Tracking device, tracking method, and storage medium storing tracking program

Abstract

The tracking device (1) for tracking a target (2) comprises: a prediction module (100) configured to predict a state distribution of the target (2) at a certain time k; a correction module (120) configured to define an accuracy of the state distribution predicted by the prediction module (100) by a likelihood function at at least one observation point of the target (2) observed at the certain time k, and correct the likelihood function according to a number M of the observation points; and an update module (140) configured to update the state distribution at the certain time k based on the likelihood function corrected by the correction module (120).

Description

[0001] Cross-references to related applications

[0002] This application claims priority to Japanese Application No. 2019-190909, filed on October 18, 2019, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates to tracking techniques for tracking targets. Background Technology

[0004] Previously, for example, non-patent document 1 disclosed so-called extended object tracking (EOT), which identifies and follows extended targets based on multiple observation points of the observed target.

[0005] Non-patent literature 1: K.Granstrom, M.Baum, and S.Reuter, "Extended objecttracking: Introduction, overview, and applications," Journal of Advances in Information Fusion, vol.12, no.2, 2017.

[0006] However, the number of observation points in EOT changes over time, affecting the prediction of the target's state. Therefore, the feasibility and accuracy of target tracking depend on the number of observation points. In particular, with a small number of observation points, the possibility of target tracking loss increases due to predictions based on limited observation data, thus requiring improvements in tracking techniques. Summary of the Invention

[0007] The purpose of this disclosure is to provide a technique for suppressing target tracking loss. Another objective of this disclosure is to provide a tracking method for suppressing target tracking loss. Yet another objective of this disclosure is to provide a tracking procedure for suppressing target tracking loss.

[0008] The technical means of this disclosure will be described below. Furthermore, the reference numerals in parentheses within this section indicate their correspondence with the specific units described in the detailed embodiments below, and do not limit the technical scope of this disclosure.

[0009] The first aspect of this disclosure is a tracking device for tracking a target, comprising:

[0010] The prediction department predicts the state distribution of the target at a specific moment.

[0011] The correction unit, for at least one observation point of the target observed at a specific time, defines the accuracy of the state distribution predicted by the prediction unit using a likelihood function, and corrects the likelihood function according to the number of observation points; and

[0012] The update unit updates the state distribution at a specific time based on the likelihood function corrected by the correction unit.

[0013] The second approach disclosed herein is a method for executing and tracking a target by a processor, comprising:

[0014] The prediction process involves predicting the state distribution of a target at a specific moment.

[0015] The calibration process involves defining the accuracy of the predicted state distribution using a likelihood function for at least one observation point of the target observed at a specific time, and then correcting the likelihood function based on the number of observation points; and

[0016] The update process involves updating the state distribution at a specific time step based on the corrected likelihood function.

[0017] The third aspect of this disclosure is a computer-readable storage medium storing a tracking program for tracking a target, the tracking program containing commands that cause a processor to execute.

[0018] The command includes:

[0019] The prediction process involves predicting the state distribution of a target at a specific moment.

[0020] The calibration process involves defining the accuracy of the predicted state distribution using a likelihood function for at least one observation point of the target observed at a specific time, and then correcting the likelihood function based on the number of observation points; and

[0021] The update process involves updating the state distribution at a specific time step based on the corrected likelihood function.

[0022] According to the first to third methods, for at least one observation point of the target observed at a specific time, the accuracy of the predicted state distribution at that specific time is defined by a likelihood function. Furthermore, by using a likelihood function corrected according to the number of observation points, the target's state distribution at a specific time can be correctly updated, thereby suppressing target tracking loss. Attached Figure Description

[0023] Figure 1 This is a block diagram showing the overall configuration of a tracking device according to one embodiment.

[0024] Figure 2 This is a block diagram showing the detailed configuration of a tracking device according to one embodiment.

[0025] Figure 3This is a feature diagram used to illustrate a prediction module of one embodiment.

[0026] Figure 4 This is a characteristic diagram used to illustrate a correction module of one embodiment.

[0027] Figure 5A This is a characteristic diagram used to illustrate a correction module of one embodiment.

[0028] Figure 5B This is a characteristic diagram used to illustrate a correction module of one embodiment.

[0029] Figure 6A This is a characteristic diagram used to illustrate a correction module of one embodiment.

[0030] Figure 6B This is a characteristic diagram used to illustrate a correction module of one embodiment.

[0031] Figure 7A This is a feature diagram used to illustrate an update module in one embodiment.

[0032] Figure 7B This is a feature diagram used to illustrate an update module in one embodiment.

[0033] Figure 7C This is a feature diagram used to illustrate an update module in one embodiment.

[0034] Figure 8 This is a flowchart illustrating a tracking method in one embodiment.

[0035] Figure 9 This is a table used to explain the effects of one implementation method. Detailed Implementation

[0036] Hereinafter, an embodiment will be described with reference to the accompanying drawings.

[0037] like Figure 1 The tracking device 1 of one embodiment shown implements EOT (Electronic Time-of-Tracking) to identify and track an extended target 2 based on at least one observation point of the observed target 2. For this purpose, the tracking device 1 is mounted on a vehicle 4 together with a sensing device 3. The sensing device 3 is, for example, a millimeter-wave radar or LIDAR (Light Detection and Ranging / Laser Imaging Detection and Ranging) capable of observing a set of reflection points of the target 2. The sensing device 3 repeatedly senses the target 2 and observes the set of reflection points at predetermined time intervals. Therefore, in the tracking device 1, the set of reflection points observed by the sensing device 3 at a specific time k is identified as at least one observation point of the target 2 at that specific time k.

[0038] The tracking device 1 is connected to the sensing device 3 via at least one of the following: a LAN (Local Area Network), a wiring harness, and an internal bus. The tracking device 1 is a dedicated computer comprising at least one memory 10 and one processor 12. The memory 10 is a non-transitory tangible storage medium, such as semiconductor memory, magnetic media, and optical media, that non-transitorily stores or stores programs and data that can be read by a computer. The processor 12 includes, for example, at least one of the following as its core: a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and a RISC (Reduced Instruction Set Computer) CPU.

[0039] The processor 12 executes multiple commands contained in the tracking program stored in the memory 10, including a prediction process, a correction process, and an update process. The prediction process predicts the state distribution of the target at a specific time. The correction process, for at least one observation point of the target observed at a specific time, defines the accuracy of the predicted state distribution using a likelihood function and corrects the likelihood function according to the number of observation points. The update process updates the state distribution of the target at the specific time based on the corrected likelihood function. Thus, the tracking device 1 constructs multiple functional modules for tracking the target 2. In this way, by having the processor execute multiple commands contained in the tracking program stored in the memory 10 for tracking the target 2, multiple functional modules can be constructed in the tracking device 1, suppressing target tracking loss. Figure 2 As shown, the multiple functional modules constructed by the tracking device 1 include a prediction module 100, a correction module 120, and an update module 140.

[0040] Figure 2 The prediction module 100 shown predicts the state of target 2 at a specific time k. As a premise of this prediction, if target 2 is one of the other vehicles surrounding vehicle 4 (refer to...), then... Figure 1 If X is the state space of target 2 predicted at a specific time k, then X is the state space of target 2 predicted at a specific time k. k According to the following formula 1 Figure 3 The observation probability distribution of the rectangular vehicle model as target 2 is defined as shown in Equation 1. k y k These represent the horizontal and vertical positions on an orthogonal coordinate system, respectively. In Equation 1, θ k s k ωk These are the direction, velocity, and yaw rate of target 2, respectively. In Equation 1, l k b k These are the length and width of target 2.

[0041] [Formula 1]

[0042] X k =(x k y k θ k s k ω k , l k b k ) T …Formula 1

[0043] On the other hand, if the index of each independent observation point of target 2 identified at a specific time k is set to m, then this index m can be defined using the following Equation 2, which is the total number of observation points M. Furthermore, if the observation state of each observation point of target 2 at a specific time k is set to z... k m Then, the state space Z of target 2 observed at a specific time k can be represented by the following equation 3. k .

[0044] [Equation 2]

[0045] m=1~M…Equation 2

[0046] [Formula 3]

[0047]

[0048] Prediction module 100 uses equations 1-3 as a premise and predicts the state distribution F(X) of target 2 at a specific time k by using the first particle filter. k |Z k Specifically, if the index of the identification sampling point in the Monte Carlo method is set to i, then this index i can be defined by the following Equation 4 using the number of particles I that constitute the total number of sampling points. Furthermore, the state distribution F(X) of target 2 predicted at a specific time k can be predicted using the following Equation 5. k |Z k In Equation 5, W k i It is a weighted function whose summation value is 1 for the sampling points from i = 1 to I. In Equation 5, δ(X) k -X k i ) is the state space X of target 2 at a specific time k. k The state space X of each sampling point of target 2 at a specific time k ki The delta function. In the processing of prediction module 100, W is calculated by update module 140 at the previous time k-1 relative to a specific time k. k－1 i As for W in Equation 5 k i Being substituted.

[0049] [Formula 4]

[0050] i=1~I…Formula 4

[0051] [Formula 5]

[0052]

[0053] exist Figure 2 In the processing of the correction module 120 shown, it determines the state distribution F(X) at a specific time k predicted by the prediction module 100 for at least one observation point observed by the sensing device 3 at a specific time k. k |Z k The accuracy of the observation (z) is improved. Therefore, in the processing of the correction module 120, at least one of various correlation processes is used to adjust the observation state z of each observation point. k m With the predicted state distribution F(X) k |Z k The state space X contained in ) k Establish a connection.

[0054] After the correlation is established through this correlation processing, in the processing of the correction module 120, a single likelihood function p(z) of the Gaussian mixture model is used. k m |X k Define the state distribution F(X) for each observation point. k |Z k The accuracy of ). Specifically, if the index of the probability density function to be identified in the Gaussian mixture model is set as j, then this index j can be defined by the following Equation 6 using the total number J of the probability density function. Moreover, for the observation state z of each observation point at a specific time k. k m It can be obtained through a single likelihood function p(z) as shown in Equation 7 below. k m |X k F(X) represents the state distribution at a specific time k. k |Z k The state space X contained in ) k The predictive reliability. In Equation 7, w j (X k ) depends on the state distribution F(X) k|Z k The state space X contained in ) k And w for j = 1 to J j (X k A weighted function whose sum is 1. In Equation 7... It is the average μ j And the observed state z in a Gaussian distribution with covariance R as a parameter k m The probability density function.

[0055] [Formula 6]

[0056] j = 1 ~ J…Equation 6

[0057] [Formula 7]

[0058]

[0059] here Figure 4 This shows an ideal distribution with an infinite number of observation points M, representing the predicted state distribution F(X) at a specific time k. k |Z k In contrast, Figure 5A This is a characteristic plot of the observation probability distribution of a rectangular vehicle model with two overlapping observation points. Figure 6A This is a characteristic diagram of the observation probability distribution of a rectangular vehicle model with ten overlapping observation points. Figure 5B F(X) is the state distribution of a rectangular vehicle model predicted at a specific time k when there are two observation points. k |Z k ). Figure 6B F(X) is the state distribution of a rectangular vehicle model predicted at a specific time k when there are ten observation points. k |Z k ). Figure 5B , 6B State distribution F(X) k |Z k ) indicates that it is expected to be more than Figure 4 The ideal distribution shown deteriorates. Furthermore, according to... Figures 4-6B It can be seen that the fewer the number of observation points M, the lower the predicted state distribution F(X). k |Z k There are concerns that the degree of degradation relative to the ideal distribution is increasing.

[0060] Therefore, in the processing of the correction module 120, the individual likelihood function p(z) of each observation point at a specific time k is corrected according to the number of observation points M. k m |X kSpecifically, in the processing of the correction module 120, the following equation 8 is used to apply the correction function η(M) corresponding to the number of observation points M, which is a single likelihood function p(z) in the above equation 7. k m |X k The covariance R, one of the parameters of the Gaussian distribution, is used for permutation correction. The correction function η(M) is pre-set, for example through simulation or experimentation, to output a value that increases the covariance R as the number of observation points M decreases. Equation 9 below shows the classification of target 2 into the other vehicles mentioned above (refer to...) based on the number of observation points M. Figure 1 This is an example of the output of the correction function η(M) under the condition of M. By dividing the output of the correction function η(M) according to the number of observation points M as in Equation 9, in this embodiment, the value of the covariance R corrected by permutation according to Equation 8 is also divided according to the number of observation points M. Using such a correction function η(M), in the processing of the correction module 120, a single likelihood function p(z) of the covariance R of the Gaussian distribution corrected according to the number of observation points M is calculated. k m |X k ).

[0061] [Formula 8]

[0062] R = η(M) · R…Equation 8

[0063] [Formula 9]

[0064]

[0065] In the processing of the correction module 120, the individual likelihood functions p(z) of each observation point after correction are also used. k m |X k To define the global likelihood function L(Z), k |X k Furthermore, through the overall likelihood function L(Z) k |X k Define the state distribution F(X) for all observation points observed at a specific time k. k |Z k The accuracy of ) is determined by the overall likelihood function L(Z) as shown in Equation 10 below. Specifically, in the processing of the correction module 120, the accuracy is determined by the overall likelihood function L(Z) as shown in Equation 10 below. k |X k ) Calculate the state distribution F(X) for all observation points observed at a specific time k. k |Z k The accuracy and reliability of the predictions are determined by the individual likelihood functions p(z) for each observation point. k m |Xk The overall likelihood function L(Z) k |X k It also becomes a function that has been corrected based on the number of observation points M.

[0066] [Formula 10]

[0067]

[0068] exist Figure 2 In the processing of the update module 140 shown, the overall likelihood function L(Z) is corrected by the processing of the correction module 120. k |X k The state distribution F(X) predicted at a specific time k by the prediction module 100 is updated. k |Z k Specifically, in the processing of update module 140, a second particle filter with a different computational processing method than the first particle filter used in the processing of prediction module 100 is used to update the predicted state distribution F(X) at a specific time k. k |Z k ).

[0069] Specifically, in the processing of update module 140, the overall likelihood function L(Z) calculated by Equation 10 above will be used. k |X k The substitution is the global likelihood function L(Z) for each sampling point. k |X k i In the processing of update module 140, the replaced function L(Z) is also made using the following equation 11. k |X k i The updated state distribution F(X) is reflected in Equation 5 above. k |Z k At this point, in the processing of update module 140, the weighting function W calculated by update module 140 at the previous time k-1 is... k－1 i As in equation 11, W k－1 i Being substituted.

[0070] [Equation 11]

[0071]

[0072] here Figure 7A , 7BTables 7C and 7C show the comparison results between this embodiment, which underwent correction using Equations 8 and 9 above, and a comparative example, where the number of observation points M is 1, 2, and 5, respectively. Furthermore, in Figure 7A , 7B In 7C, the state distribution F(X) at a specific time k was updated. k |Z k The state distribution F(X) after that k |Z k The degree of degradation was compared. Figure 7A , 7B As shown in 7C, the same vertical axis value represents the state distribution F(X). k |Z k The smaller the horizontal axis value of the degradation level, the more accurate the prediction. According to... Figure 7A , 7B As can be seen from 7C, even if the number of observation points M is reduced, compared to the comparative example, this embodiment can still improve the overall likelihood function L(Z) corresponding to the number of observation points M by correcting the overall likelihood function L(Z) of the comparative example. k |X k And perform state distribution F(X) for all observation points. k |Z k The update of ) is used to achieve accurate predictions.

[0073] Based on the above description, in this embodiment, the prediction module 100 is equivalent to a "prediction unit", the correction module 120 is equivalent to a "correction unit", and the update module 140 is equivalent to an "update unit".

[0074] The following is based on Figure 8 The process flow of the tracking method in which the tracking device 1 tracks the target 2 through the cooperation of functional modules 100, 120, and 140 will be described. Furthermore, this process begins during the tracking period from the last time k-1 to the specific time k. Additionally, in this process, "S" refers to a step of this process that executes the commands contained in the tracking program stored in memory 10.

[0075] In the processing of S101, before a specific time k, the state distribution F(X) of target 2 at that specific time k is predicted in the processing of prediction module 100. k |Z k Specifically, the state distribution F(X) of the previous time step k-1 is updated by satisfying the processing of S103 in the previous process. k－1 |Z k－1 The weighting function W) k－1 i Used in the first particle filter to predict the state distribution F(X) k |Z k ).

[0076] In the processing of S102, the state distribution F(X) predicted by the processing of S101 through this process is... k |Z k In the processing of the correction module 120, the individual likelihood function p(z) of each observation point observed at a specific time k is corrected. k m |X k ), and the global likelihood function L(Z) of all observation points observed at a specific time k. k |X k In particular, by reducing the number of observation points M at a specific time k, p(z) becomes a single likelihood function. k m |X k The covariance R, one of the parameters of the Gaussian distribution followed by the observation, increases to correct the individual likelihood function p(z) based on the number of observation points M. k m |X k Then, the corrected single likelihood function p(z) for each observation point is used. k m |X k ), correcting the overall likelihood function L(Z) k |X k ).

[0077] In the processing of S103, the overall likelihood function L(Z) is corrected based on the processing of S102. k |X k In the processing of update module 140, the state distribution F(X) at a specific time k predicted by the processing of S101 is updated. k |Z k Specifically, the state distribution F(X) of the previous time step k-1 is updated by satisfying the processing of S103 in the previous process. k－1 |Z k－1 The weighting function W) k－1 i Used in the second particle filter to update the state distribution F(X) k |Z k ).

[0078] Based on the above description, in this embodiment, the process of S101 is equivalent to a "prediction process", the process of S102 is equivalent to a "correction process", and the process of S103 is equivalent to an "update process".

[0079] (Effects)

[0080] The effects of the above-described embodiment will now be explained.

[0081] According to this embodiment, for at least one observation point of target 2 observed at a specific time k, the overall likelihood function L(Z) is used. k |X k Define the predicted state distribution F(X) at a specific time k. k |Z k The accuracy of ) is then determined by the overall likelihood function L(Z) corrected for the number of observation points M. k |X k Update the state distribution F(X) of target 2 at a specific time k. k |Z k ),like Figure 9 As shown, it can suppress tracking loss. Additionally, Figure 9 The state distribution F(X) of target 2 at a specific time k is compared between this embodiment, which has been corrected using Equations 8 and 9 above, and a comparative example, which has not been corrected using Equations 8 and 9 above. k |Z k The number of tracking loss events was compared with the number of events that were lost.

[0082] In addition, since the overall likelihood function L(Z) is corrected based on the number of observation points M, k |X k The state distribution F(X) follows a Gaussian distribution with covariance R, thus reducing the number of observation points M for the state distribution F(X). k |Z k The influence of ) can be applied. Based on this, the state distribution F(X) can be improved. k |Z k The update accuracy ensures the reliability of tracking loss suppression.

[0083] In particular, in this embodiment, where a correction function is used to classify the Gaussian distribution covariance R based on the number of observation points M, the correction operation for the covariance R is relatively simple. Therefore, it is possible to suppress tracking loss while reducing the computational load required to suppress tracking loss.

[0084] In particular, according to this embodiment, the covariance R of the Gaussian distribution increases as the number of observation points M decreases. Therefore, even with a small number of observation points M, it is possible to achieve a distribution based on the appropriately corrected global likelihood function L(Z). k |X k ), improve the state distribution F(X) k |Z k This improves the update accuracy of tracking. Therefore, it can enhance the reliability of tracking loss suppression.

[0085] Furthermore, the state distribution F(X) in this embodiment k |Z kAfter prediction using the first particle filter, the results are updated using a second particle filter, which is applied through the corrected overall likelihood function L(Z). k |X k This makes the computational processing different from the first particle filter used during prediction. Therefore, it is possible to improve the efficiency of the computation from the state distribution F(X). k |Z k The reliability of the predicted update is improved, thereby enhancing the reliability of tracking loss suppression.

[0086] Furthermore, this embodiment, which can suppress tracking loss when the target 2 is observed by a millimeter-wave radar or LIDAR mounted on the vehicle 4 as a sensing device 3, is particularly effective for surrounding detection of autonomous vehicles or highly assisted driving vehicles.

[0087] (Other implementation methods)

[0088] The above describes one embodiment, but this disclosure should not be construed as being limited to this embodiment and can be applied to various embodiments without departing from the spirit of this disclosure.

[0089] The tracking device, tracking method, and tracking program of the modified examples can also be applied to vehicles. In this case, the tracking device applied to vehicles may not be mounted on or placed on the same object as the sensing device 3.

[0090] The tracking device in the variation can also be a dedicated computer comprising at least one of digital circuitry and analog circuitry as a processor. Specifically, the digital circuitry includes at least one of the following: ASIC (Application Specific Integrated Circuit), FPGA (Field Programmable Gate Array), SOC (System on a Chip), PGA (Programmable Gate Array), and CPLD (Complex Programmable Logic Device). Furthermore, such digital circuitry may also include a memory storing the program.

[0091] In a variation, the single likelihood function p(z) containing the values ​​of each observation point can also be corrected using the function η(M). k m |X k The overall likelihood function L(Z) k |X kThe covariance R of the Gaussian distribution. This is because if equations 7 and 10 described in the above implementation are rearranged, the following equation 12 is obtained.

[0092] [Equation 12]

[0093]

[0094] In a modified example, the processing based on correction module 120 (S102) and the processing based on update module 140 (S103) can also be performed in the same step of the same module. This is because if equations 10 and 11 described in the above embodiment are rearranged, the following equation 13 is obtained.

[0095] [Equation 13]

[0096]

[0097] The state distribution F(X) of the variant example k |Z k Alternatively, after making predictions using an estimation filter other than the first particle filter, the corrected overall likelihood function L(Z) can be used. k |X k This updates the estimated filter other than the second particle filter, which is different from the one used in the prediction.

Claims

1. A tracking device for tracking a target, comprising: The prediction department predicts the state distribution of the aforementioned targets at a specific time. The correction unit corrects multiple individual likelihood functions based on the number of observation points of the target observed at the specific time mentioned above, and calculates the overall likelihood function using the corrected multiple individual likelihood functions for each observation point, and corrects the overall likelihood function based on the number of observation points, wherein... The individual likelihood functions mentioned above represent the accuracy of the predicted state distribution for each of the aforementioned observation points of the aforementioned target observed at the aforementioned specific time; the overall likelihood function represents the accuracy of the predicted state distribution for all the aforementioned observation points at the aforementioned specific time; and The update unit updates the state distribution at the specific time point based on the corrected overall likelihood function. The aforementioned correction unit uses a correction function corresponding to the number of observation points to perform permutation correction on each covariance of each Gaussian distribution followed by the multiple individual likelihood functions, and calculates the multiple individual likelihood functions after correcting the covariances of each Gaussian distribution according to the number of observation points. The aforementioned correction function is set to increase the aforementioned covariances as the number of observation points decreases.

2. The tracking device according to claim 1, wherein, The aforementioned correction unit uses the aforementioned correction function, which classifies the covariances according to the number of the aforementioned observation points, to perform permutation correction on the aforementioned covariances.

3. The tracking device according to claim 1 or 2, wherein, The aforementioned prediction unit uses a first particle filter to predict the aforementioned state distribution. The aforementioned update unit updates the state distribution using a second particle filter, which has a different computational processing than the first particle filter, based on the overall likelihood function.

4. The tracking device according to claim 1 or 2, wherein, The aforementioned tracking device is a vehicle-mounted tracking device that is integrated with the sensing device for observing the aforementioned target. The aforementioned sensing devices are millimeter-wave radars or LiDARs.

5. A tracking method, executed by a processor to track a target, comprising: The prediction process forecasts the state distribution of the aforementioned target at a specific moment. The correction process involves correcting multiple individual likelihood functions based on the number of observation points of the target observed at the specific time mentioned above, and calculating the overall likelihood function using the corrected individual likelihood functions for each observation point. The overall likelihood function is then corrected based on the number of observation points. The individual likelihood functions mentioned above represent the accuracy of the predicted state distribution for each of the aforementioned observation points of the aforementioned target observed at the aforementioned specific time; the overall likelihood function represents the accuracy of the predicted state distribution for all the aforementioned observation points at the aforementioned specific time; and The update process involves updating the state distribution at the specific time point based on the corrected overall likelihood function. The above correction process uses a correction function corresponding to the number of observation points to perform permutation correction on the covariances of each Gaussian distribution followed by the multiple individual likelihood functions, and calculates the multiple individual likelihood functions after correcting the covariances of each Gaussian distribution according to the number of observation points. The aforementioned correction function is set to increase the aforementioned covariances as the number of observation points decreases.

6. The tracking method according to claim 5, wherein, The above correction process uses the correction function, which classifies the covariances according to the number of observation points, to perform permutation correction on the covariances.

7. The tracking method according to claim 5 or 6, wherein, The above prediction process uses a first particle filter to predict the above state distribution. The above update process updates the state distribution using a second particle filter with different computational processing than the first particle filter, based on the overall likelihood function.

8. The tracking method according to claim 5 or 6, wherein, The above-described tracking method is a method for tracking the target observed by sensors mounted on the vehicle. The aforementioned sensing devices are millimeter-wave radars or LiDARs.

9. A computer-readable storage medium storing a tracking program for tracking a target, said tracking program containing commands that cause a processor to execute. The above commands include: The prediction process forecasts the state distribution of the aforementioned target at a specific moment. The correction process involves correcting multiple individual likelihood functions based on the number of observation points of the target observed at the specific time mentioned above, and calculating the overall likelihood function using the corrected individual likelihood functions for each observation point. The overall likelihood function is then corrected based on the number of observation points. The individual likelihood functions mentioned above represent the accuracy of the predicted state distribution for each of the aforementioned observation points of the aforementioned target observed at the aforementioned specific time; the overall likelihood function represents the accuracy of the predicted state distribution for all the aforementioned observation points at the aforementioned specific time; and The update process, based on the corrected overall likelihood function, updates the state distribution at the specific time point. The above correction process uses a correction function corresponding to the number of observation points to perform permutation correction on the covariances of each Gaussian distribution followed by the multiple individual likelihood functions, and calculates the multiple individual likelihood functions after correcting the covariances of each Gaussian distribution according to the number of observation points. The aforementioned correction function is set to increase the aforementioned covariances as the number of observation points decreases.

10. The computer-readable storage medium according to claim 9, wherein, The above correction process uses the correction function, which classifies the covariances according to the number of observation points, to perform permutation correction on the covariances.

11. The computer-readable storage medium according to claim 9 or 10, wherein, The above prediction process uses a first particle filter to predict the above state distribution. The above update process updates the state distribution using a second particle filter that performs different operations than the first particle filter based on the overall likelihood function.

12. The computer-readable storage medium according to claim 9 or 10, wherein, The aforementioned tracking procedure is used to track the aforementioned target as observed by sensors mounted on the vehicle. The aforementioned sensing devices are millimeter-wave radars or LiDARs.

13. A computer program product comprising a computer program, wherein, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 5 to 8.

Images