] Obtain track data, construct the Euclidean distance matrix between targets, and select targets that are smaller than the coarse threshold, and then construct the target neighborhood topology structure based on this;

[0014] Apply the distance metric between sets to the above neighborhood topology and replace the original distance matrix elements;

[0015] Calculate the corrected average Hausdorff distance for the track history set;

[0016] The above neighborhood topological structure distance and Hausdorff distance are weighted to obtain the association cost matrix, and the track association is realized through the linear assignment algorithm.

[0017] As a further solution of the present invention: the method for constructing the neighborhood topological structure is specifically:

[0018] Obtain the track information data from the two sensors, use the Euclidean distance to build the distance matrix between the targets, filter out the track pairs that have no correlation through the distance matrix and set the value to INF;

[0019] A set of neighbor targets within the respective threshold range is selected as a neighborhood track set, and a neighborhood topology structure is constructed according to the current track and the corresponding respective neighborhood track sets.

[0020] As a further solution of the present invention: the process of filtering out tracks that do not have an association relationship includes:

[0021] Set a reasonable correlation threshold to roughly correlate the tracks reported by the two sensors. If the two target tracks The following conditions are met:

[0022]

[0023] If the two tracks are associated, their distance value will not be changed. Otherwise, it indicates that the distance between the two targets is too large in space, and their distance is set to INF. In the subsequent association task, the two tracks are considered to be generated by the two targets respectively.

[0024] As a further solution of the present invention: the process of constructing the neighborhood track set and the neighborhood topology structure includes:

[0025] The physical meaning of the neighborhood track set is to fall into the target to be judged The set of all target tracks in a circular area with a radius of r and a center of The formula is as follows:

[0026]

[0027] Based on the neighborhood track set, the points in the neighborhood track set are subtracted from the current target point in turn, and the set of difference vectors is the neighborhood topology structure RET i s (r), the number of elements in the set is consistent with the number of neighboring track sets, and the definition is The target neighborhood topology is:

[0028]

[0029] in, yes The kth element in .

[0030] As a further solution of the present invention: the method of replacing the original distance matrix elements with the distance metric between sets in the neighborhood topology structure is specifically:

[0031] The cosine similarity and distance similarity are used to compare the vectors in the two neighborhood topological structures. When the cosine similarity is greater than the set threshold and the Euclidean distance is less than the set threshold, the two vectors are considered to be matched.

[0032] Record the number of pairs, apply the distance metric between sets to the two neighborhood topological structures, and replace this metric distance with the elements of the corresponding position in the original distance matrix.

[0033] As a further solution of the present invention: the distance measurement formula between the sets is:

[0034]

[0035] Among them, the neighborhood topology structure of the i-th target of sensor 1 is RET1, and the number of vectors in its set is n1. The neighborhood topology structure of the j-th target of sensor 2 is RET2, and the number of vectors in its set is n2. h is a distribution matrix with a dimension of n1×n2, and its (k,l)th element h kl (k=1,2,...,n1;l=1,2,...,n2) is:

[0036]

[0037] And the following conditions are met:

[0038]

[0039] q is the number of pairings between two vector sets, calculated as follows:

[0040]

[0041] t ik is the kth element of the topological structure of the i-th target neighborhood of sensor 1;

[0042] t jl is the lth element of the topological structure of the jth target neighborhood of sensor 2;

[0043] d(t ik ,t jl ) is t ik and t jl The basic distance measurement between them can be Mahalanobis distance or Euclidean distance as appropriate;

[0044] n1+n2-2q is the total number of isolated points in the two sets;

[0045] n1+n2-q is the sum of the number of isolated points and the number of pairs in the two sets;

[0046] p is the order parameter of the distance, which is generally set to 2;

[0047] c is the penalty distance for false alarms and missed reports, which can be adjusted appropriately based on the false alarms and missed reports of the sensor.

[0048] As a further solution of the present invention: the calculation of the corrected average Hausdorff distance of the track history set is specifically:

[0049] Drawing on the ideas of the longest common subsequence and sliding window, time information is taken into consideration, and searches are performed only within the time window. The maximum mismatch value between sets is calculated instead of the average of all minimum matches. The modified average Hausdorff distance formula is as follows:

[0050]

[0051] Among them, A and B represent two track history sets, n a is the number of matching track pairs.

[0052] As a further solution of the present invention: the track association is specifically achieved as follows:

[0053] The neighborhood topology distance and Hausdorff distance are weighted to obtain the final association cost matrix. The association cost consists of two parts: the neighbor representation between the two targets and the distance representation between the track histories.

[0054] The optimal linear allocation algorithm is used to allocate the updated association cost matrix to obtain the final association result.

[0055] As a further solution of the present invention: the final association result cost ij The calculation formula is:

[0056] cost ij =λd ij +(1-λ)hausdorff ij ;

[0057] Among them, λ is the linear combination parameter, d ij Indicates calculating the neighborhood topological distance between two targets, hausdorff ij represents the Hausdorff distance between two targets, cost ij The larger the value, the smaller the similarity between the two targets. ij The smaller the value, the greater the similarity between the two targets.

[0058] Beneficial effects of the present invention:

[0059] 1. Since the traditional method based on traversal optimization of statistical distance matrix between tracks has slow solution speed and low association accuracy, the present invention proposes a track association algorithm based on spatiotemporal characteristics, which transforms the association problem between tracks into a matching problem of neighborhood topological structure, and uses the distance metric between sets to evaluate the similarity relationship between two targets, and then completes the association of targets through a simple linear allocation algorithm.

[0060] 2. The neighborhood topology structure proposed in the present invention can not only use the relative position relationship between targets as a reference for association judgment to effectively distinguish targets, but also offset the influence of system errors through relative positions.

[0061] 3. In view of the shortcoming that the Hausdorff distance is insensitive to noise, the present invention proposes a modified Hausdorff distance, which can reduce the impact of noise and reduce the probability of mismatching.

[0062] The present invention will be further described below in conjunction with the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a schematic diagram of the processing flow of the track association algorithm based on spatiotemporal characteristics of the present invention;

[0064] Figure 2 A schematic diagram of constructing a neighborhood topology structure according to the current track and the corresponding respective neighborhood track sets in the present invention;

[0065] Figure 3 It is a schematic diagram of the target matching process of the present invention;

[0066] Figure 4 A schematic diagram of calculating the one-way Hausdorff distance for the present invention;

[0067] Figure 5 The schematic diagram of the target scene of formation flight with 8 batches of targets in the present invention, where (a) is the real track and (b) is the noise track;

[0068] Figure 6 The schematic diagram of the target scene of formation flight with 20 targets in the present invention, where (a) is the real track and (b) is the noise track;

[0069] Figure 7 The schematic diagram of the target scene of formation flight with 30 targets in the present invention, where (a) is the real track and (b) is the noise track;

[0070] Figure 8 The schematic diagram of the target scene of formation flight with 50 targets in the present invention, where (a) is the real track and (b) is the noise track;

[0071] Fig. 9 The visualization results of the track association algorithm based on neighborhood topology and Hausdorff distance adopted in the present invention are shown in Figure 1, where (a) is the display diagram for a scenario with 8 batches of targets, (b) is the display diagram for a scenario with 20 batches of targets, (c) is the display diagram for a scenario with 30 batches of targets, and (d) is the display diagram for a scenario with 50 batches of targets. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0073] See attached Figure 1 The present invention discloses a track association algorithm based on spatiotemporal features, and the specific steps are as follows:

[0074] (1) Construct the neighborhood topology structure of the target.

[0075] The present invention converts the correlation problem between tracks into a neighborhood topology matching problem, finds two targets in the inter-target distance matrix that are less than a coarse threshold, calculates their target neighborhood topology structures respectively, and compares and analyzes the target to be associated with other adjacent targets to determine the relative position relationship between them. The relative position relationship between targets can be used as a reference for association judgment to effectively distinguish targets, and the influence of system errors can be offset by relative positions.

[0076] (1.1) Coarse threshold division.

[0077] The acquired track data is statistically calculated for distance, and the distance matrix between targets is constructed using Euclidean distance. Euclidean distance is used to measure the straight-line distance between two points X and Y in space. The calculation formula is as follows:

[0078]

[0079] Among them, X=(x1,y1),Y=(x2,y2).

[0080] Set a reasonable correlation gate to roughly correlate the tracks reported by the two sensors, and filter out the track pairs that have no correlation through the distance matrix. The statistical distance between them is greater than the set threshold, which means that the following conditions are met:

[0081]

[0082] It is considered that there is no association relationship between the two, and the value is set to the maximum value, narrowing the range of possible association pairs and reducing the amount of calculation for association solution.

[0083] (1.2) Neighborhood topology.

[0084] For two targets whose distance matrix between targets is smaller than the coarse threshold, the neighbor target sets within their respective threshold ranges are first selected as the neighborhood track sets. The physical meaning of the neighborhood track set is to fall into the target to be judged The set of all target tracks in a circular area with a radius of r and a center of The formula is as follows:

[0085]

[0086] On the basis of the neighborhood track set, the neighborhood topology structure is constructed according to the current track and the corresponding neighborhood track set, such as Figure 2 As shown in the figure, the neighborhood topology of target A is Similarly, the target neighborhood topology of a is The specific method is to subtract the points in the neighborhood track set from the current target point in turn to form a difference vector The collection of neighborhood topology RET i s (r), definition The target neighborhood topology is:

[0087]

[0088] in, yes The kth element in the set, the number of elements in the set is consistent with the number of neighborhood track sets.

[0089] (2) Calculate the association cost matrix.

[0090] The distance metric between sets is introduced to measure the association relationship of the neighborhood topology structure to measure the similarity between two targets. The larger the distance, the smaller the similarity. At the same time, for the historical status message of the track, the proposed spatiotemporal average Hausdorff distance is used to calculate the distance between the two tracks to measure the mismatch between the two tracks. The neighborhood topology structure distance and the spatiotemporal average Hausdorff distance are weighted to obtain the final association cost matrix. The association cost consists of two parts, namely the neighbor representation between the two targets and the distance representation between the track history. The combination of the two can make full use of the characteristics of the track situation and make it easier to distinguish different targets.

[0091] (2.1) Neighborhood topological distance.

[0092] Based on OSPA distance, the present invention proposes a new distance measurement method between sets, which eliminates the problem of OSPA distance in dealing with isolated elements. The distance measurement between sets is realized by calculating vector similarity to match the same vectors, and then according to the number of matching points, the number of unmatched points, the false alarm and missed report penalty parameters and the distance order parameters.

[0093] Since cosine similarity is more sensitive to angle changes and Euclidean distance similarity is more sensitive to vector length, vector similarity matching uses cosine similarity and distance similarity to compare vectors in two neighborhood topological structures. When the cosine similarity is greater than the set threshold and the Euclidean distance is less than the set threshold, the two vectors are considered to be matched. Figure 3 This is a schematic diagram of target matching. First, the vector angles of the elements in the set are calculated, and they are sorted in the order of increasing angles with the north direction to obtain the topological structure of the two neighborhoods. Select two vectors in the two neighborhood topological structures for matching in turn. If the match is successful, continue to match the remaining vectors and skip the vectors with smaller angle values ​​until all vectors in the two sets are traversed and the number of pairings q is recorded.

[0094] The distance measurement between sets is applied to two neighborhood topological structures. The neighborhood topological structure of the i-th target of sensor 1 is set to RET1, and the number of vectors in its set is n1. The neighborhood topological structure of the j-th target of sensor 2 is set to RET2, and the number of vectors in its set is n2. The calculation formula for measuring the distance from target i to target j is as follows:

[0095]

[0096] Where h is a distribution matrix with dimension n1×n2, and its (k,l)th element

[0097] h kl (k=1,2,...,n1;l=1,2,...,n2) is:

[0098]

[0099] And the following conditions are met:

[0100]

[0101] q is the number of pairings between two vector sets, calculated as follows:

[0102]

[0103] t ik is the kth element of the topological structure of the i-th target neighborhood of sensor 1;

[0104] t jl is the lth element of the topological structure of the jth target neighborhood of sensor 2;

[0105] d(t ik ,t jl ) is t ik and t jl The basic distance measurement between them can be Mahalanobis distance or Euclidean distance as appropriate;

[0106] n1+n2-2q is the number of isolated targets in the two sets;

[0107] n1+n2-q is the total number of targets in the two sets;

[0108] p is the order parameter of the distance, which is generally set to 2;

[0109] c is the penalty distance for false alarms and missed reports, which can be adjusted appropriately based on the false alarms and missed reports of the sensor.

[0110] (2.2) Average Hausdorff distance in space and time.

[0111] Since the track history information is a trajectory with time and space characteristics, the time information and space information of the trajectory are important features to measure whether two trajectories are similar. The time information can reflect the changing trend of the trajectory, while the space information can reflect the moving range and direction of the trajectory. In view of the above characteristics, the comprehensive use of the time and space information contained in the trajectory and the adoption of appropriate similarity measurement between trajectories can significantly distinguish the tracks generated by different target movements and improve the accuracy of track association.

[0112] The present invention uses the proposed spatiotemporal average Hausdorff distance to calculate the distance between two trajectories. Figure 4 This is an example of calculating the one-way Hausdorff distance. In view of the problem that the traditional Hausdorff distance ignores time information and is insensitive to noise, the Hausdorff distance is optimized by combining the ideas of the longest common subsequence (LCSS) and sliding window (SW). When looking for the minimum match, since the target that is not in the time window cannot become the feature of its minimum match point, it only needs to be searched within its time window, reducing the amount of calculation. At the same time, in view of the shortcoming of the Hausdorff distance being insensitive to noise, the maximum mismatch value between sets is calculated instead of the mean of all minimum matches, which can reduce the impact of noise and reduce the probability of mismatch.

[0113] The improved spatiotemporal average Hausdorff distance calculation formula is as follows:

[0114]

[0115] Among them, A and B represent two track history sets, n a is the number of matching track pairs.

[0116] (2.3)Associated costs.

[0117] The neighborhood topological distance d obtained above ij and the space-time average Hausdorff distance ij Weighted to get the final associated cost cost ij , the calculation formula is as follows:

[0118] cost ij =λd ij +(1-λ)hausdorff ij;

[0119] Among them, λ is the parameter of the linear combination setting, and the cost of the two tracks that do not satisfy the rough association in (1.1) is set to cost ij =∞.

[0120] After obtaining the complete association cost matrix cost, cost ij The value represents the similarity between target i of sensor 1 and target j of sensor 2. ij The larger the value, the smaller the similarity between the two targets. ij The smaller the value, the greater the similarity between the two targets. Through analysis, we can get the association cost cost ij It consists of two parts. The first part is represented by the neighbors between the two targets, which measures the similarity of the surrounding neighbors; the second part is represented by the distance between the track histories, which measures the similarity of the track histories. By taking the neighbors of the current target and the track history as the conditions for similarity judgment, it is possible to more accurately judge whether the two tracks are generated by the same target, so that cost ij The elements in the model are quite different, which can increase the differences between different targets and the similarities between the same targets.

[0121] (3) Complete the track association task based on the target neighborhood topology and Hausdorff distance.

[0122] After obtaining the association cost matrix, the optimal linear assignment algorithm is applied to obtain the final estimation result of track association. In addition, in order to verify the association effect of the track association algorithm based on spatiotemporal features proposed in the present invention, the track association algorithm based on spatiotemporal features is tested and compared with the nearest neighbor algorithm (NN) and the global nearest neighbor algorithm (GNN) in four simulation scenarios.

[0123] (3.1) Linear assignment algorithm.

[0124] For the association cost matrix, a global minimum cost solution is further solved by a simple linear allocation algorithm, that is, a higher association accuracy can be achieved. Optionally, the linear allocation algorithm includes the nearest neighbor method (NN), the global nearest neighbor method (GNN) using the Munkres algorithm (Kuhn-Munkres, KM), etc. In this embodiment, GNN is selected to solve the association cost matrix.

[0125] (3.2) Simulation scene setting.

[0126] To simulate the detection of targets on a two-dimensional plane, assume that the common space detected by the two sensors is a square area of ​​[0km~60km]×[0km~60km], and the sensors are located at (0km,-20km) and (60km,-20km). A group of targets with different numbers that meet the uniform distribution are generated by random generation as the real target track, and then the targets are converted into the polar coordinate system of each radar to simulate the slant range and azimuth of the return data when the real target is detected. In this coordinate system, the slant range systematic error, azimuth systematic error, slant range random error and azimuth random error are added to the sensor. Among them, the systematic error is set to 50m~200m by artificial slant range systematic error and 0.05° by azimuth systematic error. The random error satisfies the Gaussian distribution. The mean of the slant range error is set to 0m, the standard deviation is 100m, and the mean of the azimuth error is set to 0° and the standard deviation is 0.01°. This can simply simulate the track measurement value of the sensor detecting the target.

[0127] In order to simulate the situation where the number of tracks reported by the two sensors is inconsistent due to missed reports, the detection probabilities of the two sensors are set to P D =0.95. At the same time, in order to simulate the situation where the number of targets is consistent due to false alarms, the number of false alarms of each sensor follows the Poisson distribution with a mean of λ=2. Some points will be randomly added to simulate this situation. The sensor time sampling interval is set to 2s. The simulation experiment tests 150 moments. In each scanning cycle, the false alarm track randomly appears in the sensor public monitoring airspace. The four simulation scenarios are set as follows:

[0128] (1) A formation flight target scenario with 8 targets. The initial positions of the targets are: (0km, 40km), (2km, 42km), (4km, 44km), (6km, 46km), (0km, 20km), (2km, 18km), (4km, 16km), (6km, 14km). The speeds of the first 4 targets are (0.2km / s, -0.1km / s), and the speeds of the last 4 targets are (0.2km / s, 0.1km / s). Figure 5 (a) is the real track, and (b) is the noise track.

[0129] (2) The targets move in a uniform linear motion with an initial speed randomly distributed between 200 m / s and 300 m / s. The number of targets is 20 batches. Figure 6 (a) is the real track, and (b) is the noise track.

[0130] (3) The targets move in a uniform linear motion with an initial speed randomly distributed between 200 m / s and 300 m / s. The number of targets is 30 batches. Figure 7(a) is the real track, and (b) is the noise track.

[0131] (4) The targets move in a uniform linear motion with an initial speed randomly distributed between 200 m / s and 300 m / s. The number of targets is 50 batches. Figure 8 (a) is the real track, and (b) is the noise track.

[0132] For constructing the target neighborhood topology, the reference radius R is selected as 40 km. When calculating the OSPA distance, the basic distance metric uses the Euclidean distance, the metric order parameter is set to 2, the false alarm penalty cost c is set to 2 km, and the associated wave gate threshold is set to 4 km. For the linear combination of OSPA distance and Hausdorff distance, the parameter λ is set to 0.5.

[0133] (3.3) Verification and analysis of algorithm results.

[0134] In order to verify the effectiveness of the algorithm, the classic nearest neighbor track association algorithm and the track association algorithm based on the target neighborhood topology structure proposed in this invention were compared in the same scenario by repeating the experiment 50 times, as follows:

[0135] (1) Nearest Neighbor (NN) algorithm;

[0136] (2) Global Nearest Neighbor (GNN), which uses the Hungarian algorithm to achieve optimal matching when the number of targets is large;

[0137] (3) A track association algorithm based on spatiotemporal characteristics is used to solve the association cost matrix using GNN.

[0138] In order to measure the stability of the algorithm in different scenarios, the system error was changed from 50m to 200m in steps of 50m, and the accuracy of the association in different scenarios was tested. When the system error was 50m, the accuracy of each algorithm is shown in Table 1:

[0139] Table 1 Comparison of average accuracy when the system error is 50m (unit %)

[0140]

[0141] When the system error is 100m, the accuracy of each algorithm is shown in Table 2:

[0142] Table 2 Comparison of average accuracy when the system error is 100m (unit %)

[0143]

[0144]

[0145] When the system error is 150m, the accuracy of each algorithm is shown in Table 3:

[0146] Table 3 Comparison of average accuracy when the system error is 150m (unit %)

[0147]

[0148] When the system error is 200m, the accuracy of each algorithm is shown in Table 4:

[0149] Table 4 Average accuracy comparison when the system error is 200m (unit %)

[0150]

[0151] By observing the data in Tables 1 to 4, it can be seen that compared with the traditional track association algorithms, namely the global nearest neighbor algorithm and the nearest neighbor method, the track association algorithm based on spatiotemporal characteristics is less affected by the system error because it uses the relative position between targets and the Hausdorff distance between track histories as the criteria for association judgment. The increase in the system error has a significantly smaller impact on the association effect than the traditional track association algorithm, which shows the algorithm's robustness.

[0152] When the system error is 50m, 100m, 150m and 200m, and the number of targets is 8, 20, 30 and 50 batches respectively, the track association algorithm based on neighborhood topology and Hausdorff distance is visualized in four scenarios as shown below: Fig. 9 , where the circle and triangle represent the tracks of the target acquired by sensor 1 and sensor 2 respectively, and the short black line represents the association relationship obtained after the track association algorithm.

[0153] In addition, in order to verify the association solving speed of this algorithm, the time consumption of NN, GNN and the track association algorithm based on neighborhood topology and Hausdorff distance in processing association is compared (GNN and NN are used for association solving respectively). The following Table 5 is the time consumption analysis of each algorithm in solving association:

[0154] Table 5 Time consumption for calculating association relationships under different target numbers (unit: ms)

[0155]

[0156]

[0157] Compared with other association algorithms based on neighborhood topology, an additional association cost matrix needs to be calculated, which requires additional computational time. From the analysis of the above table, it can be seen that when the number of targets is small, the global nearest neighbor algorithm, namely the Hungarian algorithm, can achieve good time-consuming results, but as the number of targets continues to increase, the number of iterations required for the Hungarian algorithm will also increase, resulting in a serious increase in time consumption. Compared with the nearest neighbor allocation algorithm, since its algorithm is relatively simple, even if the time consumption of calculating the association cost matrix between targets is added, the time consumption is within a controllable range.

[0158] The following conclusions can be drawn from the above experiments:

[0159] (1) The track association algorithm based on the neighborhood topology structure can better adapt to the increase in the number of targets and dense environments. When the number of target batches increases to 50, the track association accuracy can still reach more than 90%.

[0160] (2) When the system error changes, the traditional track association algorithm is very sensitive to the system error. When the system error increases, the association effect decreases rapidly. However, the track association algorithm based on spatiotemporal characteristics is insensitive to the system error. As the system error increases, the association accuracy decreases less, which can effectively resist the influence of the system error. At the same time, when the number of targets increases, it can also effectively distinguish the targets through their relative position relationship, with a higher accuracy rate.

[0161] (3) When there are false alarms and missed targets in the scene, the traditional track association algorithm is very sensitive to scenes with inconsistent target numbers, while the track association algorithm based on spatiotemporal characteristics still has good tolerance when the number of targets does not match.

[0162] In summary, the track association algorithm based on spatiotemporal features proposed in the present invention can effectively complete the multi-target track association task and has good association effect and robustness. Thus, the purpose of the present invention is achieved.

[0163] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A track association algorithm based on spatiotemporal features, characterized in that: The steps include: Obtain track data, construct the Euclidean distance matrix between targets, and select targets that are smaller than the coarse threshold, and then construct the target neighborhood topology structure based on this; Apply the distance metric between sets to the above neighborhood topology and replace the original distance matrix elements; Calculate the corrected average Hausdorff distance for the track history set; The above neighborhood topological structure distance and Hausdorff distance are weighted to obtain the association cost matrix, and the track association is realized through the linear assignment algorithm.

2. The track association algorithm based on spatiotemporal features according to claim 1, characterized in that: The method for constructing the neighborhood topology structure is specifically as follows: Obtain the track information data from the two sensors, use the Euclidean distance to build the distance matrix between the targets, filter out the track pairs that have no correlation through the distance matrix and set the value to INF; A set of neighbor targets within the respective threshold range is selected as a neighborhood track set, and a neighborhood topology structure is constructed according to the current track and the corresponding respective neighborhood track sets.

3. The track association algorithm based on spatiotemporal features according to claim 2 is characterized in that: The process of filtering out tracks that do not have an associated relationship includes: Set a reasonable correlation threshold to roughly correlate the tracks reported by the two sensors. If the two target tracks The following conditions are met: If the two tracks are associated, their distance value will not be changed. Otherwise, it indicates that the distance between the two targets is too large in space, and their distance is set to INF. In the subsequent association task, the two tracks are considered to be generated by the two targets respectively.

4. The track association algorithm based on spatiotemporal features according to claim 3 is characterized in that: The process of constructing the neighborhood track set and the neighborhood topology structure includes: The physical meaning of the neighborhood track set is to fall into the target to be judged The set of all target tracks in a circular area with a radius of r and a center of The formula is as follows: Based on the neighborhood track set, the points in the neighborhood track set are subtracted from the current target point in turn, and the set of difference vectors is the neighborhood topology structure RET i s (r), the number of elements in the set is consistent with the number of neighboring track sets, and the definition is The target neighborhood topology is: in, yes The kth element in .

5. The track association algorithm based on spatiotemporal features according to claim 1, characterized in that: The method of using the distance metric between sets to replace the original distance matrix elements in the neighborhood topology structure is specifically: The cosine similarity and distance similarity are used to compare the vectors in the two neighborhood topological structures. When the cosine similarity is greater than the set threshold and the Euclidean distance is less than the set threshold, the two vectors are considered to be matched. Record the number of pairs, apply the distance metric between sets to the two neighborhood topological structures, and replace this metric distance with the elements of the corresponding position in the original distance matrix.

6. The track association algorithm based on spatiotemporal features according to claim 5 is characterized in that: The distance measurement formula between the sets is: Among them, the neighborhood topology structure of the i-th target of sensor 1 is RET1, and the number of vectors in its set is n1. The neighborhood topology structure of the j-th target of sensor 2 is RET2, and the number of vectors in its set is n2. h is a distribution matrix with a dimension of n1×n2, and its (k,l)th element h kl (k=1,2,...,n1;l=1,2,...,n2) is: And the following conditions are met: q is the number of pairings between two vector sets, calculated as follows: t ik is the kth element of the topological structure of the i-th target neighborhood of sensor 1; t jl is the lth element of the topological structure of the jth target neighborhood of sensor 2; d(t ik ,t jl ) is t ik and t jl The basic distance measurement between them can be Mahalanobis distance or Euclidean distance as appropriate; n1+n2-2q is the total number of isolated points in the two sets; n1+n2-q is the sum of the number of isolated points and the number of pairs in the two sets; p is the order parameter of the distance, which is generally set to 2; c is the penalty distance for false alarms and missed reports, which can be adjusted appropriately based on the false alarms and missed reports of the sensor.

7. The track association algorithm based on spatiotemporal features according to claim 1, characterized in that: The modified average Hausdorff distance of the track history set is calculated as follows: Drawing on the ideas of the longest common subsequence and sliding window, time information is taken into consideration, and searches are performed only within the time window. The maximum mismatch value between sets is calculated instead of the average of all minimum matches. The modified average Hausdorff distance formula is as follows: Among them, A and B represent two track history sets, n a is the number of matching track pairs.

8. The track association algorithm based on spatiotemporal features according to claim 1, characterized in that: The specific implementation of track association is as follows: The neighborhood topology distance and Hausdorff distance are weighted to obtain the final association cost matrix. The association cost consists of two parts: the neighbor representation between the two targets and the distance representation between the track histories. The optimal linear allocation algorithm is used to allocate the updated association cost matrix to obtain the final association result.

9. The track association algorithm based on spatiotemporal features according to claim 8, characterized in that: The final association result cost ij The calculation formula is: cost ij =λd ij +(1-λ)hausdorff ij ; Among them, λ is the linear combination parameter, d ij Indicates calculating the neighborhood topological distance between two targets, hausdorff ij represents the Hausdorff distance between two targets, cost ij The larger the value, the smaller the similarity between the two targets. ij The smaller the value, the greater the similarity between the two targets.

Images