A multi-target tracking method based on aerial biological motion state feature assistance

Abstract

The present application relates to a kind of multi-target tracking method based on aerial biological motion situation feature auxiliary, belong to radar tracking technical field, this method is constructed by biological three-dimensional space distribution situation matrix according to the three-dimensional space distribution of biological in radar scanning period;Then the shift of the inner product maximum of cyclic convolution calculation matrix is calculated using the three-dimensional situation matrix of two successive scanning periods of radar, biological motion velocity estimation is obtained according to the shift;Finally, the estimated motion situation information and track head position information are combined to predict target state, and the correlation gate is established to assist the correlation tracking of biological target using the predicted position information.In the embodiment, compared with traditional tracking algorithm using simulation experimental data, the effectiveness of the proposed method is verified.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of multi-target tracking method based on aerial biological motion situation feature auxiliary, belong to radar tracking technical field, especially to a kind of multi-target tracking method based on aerial biological motion situation feature (i.e. migration biological group directional motion speed) auxiliary. BACKGROUND

[0002] Radar target tracking is the core link of radar data processing, its purpose is to realize the sustained monitoring of target, generates stable and reliable track, mainly involves track initiation algorithm, tracking filter algorithm and data association algorithm and other key technologies.

[0003] Track initiation is the primary problem to be processed in target tracking, its performance directly affects the quality of tracking track. Correct and fast track initiation helps target tracking to proceed smoothly, and also prevents the combination explosion caused by selecting too many echoes in association, reduces the amount of calculation. If track initiation is wrong, the correctness of target tracking will be out of question.

[0004] The accurate tracking of aerial biological targets is the core link of monitoring their migration trajectory. The traditional tracking algorithm only uses individual target position information to establish an association gate when initiating a track, which requires a large gate range to ensure that the highly maneuverable migration biological targets fall within the association gate. However, the number of biological targets flying at low altitude is large, and they are densely distributed, and there is a lot of clutter interference in the low-altitude flight environment. A large gate will result in an increase in the number of interference targets within it, and false association is likely to occur during tracking, affecting the performance of biological target tracking. Therefore, for densely distributed biological multi-targets, a tracking algorithm is needed that can effectively improve the association probability. SUMMARY

[0005] Therefore, the present application provides a multi-target tracking algorithm based on aerial biological motion situation feature auxiliary, which first analyzes the behavior characteristics of migration insects, birds and other biological targets flying in the air, such as gathering into layers and directional movement. The radar scanning range is divided into small grid cells by setting the grid size and grid size. According to the three-dimensional spatial distribution of biological targets within each grid cell, a biological three-dimensional spatial distribution situation matrix is constructed. Then, the maximum offset of the matrix inner product is calculated by using the three-dimensional situation matrix of the radar in two consecutive scanning periods. The estimated biological motion speed is obtained according to the offset and the radar rotation speed information. Finally, the estimated overall motion situation information of the biological target and the track head position information are combined to predict the target state, and the predicted position information is used to assist in establishing the association gate for biological target association tracking.

[0006] The technical solution of the present application is:

[0007] A multi-target tracking method based on aerial biological motion state feature assistance, the steps of the method comprising:

[0008] First, estimate the biological target motion state;

[0009] Second, according to the biological target motion state estimated in the first step, auxiliary multi-target tracking is carried out, and multi-target tracking based on aerial biological motion state feature assistance is completed.

[0010] In the first step, the motion state of the biological target is estimated using the measurement information of the biological target, and the specific method is:

[0011] (1) Scan the biological target using radar, and all measurements obtained by the Nth screen of the radar are denoted as:

[0012]

[0013] In the formula, z i (N) is the ith measurement in the Nth screen measurement, m N is the total number of measurements of the Nth screen, i=1, 2, 3, …m N ;

[0014] (2) Convert the measurement in polar coordinates obtained in step (1) to rectangular coordinates, denoted as z i =(x,y,z) i ;

[0015]

[0016] (3) The motion state of the biological target is defined as: the motion speed v=(v x ,v y ,v z ) of the target from the N-1th screen to the Nth screen in the radar observation process, that is:

[0017] Z N =Z N-1 +v·T (3)

[0018] Wherein, T is the time interval between the N-1th screen and the Nth screen, that is, the scanning period of the radar;

[0019] (4) The relationship between the two screens of measurements should be rewritten as:

[0020]

[0021] Wherein, represents the differentiation of the migratory biological multi-target, Z diff contains the error of the biological flying into and flying out of the radar airspace and the radar detection;

[0022] The method for obtaining the overall motion trend v of the biological target is as follows: first, the radar scanning range is divided into small airspace grids by setting the grid airspace range and the grid size, and the number of biological targets in each grid is counted according to the three-dimensional space distribution of the biological targets in the radar scanning period to construct a biological three-dimensional space distribution trend matrix; then, the offset amount at which the matrix inner product is maximum is calculated by using the cyclic convolution of the three-dimensional trend matrices of two consecutive radar scanning periods, and the biological motion speed estimation is obtained according to the offset amount and the radar rotation speed, which is specifically as follows:

[0023] By griding the radar scanning airspace and counting the number of biological targets in the grid, a space trend matrix Q N-1 and Q N is formed to reflect the distribution of the biological targets in the space during the radar scanning, so as to reduce the slight differences in individual motion and the radar measurement deviation and ignore the influence of the same on the estimation result. Thus, the overall position offset between the measurement set Z N-1 and Z N is converted into the grid offset between the space trend matrices Q N-1 and Q N . Obviously, the conversion effectively avoids the influence of some errors, but since there is a difference in the number of targets detected by the two screens of the radar, the grid offset cannot be directly obtained from the statistical results of Q N-1 and Q N to estimate the overall motion trend. Therefore, in this section, the idea of convolution is combined to take the maximum offset of the inner product of Q N-1 after moving and Q N as the offset value estimation, and then the target motion trend is obtained.

[0024] The method for estimating the offset value is as follows:

[0025] Since the migration of the biological targets in the airspace is continuous, in order to accurately reflect the motion of the targets between the adjacent two screens, the method of cyclic convolution is first used to expand the space of the trend matrix Q N , and the periodic edge filling is used to process the boundary. Since the biological targets have a certain speed limit during migration, it is not necessary to periodically extend the matrix, but only to determine the expansion range of each dimension of the matrix according to the estimated maximum speed of the biological motion:

[0026] (Rgx,Rgy,Rgz) t =(v x ,v y ,v z ) max ·T / Rg (6)

[0027] where Rg is the size of the grid, (Rgx,Rgy,Rgz) t is the matrix Q NThe expansion range of each dimension.

[0028] Then the matrix Q N-1 is slid on Q' N , and the biological target motion trend matrix is obtained by convolution:

[0029]

[0030] Finally, the maximum coordinate (x, y, z) of the biological motion trend matrix V s is obtained. max That is, Q N-1 moves in three dimensions, and the motion trend of the biological target in the time from the N-1 screen to the N screen can be calculated according to the radar scanning period and (x, y, z) max .

[0031] (v x , v y , v z ) = [(x, y, z) max - (x, y, z) o ]·Rg / T (8)

[0032] Wherein, (x, y, z) o represents the center coordinate of Q' N , that is, the original position of the biological target trend matrix Q N of the N screen.

[0033] In the second step, the multi-target tracking assisted by the motion trend is specifically:

[0034] First, the target position is predicted according to the biological target motion trend obtained in the first step and the track head X(N-1) position information, that is:

[0035] X(N)' = X(N-1) + v·T (9)

[0036] Then, the predicted position information X(N)' is used to assist in establishing the association gate.

[0037] The present application has the following beneficial effects:

[0038] The application is a multi-target tracking method based on aerial biological motion state feature assistance, which provides an effective means for radar aerial biological multi-target tracking. Compared with the traditional method, the method combines the estimated overall motion state of the biological target to establish a correlation gate for correlation tracking, which can effectively reduce the size of the correlation gate at the start of the track, thereby greatly reducing the number of other targets or interference falling into the initial correlation gate in the multi-target tracking process, thereby improving the accuracy of the track start and improving the precision performance of the entire tracking system. In summary, the method is more superior in radar aerial biological multi-target tracking. The application discloses a multi-target tracking algorithm based on aerial biological motion state feature assistance, which provides an effective means for aerial biological target tracking in a radar; the method constructs a biological three-dimensional space distribution state matrix according to the three-dimensional space distribution of the biological target in the radar scanning period; then the three-dimensional state matrix of the radar in two consecutive scanning periods is used to calculate the maximum inner product of the matrix, and the offset is obtained according to the offset; finally, the estimated motion state information and the track head position information are combined to predict the target state, and the predicted position information is used to assist in establishing a correlation gate for correlation tracking of the biological target. In the embodiment, simulation experimental data is used to compare with the traditional tracking algorithm, and the effectiveness of the method is verified. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 It is a multi-target tracking flowchart based on motion state feature assistance;

[0040] Figure 2 It is a biological target motion state estimation flowchart;

[0041] Figure 3 It is a track head correlation gate schematic diagram of the traditional method;

[0042] Figure 4 It is a correlation gate schematic diagram before and after motion state assistance;

[0043] Figure 5 (a) is the tracking result of the traditional method;

[0044] Figure 5 (b) is the tracking result of the proposed method;

[0045] Figure 5 (c) is the ideal track of the real target. DETAILED DESCRIPTION

[0046] The application will be described in detail below with reference to the drawings and examples. As Figure 1As shown, the motion trend feature assisted multi-target tracking algorithm adds two modules of motion trend estimation and motion trend assisted association on the basis of traditional tracking algorithm, obtains the directional flight motion trend of biological target through the motion trend estimation algorithm, and uses the overall motion trend of the biological target as prior information to assist in establishing an association gate for association tracking. The overall process can be divided into the following two parts for detailed introduction:

[0047] Biological target motion trend estimation:

[0048] The biological target motion trend estimation algorithm does not need to form a stable track, but can estimate the overall motion trend of the target by analyzing the distribution relationship of the historical measurement of the target and the measurement of the current scanning period of the radar in the airspace, and combining the convolution idea, and then provides the key group velocity prior for subsequent track initiation and association.

[0049] After detection and measurement processing, the target measurement information can be extracted from the radar echo. All measurements obtained by the radar in the Nth screen are denoted as:

[0050]

[0051] In the formula, z i (N) is the ith measurement in the Nth screen measurement set, m N is the total number of measurements in the Nth screen.

[0052] The measurement obtained by the general radar is the position observation value of the target in the polar coordinate system, that is, z i =(range,azi,ele) i . Therefore, the observation value in the polar coordinate system can be converted to the rectangular coordinate system according to (2), and is denoted as z i =(x,y,z) i .

[0053]

[0054] In the rectangular coordinate system, the overall motion trend of the biological target is defined as: the motion velocity v = (v x ,v y ,v z ) of the target from the N-1th screen to the Nth screen in the radar observation process, that is:

[0055] Z N = Z N-1 +v·T (12)

[0056] Where, T is the time interval between the N-1th screen and the Nth screen, that is, the scanning period of the radar.

[0057] The formula (12) represents an ideal state, i.e. the measurement set Z N-1 is different from Z N . In addition to the motion displacement, there is no other deviation. However, due to the fact that the actual radar scanning space range is large, there are the following three problems: ① there is a large difference in the motion of individual migratory organisms; ② there is an error in the detection and measurement of the target by the radar in actual operation; ③ migratory organisms continuously fly into and fly out of the radar scanning space, so that in most cases Z N-1 is different from Z N . Therefore, in the actual situation, the relationship between the two screen measurements should be rewritten as:

[0058]

[0059] wherein, represents the difference of the multi-target of migratory organisms, Z diff contains the error of the organisms flying into and flying out of the radar space and the error of the radar detection, and these unavoidable errors result in that the direct method cannot be used to obtain the overall motion situation v of the target in the entire scanning space of the radar.

[0060] In order to minimize the influence of these errors so that they can be ignored, a biological motion situation estimation method based on space distribution situation matrix is proposed. The method consists of two parts of constructing biological distribution situation matrix by space gridding and estimating motion situation by circular convolution. Firstly, the radar scanning range is divided into small space grids by setting the gridded space range and grid size, and the number of organisms in each grid is counted according to the three-dimensional space distribution of organisms in the radar scanning period to construct the three-dimensional space distribution situation matrix of organisms; then the offset amount when the matrix inner product is maximum is calculated by circular convolution of the three-dimensional situation matrices of two consecutive scanning periods of the radar, and the biological motion speed estimation is obtained according to the offset amount and the rotation speed of the radar.

[0061] By gridding the radar scanning space and counting the number of biological targets in the grid, the space situation matrix Q N-1 and Q N are formed, reflecting the distribution of biological targets in space during radar scanning, so as to reduce the slight difference in individual motion and the error in radar measurement and ignore the influence of the same on the estimation result. Therefore, the overall position offset between the measurement sets Z N-1 and Z N is converted into the grid offset between the space situation matrices Q N-1 and Q N . Obviously, after the conversion, the influence of some errors is effectively avoided, but due to the difference in the number of targets detected by the radar in two screens, the grid offset cannot be directly obtained by Q N-1 and Q NThe statistical results of the grid offset are used to estimate the overall motion trend. To this end, the idea of convolution is combined in this section to make Q N-1 After moving and Q N The inner product of the maximum offset is estimated as the offset value, and the target motion trend is obtained.

[0062] For one-dimensional discrete function f and the convolution of function g, the mathematical definition is:

[0063]

[0064] Convolution is a kind of weighted integral operation, and its essence is to obtain the output result by weighted superposition of function f and a set of weight functions g. The weight function here can represent the influence of each point in the input signal on the output result, that is, the weight. In the mathematical definition, the weight is usually associated with a certain function, which is n-x here, and can be regarded as a constraint condition. This constraint condition can be expressed as any function of the weight, and it does not necessarily need to use a linear weight function as in the common convolution operation. In different application fields, the form of the weight function can be modified to meet different needs. For example, in the field of signal processing, n-x is usually used to represent that the weight of x is related to the distance from x to n, in order to consider the influence of the decay time of the signal on the output result. In the field of image processing, the weight of each pixel is usually related to its relative position, so the weight function g can be expressed as a matrix to ensure that the value of each position is only multiplied by the corresponding weight. Based on the characteristics of convolution, the grid can be regarded as a spatial pixel, and the convolution operation can be compared with the convolution operation in image processing to obtain the final position of the target motion in time T. The specific operation is as follows:

[0065] Since the migration of organisms in space is continuous, in order to accurately reflect the motion of the target between adjacent two screens, the method of circular convolution is first used to expand the trend matrix Q N in space, and periodic edge filling is used to handle the boundary. Since the organisms have a certain speed limit during migration, it is not necessary to periodically extend the matrix, but only to determine the expansion range of each dimension of the matrix according to the estimated maximum speed of the organism motion:

[0066] (Rgx,Rgy,Rgz) t =(v x ,v y ,v z ) max ·T / Rg (15)

[0067] where Rg is the size of the grid, and (Rgx,Rgy,Rgz) t is the expansion range of the matrix Q N in each dimension.

[0068] Next, matrix Q N-1 In Q' N The biological motion state matrix is ​​obtained by sliding the convolution upwards:

[0069]

[0070] Finally, the biological motion state matrix V was calculated. s The maximum coordinates are (x, y, z). max Q N-1 The movement coordinates in three dimensions are determined by the radar scan cycle and (x, y, z). max The movement pattern of the organism during the time interval from screen N-1 to screen N can be calculated:

[0071] (v x ,v y ,v z )=[(x,y,z) max -(x,y,z) o ]·Rg / T (17)

[0072] Where (x,y,z) o Indicates Q' N The center coordinates are the biological target situation matrix Q of the Nth screen. N The original position.

[0073] Using two consecutive screens of migratory insect data, the flowchart of the biological situation estimation method is as follows. Figure 2 As shown.

[0074] Multi-target tracking with motion state assistance

[0075] like Figure 3 As shown, the traditional algorithm only uses the target position information to establish the track head (that is, the track head only contains the target position information, which is represented by the symbol X(N-1)={x,y,z}), and then establishes the correlation gate with the track head position as the center, and performs data correlation on the relevant measurements falling within the gate.

[0076] Due to the movement of the target, the gate range must be expanded to include the target in the associated gate of the trackhead. However, when tracking multiple targets, a large number of interfering targets and false alarm points will exist in the large-range associated gate, making it very difficult to correctly associate the target. Therefore, traditional algorithms often cannot meet the tracking performance requirements.

[0077] To solve the problem of traditional algorithm, a multi-target tracking algorithm based on biological motion state auxiliary is proposed by using the above motion state estimation result. Compared with the traditional tracking algorithm, the algorithm uses the characteristics of biological directional motion, first combines the overall motion state information of the biological target and the track head X(N-1) position information to predict the target state, that is:

[0078] X(N)' = X(N-1) + v · T (18)

[0079] Then, the predicted position information X(N)' is used to assist in establishing the association gate, as shown in Figure 4

[0080] The following will illustrate the implementation steps with specific examples:

[0081] To verify the effectiveness of the tracking method described above, the simulation experiment data is used to complete the detection of the aerial biological target by using the multi-target tracking algorithm based on the motion state characteristics of the aerial biological motion state auxiliary. The simulation parameters are shown in Table 1:

[0082] Table 1 Simulation parameters

[0083]

[0084] The 2 / 3 logic method is used to quickly start the track in the track initiation stage, and the nearest neighbor method is used for track and plot matching in the data association stage. Under the uniform linear motion model, the Kalman filter method is used to process 20 screen plots in the filtering module. The multi-target tracking situation before and after the motion state feature auxiliary is as follows Figure 5 , wherein (a) represents the tracking situation of the 20-screen target of the traditional motion state auxiliary tracking algorithm, (b) represents the tracking result of the tracking algorithm under the motion state auxiliary, and (c) represents the ideal real track generated by simulation. Since the unassociated track is not immediately deleted, but is updated and scored according to the scoring mechanism, the track is deleted only when the score is lower than the threshold, so there are tracks that exceed the preset radar blind area and scanning range in (a) and (b).

[0085] The tracking results are quantitatively analyzed, and the tracking performance is compared.

[0086] ​Firstly, the number of tracks in three cases is counted. The results show that: in (c), there are 2606 real tracks, which are more than the measurement number 2000 per screen, because the target motion exceeds the preset radar scanning limit and generates new point track measurements to generate tracks; in (b), there are 2996 tracks; in (a), there are 3369 tracks. Ideally, the number of tracks is the least, and the number of tracks in case (a) is more than that in case (b), which shows that there are more "short tracks" in (a) than in (b). Therefore, it can be seen that compared with the traditional algorithm, the tracking effect of the tracking algorithm assisted by motion situation has been greatly improved, and more accurate target track number can be obtained.

[0087] Next, the tracking performance indicators after 100,000 times of Monte Carlo simulation are counted, and the evaluation indicators are defined as follows:

[0088] (1) Real track: if the first confirmed track (including the terminated confirmed track) recovered by the association tracking algorithm satisfies formula (3.1), that is, the number of frames in which the position root mean square error is within the set maximum error tolerance is not less than frames, the first confirmed track is considered to be the real track of the target.

[0089]

[0090] In the formula, x i,l , y i,l , z i,l represents the real position of the target in the lth frame in the ith Monte Carlo simulation, l=1,2...L represents that the target exists in the lth frame.

[0091] (2) Number of false tracks (NCFT): for the confirmed track that does not satisfy the real track judgment condition of formula (3.1), it is judged as a false track, and the number of false tracks in the whole tracking process is counted, and the evaluation index is obtained through Monte Carlo experiment.

[0092] (3) Success tracking probability (Ps): for single target tracking, the success of recovering real track in the experiment is defined as successful tracking of the target in this experiment; for multi-target tracking, the proportion of the number of real tracks recovered in the experiment to the total number of real tracks is defined, and then the index is obtained by statistical average of Monte Carlo experiment, which is represented by symbol Ps in this paper.

[0093] (4) False track per unit time (False-per): the average number of false tracks per frame in Monte Carlo statistical experiment, which is represented by False-per in this paper, and False-per is calculated as follows:

[0094]

[0095] Wherein, L represents the total number of frames of target appearance, N t The number of false tracks in the t-th frame.

[0096] (5) False_dec: the ratio of the number of false tracks per unit time in the tracks recovered by algorithm 1 to that by algorithm 2, denoted by False_dec in this paper:

[0097]

[0098] Wherein, False_per1 and False_per2 respectively represent the number of false tracks per unit time in the tracks recovered by algorithm 1 and algorithm 2.

[0099] The tracking performance indicators after the statistical simulation experiment are shown in Table 2.

[0100] Table 2 Comparison of tracking algorithm performance

[0101]

[0102] It can be seen that the number of successfully tracked targets is significantly increased by using the motion trend assisted tracking algorithm, and compared with the traditional algorithm, the number of real tracks is increased by 61.5%; the number of false tracks is significantly reduced, and the reduction rate per unit time is increased by 73%, verifying the effectiveness of the algorithm.

[0103] The method of the application can be applied to realize efficient tracking of multiple targets of migratory organisms in the air by radar.

[0104] To sum up, the above is only a preferred embodiment of the application, and is not used to limit the protection scope of the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for multi-target tracking based on aerial biological motion pattern feature assistance, characterized in that The steps of the method include: The first step is to estimate the motion state of the biological target by using the measurement information of the biological target; The second step is to assist in multi-target tracking according to the motion state of the biological target estimated in the first step, and complete multi-target tracking assisted by the motion state characteristics of the biological target in the air; In the first step, the motion situation of the biological target is estimated The specific method is that: (1) Scan the biological target by using the radar to obtain all measurements of the biological target on the Nth screen of the radar; (2) Convert the measurements obtained in step (1) to the rectangular coordinate system; (3) Divide the scanning range of the radar into a small grid in the space by setting the grid space range and the grid size, and count the number of biological targets in each grid according to the three-dimensional space distribution of the biological targets in the radar scanning period to construct a two-screen biological target three-dimensional space distribution state matrix; (4) Calculate the offset amount when the matrix inner product is maximum by using the cyclic convolution of the biological target three-dimensional space distribution state matrix in two consecutive scanning periods of the radar, and obtain the motion state of the biological target according to the offset amount and the rotation speed of the radar; In the step (3), the biological target three-dimensional space distribution situation matrix After moving The maximum offset value of the inner product as the measurement set And The overall position offset between The method for obtaining the offset value is: (I): the method of cyclic convolution is used to expand the situation matrix in the spatial domain to form a matrix , and the specific method is to use periodic edge filling to process the boundary, and then determine the expansion range of each dimension of the matrix according to the estimated maximum speed of biological motion: wherein the size of the grid is set, the matrix the extent of the expansion in each dimension; (II) matrix In Sliding convolution, get biological target movement posture matrix: (III) finding the maximum coordinate of the biological motion posture matrix in three-dimensional movement coordinates, according to the radar scanning period and the first screen to the first screen time of the biological target motion posture:​​ wherein, denotes the center coordinates of the Nth screen, i.e. the screen biological target situation matrix the original position of the Nth screen, is the time interval between the N-1th screen and the Nth screen, i.e. the scanning period of the radar.

2. The multi-target tracking method assisted by the motion state characteristics of the biological target in the air according to claim 1, characterized in that: In step (1), all measurements of the biological target on the Nth screen of the radar are recorded as: In the formula, is the i-th measurement in the N-th screen measurement set, specifically , , is the total number of measurements in the N-th screen, i = 1, 2, 3, … .

3. The multi-target tracking method assisted by the motion state characteristics of the biological target in the air according to claim 2, characterized in that: In the step (2), the measurement in the rectangular coordinates is denoted as ; 。 4. The multi-target tracking method assisted by the motion state characteristics of the biological target in the air according to claim 1, characterized in that: In the second step, the multi-target tracking assisted by the motion state is specifically: First, according to the motion situation and the track head of the biological target obtained in the first step Position information predicts the target position That is: The predicted position information is then used The auxiliary association gate is used to assist in association tracking.

Images