Time asynchronous multi-source trajectory association method for dynamic target

Abstract

The invention discloses a time asynchronous multi-source trajectory association method for a dynamic target. The method specifically comprises the following steps: using a time dynamic regular cumulant of Cauchy distribution membership degrees among trajectories as quantitative similarity among the trajectories; the method comprises the following steps of: extracting partial track points of tracks of different sources, which are overlapped in time; for any track feature, solving a Cauchy distribution membership degree for each pair of track points between every two pieces of track data to measure the similarity of the two track points; feature similarity weighted summation is carried out; the similarity among all different source trajectories is calculated, and then trajectory association distribution is performed based on a system optimization principle, that is, trajectory association distribution is performed by maximizing the sum of the similarities among the different source trajectories obtained after trajectory association. According to the method, different source tracks of any dynamic target can be associated, and high association accuracy is still kept under the conditions of time asynchronization, sampling time interval isomerism and sensor random error.

Description

Technical Field

[0001] The present invention relates to target tracking technology in a multi-sensor scenario, and in particular to a time-asynchronous multi-source trajectory association method for a dynamic target. Background Art

[0002] In recent years, dynamic target tracking has become an important task in the fields of vehicle monitoring, autonomous driving, aerospace engineering, and robotics [1-2]. Since target tracking sensors have certain errors to a certain extent, in the task of multi-source trajectory association of dynamic targets, it is usually necessary to use two or even more sensors of the same or different types to track the dynamic target and fuse the multi-source trajectory data from these sensors. Among them, the process of associating multi-source trajectories from multiple sensors or data sources from the same dynamic target is usually called trajectory association, which is an important step in the data fusion of multi-sensor systems. The purpose of using multiple sensors for dynamic target tracking is to achieve accurate tracking of dynamic targets through multi-source data processing technology.

[0003] Multi-source trajectory data of dynamic targets have both spatial and temporal characteristics. In recent years, trajectory association methods based on spatial registration have made significant progress in time-synchronized multi-sensor systems with certain sensor errors, especially in trajectory association in multi-radar systems. By taking all target positions at the same time as point cloud data, trajectory point drift processing can be robustly eliminated through point cloud registration, such as non-rigid matching [3], coherent point drift [4], covariance matrix adaptive evolution strategy estimation of sensor error [5], maximum likelihood residual bias estimation registration [6], and relative coordinate assignment matrix [7]. However, due to different sensor types and configurations, the sampling intervals of multi-source trajectory data are often different, such as the use of radar tracks and AIS tracks in ship monitoring, which increases the complexity of trajectory association. Due to time asynchrony and time interval heterogeneity, it is not feasible to directly drift the positions of multiple trajectory points from different sources at the same time. Therefore, methods based on spatial registration have limited effect in scenarios with time asynchrony and time interval heterogeneity.

[0004] There are usually two types of solutions for multi-source trajectory association in scenarios with temporal asynchrony and heterogeneous time intervals. The first type of methods uses interpolation methods for temporal registration, converting temporal asynchronous data into temporal synchronous data before measuring the similarity of information such as position, speed, direction, etc. of trajectory points based on statistical or fuzzy theory methods, such as fuzzy theory-based methods [8], linear minimum variance unbiased estimation [9], membership and clustering-based methods

[10] , and grey relational degree

[11] . Temporal registration can solve the problem of slight temporal asynchrony caused by clock offset. In addition, statistical or fuzzy theory methods can utilize multiple features by simply weighting the similarity of all features. However, in the presence of severe temporal asynchrony and heterogeneous time intervals (which is common in many trajectory association scenarios), temporal registration is difficult to provide satisfactory results. The second type of methods directly utilizes the time series information and topological properties of trajectories and measures trajectory similarity based on the location information of trajectories. For example, trajectories are associated as graphs using graph neural networks

[12] , methods based on average trajectory point distances [13-14], methods based on iterative discrete sequential detection

[15] , and methods based on dynamic programming to accumulate the distance between trajectory points

[16] . Because they do not involve time processing, these methods are effective in the case of time asynchrony and heterogeneous time intervals. However, these methods ignore the valuable information carried by other trajectory features (such as speed and direction), and are therefore easily affected by random errors. It is difficult to maintain stable associations in scenarios with high trajectory complexity.

[0005] References:

[0006] [1]YUAN T, KRISHNAN K, CHEN Q, et al. 2017. Object Matching for Inter-Vehicle Communication Systems—An IMM-Based Track Association Approach WithSequential Multiple Hypothesis Test[J / OL]. IEEE Transactions on IntelligentTransportation Systems, 18(12):3501-3512.

[0007] [2]AEBERHARD M, SCHLICHTHARLE S, KAEMPCHEN N, et al. 2012. Track-to-Track Fusion With Asynchronous Sensors Using Information Matrix Fusion for Surround Environment Perception[J / OL]. IEEE Transactions on Intelligent Transportation Systems, 13(4): 1717-1726.

[0008] [3]ZHU H, WANG W, WANG C. 2016. Robust track-to-track association in the presence of sensor biases and missed detections[J / OL]. Information Fusion, 27: 33-40.

[0009] [4]ZHU H, WANG M, YUEN K V, et al. 2017. Track-to-Track Association by Coherent Point Drift[J / OL]. IEEE Signal Processing Letters, 24(5): 643-647.

[0010] [5]TOKTAA, HOCAOGLU A K. 2019. Sensor Bias Estimation for Track-to-Track Association[J / OL]. IEEE Signal Processing Letters, 26(10): 1426-1430.

[0011] [6]WANG J, ZENG Y, WEI S, etc. 2021. Multi-Sensor Track-to-Track Association and Spatial Registration Algorithm Under Incomplete Measurements[J / OL]. IEEE Transactions on Signal Processing, 69: 3337-3350.

[0012] [7]CHEN J,ZHANG Y,YANG Y,et al.2021.Multi-target Track-to-TrackAssociation Based on Relative Coordinate Assignment Matrix[C / OL] / / 2021 33rdChinese Control and Decision Conference(CCDC).Kunming,China:IEEE:1255-1260[2023-09-05].

[0013] [8]AZIZ AM.2007.Fuzzy track-to-track association and track fusionapproach in distributed multisensor–multitarget multiple-attributeenvironment[J / OL].Signal Processing,87(6):1474-1492.

[0014] [9]TALEBI H,HEMMATYAR A M A.2014.Asynchronous Track-to-Track Fusionby Direct Estimation of Time of Sample in Sensor Networks[J / OL].IEEE SensorsJournal,14(1):210-217.

[0015]

[10] HONG S,PENG D,SHI Y.2018.Track-to-track association using fuzzymembership function and clustering for distributed information fusion[C / OL] / / 2018 37th Chinese Control Conference(CCC).Wuhan:IEEE:4028-4032[2023-09-05].

[0016]

[11] Liu C,Xu T,Yao T,et al.Data Association of AIS and Radar Based onMulti-factor Fuzzy Judgment and Gray Correlation Grade[C] / / Internationalconference on communications,signal processing,and systems.2019.DOI:10.1007 / 978-981-10-6571-2_158.

[0017]

[12] YANG Y, YANG F, SUN L, et al. 2022. Multi-target association algorithm of AIS-radar tracks using graph matching-based deep neural network [J / OL]. Ocean Engineering, 266: 112208.

[0018]

[13] Yi X, Zeng R. 2022. Asynchronous track-to-track association algorithm based on k means distance of nearest neighbors[J / OL]. XiTong Gong Cheng Yu Dian Zi Ji Shu / Systems Engineering and Electronics, 44(11): 3515-3521.

[0019]

[14] XIAO Y,JIANYUE H,XIN G.2015.An Asynchronous Track-to-trackAssociation Algorithm without Time Alignment[J / OL].Procedia Engineering,99:1120-1125.

[0020]

[15] Guan X, Guo Jianen. Anti-bias track association algorithm based on sequential detection of iterative discrete degree[J]. Systems Engineering and Electronics, 2022(008):044. DOI:10.12305 / j.issn.1001-506X.2022.08.14. [GUAN X, GUO J.2022. Anti-bias track association algorithm based on sequential detection of iterative discrete degree[J].Xi Tong Gong Cheng Yu Dian Zi Ji Shu / Systems Engineering and Electronics, 44(8):2498-2505.

[0021]

[16] YANTING Y,YAN L,YANBO Y,et al.2015.Asynchronous track-to-trackassociation algorithm based on dynamic time warping distance[C / OL] / / 2015 34thChinese Control Conference(CCC).Hangzhou,China:IEEE:4772-4777[2023-08-14].

[0022]

[17] CUI Y, XU P, GONG C, et al. 2023. Multisource Track Association Dataset Based on the Global AIS[J / OL]. Journal of Electronics and Information Technology, 45(2): 746-756. https: / / doi.org / 10.11999 / JEIT221202. Summary of the invention

[0023] For the problem that trajectory data with time-asynchronous and heterogeneous sampling time intervals cannot be simultaneously associated with all trajectories through trajectory point drift, the present invention provides a time-asynchronous multi-source trajectory association method for dynamic targets.

[0024] A time-asynchronous multi-source trajectory association method for a dynamic target of the present invention is specifically as follows:

[0025] For two trajectories A and B from different sources, firstly, the part of trajectory points that overlap in time are extracted.

[0026] a i and b j are respectively a track point of the temporally overlapping portion of track A and track B.

[0027] a i ∈Ai=1,2,…,m

[0028] b j ∈Bj=1,2,…,n

[0029] For any trajectory feature, each pair of trajectory points a between every two trajectory data i and b j , we can find a Cauchy distribution membership ξ i.j To measure the similarity of these two trajectory points:

[0030]

[0031] Among them, δ i,j is the trajectory point a i and b j The difference of the trajectory characteristics; λ is the spread of the trajectory characteristics.

[0032]

[0033] DTW(A,B) represents the temporal dynamic warping accumulation of the feature difference between trajectory A and trajectory B.

[0034] Construct a matrix E = (ε i,j ) (m+1)×(n+1) ,

[0035]

[0036] Among them, the element ε i,j =min(ε i-1,j-1 ,ε i,j-1 ,ε i-1,j )+δ i,j ; The element in the lower right corner is the time dynamic warping accumulative amount of the feature difference between trajectory A and trajectory B, that is, DTW(A,B)=ε m,n .

[0037] The similarity of this feature between two trajectories is obtained by accumulating the Cauchy distribution membership between all pairs of trajectory points using time dynamics warping.

[0038] Construct a matrix E′=(ε′ i,j ) (m+1)×(n+1) ,

[0039]

[0040] Among them, the element ε′ i,j =max(ε′ i-1,j-1 ,ε′ i,j-1 ,ε′ i-1,j )+ξ(a i ,b j ); the element in the lower right corner is the time-dynamic regularized accumulative amount of the Cauchy distribution membership of the feature of trajectory A and trajectory B, that is, the quantitative similarity of the feature of trajectory A and trajectory B.

[0041] Position feature similarity Speed ​​feature similarity Directional feature similarity By weighted summation, we get the similarity M between trajectory A and trajectory B based on multiple factors. A,B .

[0042] After calculating the similarities between all different source trajectories, trajectory association allocation is then performed based on the system optimal principle, that is, trajectory association allocation is performed by maximizing the sum of similarities between different source trajectories obtained after trajectory association.

[0043] max S⊙V

[0044] The constraints are:

[0045]

[0046]

[0047] Among them, S is a u×v matrix, whose elements are the comprehensive multi-factor similarity M between trajectory x and trajectory y x,y ; V is a decision variable matrix of u×v, where element v x,y The value is 0 or 1, where 1 indicates that trajectory x and trajectory y are associated; u is the number of trajectories from one trajectory source, and v is the number of trajectories from the other trajectory source.

[0048] The beneficial technical effects of the present invention are:

[0049] The present invention can associate any different source trajectories and maintain a high association accuracy in the case of time asynchrony, heterogeneous sampling time intervals, and random errors in sensors. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is the trajectory similarity measurement process of the present invention.

[0051] Figure 2 It is the association accuracy of the method of the present invention under different scenarios and error levels in the simulation experiment. DETAILED DESCRIPTION

[0052] The present invention is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0053] For trajectory data with time-asynchronous and heterogeneous sampling time intervals, it is not possible to simultaneously associate all trajectories through trajectory point drift. In this case, trajectory association can only be performed by measuring the similarity between different source trajectory pairs one by one.

[0054] A time-asynchronous multi-source trajectory association method for a dynamic target of the present invention is specifically as follows:

[0055] like Figure 1 As shown, the present invention uses the time-dynamically regularized cumulative amount of the Cauchy distribution membership between trajectories as the quantitative similarity between trajectories.

[0056] For two trajectories A and B from different sources, firstly, the part of trajectory points that overlap in time are extracted.

[0057] a i and b j are respectively a track point of the temporally overlapping portion of track A and track B.

[0058] a i ∈Ai=1,2,…,m

[0059] b j ∈Bj=1,2,…,n

[0060] For any trajectory feature (such as position, speed, direction), each pair of trajectory points a between two trajectory data i and b j , we can find a Cauchy distribution membership ξ i.j To measure the similarity of these two trajectory points:

[0061]

[0062] Among them, δ i,j is the trajectory point a i and b j The difference of the trajectory features, for example, the difference of the position feature is the distance, the difference of the speed feature is the absolute value of the speed difference of the two trajectory points, and the difference of the direction feature is the angle between the azimuths of the two trajectory points. λ is the spread of the trajectory feature.

[0063]

[0064] DTW(A,B) represents the temporal dynamic warping accumulation of the feature difference between trajectory A and trajectory B.

[0065] Construct a matrix E = (ε i,j ) (m+1)×(n+1) ,

[0066]

[0067] Among them, element ε i,j =min(ε i-1,j-1 ,ε i,j-1 ,ε i-1,j )+δ i,j ; The element in the lower right corner is the time dynamic warping accumulative amount of the feature difference between trajectory A and trajectory B, that is, DTW(A,B)=ε m,n .

[0068] The similarity of this feature between two trajectories is obtained by accumulating the Cauchy distribution membership between all pairs of trajectory points using time dynamics warping.

[0069] Construct a matrix E′=(ε′ i,j ) (m+1)×(n+1) ,

[0070]

[0071] Among them, the element ε′ i,j =max(ε′ i-1,j-1 ,v′ i,j-1 ,ε′ i-1,j )+ξ(a i ,b j ); the element in the lower right corner is the time-dynamic regularized accumulative amount of the Cauchy distribution membership of the feature of trajectory A and trajectory B, that is, the quantitative similarity of the feature of trajectory A and trajectory B.

[0072] Position feature similarity Speed ​​feature similarity Directional feature similarity By weighted summation, we get the similarity M between trajectory A and trajectory B based on multiple factors. A,B .

[0073] After calculating the similarities between all different source trajectories, trajectory association allocation is then performed based on the system optimal principle, that is, trajectory association allocation is performed by maximizing the sum of similarities between different source trajectories obtained after trajectory association.

[0074]

[0075] The constraints are:

[0076]

[0077] Among them, M x,y is the comprehensive multi-factor similarity between trajectory x and trajectory y; v x,y is a decision variable with a value of 0 or 1, where a value of 1 indicates associating trajectory x and trajectory y; u is the number of trajectories from one trajectory source, and v is the number of trajectories from another trajectory source.

[0078] In most practical scenarios, the trajectories of different sources without temporal overlap are not associated. However, if there is a trajectory tracking interruption in the actual working condition, the method of the present invention can also be used to associate the trajectories before and after the interruption.

[0079] Simulation experiment:

[0080] The present invention is applied to the correlation of radar and AIS tracks in ship detection. Both track sources have three track features: position, speed and direction. The original data of ship tracks comes from the literature

[17] . Ten scenes are randomly selected from the original data for simulation experiments. For each scene, the original track data is sampled into two different source track data, one of which randomly retains 50%-60% of the track points of the original track, and the other track source randomly retains 40%-50% of the track points of the original track. And on the generated two different source track data, random errors that obey Gaussian distribution are added to the five bytes of longitude, latitude, speed, heading and time. And the experimental data are divided into 11 different error levels from 0 to 10 according to the standard deviation of the Gaussian distribution. By using the method of the present invention and setting all feature weights to 1, the correlation results of track pairing can be obtained, and each pairing result has a track from two sampling sets. The specific correlation results of each scene and error level are as follows. Figure 2 Taking all scenarios and error levels into consideration, the proposed method achieves an association accuracy of 97.33%.

Claims

1. A method for associating time-asynchronous multi-source trajectories of dynamic targets, characterized in that: Specifically: For two trajectories A and B from different sources, firstly extract the trajectory points that overlap in time; a i and b j are respectively a trajectory point of the temporally overlapping portion of trajectory A and trajectory B; a i ∈Ai=1,2,…,m b j ∈Bj=1,2,…,n For any trajectory feature, each pair of trajectory points a between every two trajectory data i and b j , we can find a Cauchy distribution membership ξ i.j To measure the similarity of these two trajectory points: Among them, δ i,j is the trajectory point a i and b j The difference of the trajectory characteristics; λ is the spread of the trajectory characteristics; DTW(A,B) represents the temporal dynamic warping accumulation of the feature difference between trajectory A and trajectory B; Construct a matrix E = (ε i,j ) (m+1)×(n+1) , Among them, the element ε i,j =min(ε i-1,j-1 ,ε i,j-1 ,ε i-1,j )+δ i,j ; The element in the lower right corner is the time dynamic warping accumulative amount of the feature difference between trajectory A and trajectory B, that is, DTW(A,B)=ε m,n ; For the similarity of this feature between two trajectories, the Cauchy distribution membership between all pairs of trajectory points is accumulated using time dynamic regularization; Construct a matrix E′=(ε′ i,j ) (m+1)×(n+1) , Among them, the element ε′ i,j =max(ε′ i-1,j-1 ,ε′ i,j-1 ,ε′ i-1,j )+ξ(a i ,b j ); the element in the lower right corner is the time-dynamic regularization cumulant of the Cauchy distribution membership of the feature of trajectory A and trajectory B, that is, the quantitative similarity of the feature of trajectory A and trajectory B; Position feature similarity Speed ​​feature similarity Directional feature similarity By weighted summation, we can get the similarity M between trajectory A and trajectory B based on multiple factors. A,B ; After calculating the similarities between all different source trajectories, the trajectory association is then allocated based on the system optimal principle, that is, the trajectory association is allocated by maximizing the sum of similarities between different source trajectories obtained after trajectory association. The constraints are: Among them, M x,y is the comprehensive multi-factor similarity between trajectory x and trajectory y; v x,y is a decision variable with a value of 0 or 1, where a value of 1 indicates associating trajectory x and trajectory y; u is the number of trajectories from one trajectory source, and v is the number of trajectories from another trajectory source.

Images