Multi-side unit cross-domain target fusion tracking system and method in foggy day scene

Abstract

The invention discloses a multi-side unit cross-domain target fusion tracking system and method in a foggy day scene. The system is composed of a target detection module, a fusion tracking module and a communication module. The target detection module preprocesses original point cloud, including background filtering and ground segmentation to obtain non-ground point cloud containing noise, then filters out fog noise point cloud by adopting a point cloud space-time sequence denoising method, and extracts a target through clustering; the fusion tracking module carries out space-time fusion on the measurement sets published by the adjacent road side units and received by the road target measurement and communication module to obtain an augmented target measurement set, and a target trajectory is updated through a PHD filter; the communication module is used for receiving and sending the target measurement set between the road side units and sending the road traffic target information to the vehicle. According to the invention, continuous and stable tracking of the road target in the foggy day is realized through point cloud space-time sequence denoising and multi-road-side unit cross-domain target fusion tracking, and the safety of vehicle driving in the foggy day is enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of automation and computer technology, and in particular relates to a multi-road side unit cross-domain target fusion tracking system and method in foggy scenes. Background Art

[0002] Roadside perception is an important part of the vehicle-road cooperative system, and target tracking technology is the key to ensure that the roadside perception system can operate efficiently and safely. Especially in adverse weather conditions such as fog, improving the accuracy and precision of target tracking has great practical significance for roadside environmental perception performance.

[0003] Chinese patent application: Denoising method and device for foggy point cloud for unmanned driving control system (application number: CN202410239016.1) discloses a method for denoising foggy point cloud, which uses different point cloud intensity rules to cluster according to the distance of point cloud data from the center point, and determines its label (real point cloud or noise point cloud) by analyzing the centroid point and point attributes contained in each cluster, and removes the clusters marked as noise from the data. This method only considers the intensity value of foggy point cloud at the same distance at the instantaneous moment, and does not consider the point cloud information in the time dimension, which can easily cause normal point clouds to be misdetected as foggy point clouds.

[0004] Chinese patent application: A target trajectory fusion method based on multi-point roadside sensors (application number: CN202410009145.1) discloses a target trajectory fusion system based on multi-point roadside sensors. Multiple roadside devices cache point perception information in a central server, and perform trajectory matching and data fusion after time alignment. The system uses a centralized server to process fused data and only uses the point cloud at a certain instant, without considering the contextual features of the point cloud.

[0005] Chinese patent application: A target tracking system, method and storage medium for roadside unit collaboration (application number: CN202210171778.3) discloses a target tracking system for multiple roadside units, which uses adjacent roadside units to share target trajectory information and reduce the judgment of new target tracking start. However, it does not use the target detection data of adjacent roadside units. Summary of the invention

[0006] The present invention aims to solve the above problems of the prior art. A multi-road side unit cross-domain target fusion tracking system and method in foggy scenes are proposed. The technical solution of the present invention is as follows:

[0007] A multi-road side unit cross-domain target fusion tracking system in foggy scenes, comprising:

[0008] Target detection module, fusion tracking module and communication module. The target detection module is used to preprocess the original point cloud. The preprocessing includes background filtering and ground segmentation to obtain non-ground point cloud containing noise, and then the point cloud spatiotemporal sequence denoising method is used to filter out the fog noise point cloud, and then the target is extracted by clustering;

[0009] The fusion tracking module is used to perform spatiotemporal fusion of the road target measurement and the measurement set released by the adjacent roadside unit received by the communication module to obtain an augmented target measurement set, and then update the target trajectory through probability hypothesis density filtering PHD.

[0010] The communication module is used to receive and send target measurement sets between roadside units and send road target information to vehicles.

[0011] A tracking method based on the system comprises the following steps:

[0012] 2.1 Point cloud collection and preprocessing: The roadside laser radar collects the surrounding environment point cloud data at the current moment, and then preprocesses the collected raw point cloud data, including background filtering and ground segmentation, to obtain the non-ground point cloud P containing noise. N ;

[0013] 2.2 Point cloud spatiotemporal sequence denoising: Continuous multi-frame point cloud data are superimposed and assigned to different voxel blocks according to the point cloud index, and then the point cloud intensity of all point clouds is compensated. The Gaussian convolution kernel is used to traverse all voxel blocks to calculate the smooth intensity value of the central voxel and update it to the voxel block and normalize it. The normalized smooth intensity value of the central voxel I n Compared with the given threshold T, the denoised point cloud set P′ is obtained N ;

[0014] 2.3 Target Extraction: De-noised Point Cloud P′ N Perform clustering and extract the target to the observation set Z k And broadcast to adjacent roadside units;

[0015] 2.4 Measurement set transmission: The roadside unit transmits the target measurement set to the adjacent roadside unit;

[0016] 2.5 Measurement set reception: The roadside unit receives the target measurement set of the adjacent node;

[0017] 2.6 Spatiotemporal fusion: First, create a suitable sliding time window according to the time series, put the target measurements of other nodes that meet the time series into the window, and convert the target measurements in the window to the current roadside lidar coordinate system according to the relative position of the roadside equipment, and fuse them with the own observation data to obtain the augmented target measurement set Z′ k ;

[0018] 2.7 Target trajectory extraction: The roadside unit predicts the existing target strength based on the state transition model and the target survival probability, introduces the new target strength, and then updates the target strength through observation data. The predicted strength is combined with the measurement information to generate the posterior strength. Finally, the target state estimate is extracted through the updated strength function.

[0019] 2.8 Target status transmission: The roadside unit sends road target information to the vehicle;

[0020] 2.9 Repeat steps 2.1 to 2.8;

[0021] Furthermore, the step 2.2 of the spatiotemporal sequence point cloud denoising method specifically includes:

[0022] 3.1 Point cloud integration: In the time series, statistics [t k-len ~t k-1 ] and compare it with the original point cloud at the current moment Superposition is performed to calculate the new point cloud after the point cloud accumulation

[0023]

[0024] 3.2 Voxelization: Taking the roadside LiDAR coordinate system as the center, divide the voxel into a voxel of size d x ×d y ×d z For each point cloud p, it is mapped to the voxel block according to its index coordinates. The index coordinates (i, j, k) are calculated as follows:

[0025]

[0026] 3.3 Point cloud intensity compensation: According to the point cloud position and atmospheric attenuation coefficient, compensation is made for the point cloud intensity at different positions:

[0027] I i =I·d 2 ·R·e -2αd #(3)

[0028] Among them, I is the original intensity of the point cloud, d is the distance from the point cloud to the lidar, R is the reflectivity of the target object, and α is the atmospheric attenuation coefficient. Under the condition that the atmospheric visibility is V meters, the commonly used α approximate formula is expressed as:

[0029]

[0030] 3.4 Point cloud filtering: For each voxel block, first calculate the distance d between each neighborhood voxel and the central voxel i , and then according to the distance d iThe weight W of each neighborhood voxel is calculated using the Gaussian convolution kernel formula i :

[0031]

[0032] According to the weight, the intensity value I of the neighboring voxels i Perform weighted averaging to obtain the smooth intensity value I of the central voxel s :

[0033]

[0034] 3.5 Normalization: After smoothing, select the smallest point cloud intensity I in the voxel block min and the maximum point cloud intensity I max , and then the smoothing intensity value I of the central voxel s Normalize and get the normalized intensity value I n :

[0035]

[0036] Update the smooth intensity value of the central voxel to the voxel block, and then normalize the smooth intensity value I n Compared with the given threshold T, the voxels smaller than the given threshold T are removed to obtain the filtered point cloud set P′ N It can be used for target detection.

[0037] Furthermore, the target trajectory extraction method in step 2.7 comprises the following steps:

[0038] 4.1 Establishing motion model: target motion state Where x is the target x position, y is the target y position, is the target x speed, is the target y velocity, and the state transfer equation is:

[0039] X k =F k X k-1 +Γ k v k #(8)

[0040] Among them, F k is the state transfer matrix, Γ k is the noise transfer matrix, v k Yes Obedience Gaussian white noise, L k represents the variance of the noise;

[0041] Its measurement equation:

[0042] Z k =Hk X k +ω k #(9)

[0043] Among them, H k is the observation matrix, ω k Yes Obedience Gaussian white noise;

[0044] 4.2 Intensity prediction: Intensity function of multi-objective state at time k-1:

[0045]

[0046] Among them, J k-1 represents the number of Gaussian components in the Gaussian mixture model at time k-1, is the weight of the Gaussian density, It means the mean The covariance is Gaussian distribution of

[0047] Combined with the state transfer equation, the multi-target prediction strength at time k is:

[0048]

[0049] in, represents the prediction weight of the i-th target at time k, represents the predicted mean value of the i-th target at time k, represents the predicted covariance of the i-th target at time k, The predicted mean is The covariance is Gaussian distribution of

[0050] 4.3 New target strength calculation: based on measurement set Z k-1 The new target measurement set that may be included in the k-1 moment is obtained as follows:

[0051] Z B,k-1 =Z k-1 -Z S,k-1 #(15)

[0052] Among them, Z S,k-1 represents the surviving target measurement set at time k-1, then the number of new targets at time k-1 is:

[0053]

[0054] Based on the measurement set Z at k-1 time k-1 , the new target strength γ k|k-1 (x):

[0055]

[0056] Then the target strength v after the prediction step is k|k-1 (x):

[0057] v k|k-1 (x) = v S,k|k-1 (x)+γ k|k-1 (x)#(18)

[0058] 4.3 Intensity update: based on the target predicted intensity v k|k-1 (x), survival target prediction strength v S,k|k-1 (x), prediction strength of new target γ k|k-1 (x), survival target measurement set Z S,k|k-1 , new target measurement set Z B,k|k-1 , then the posterior target strength v at time k k|k (x) is:

[0059]

[0060] 4.4 Gaussian component pruning and merging: Given a pruning threshold D th , merging threshold M th , for all Gaussian components if Then delete the component from the posterior distribution;

[0061] If the distance between the two Gaussian components is less than the given merging threshold D th , then the two components are considered to represent the same target and should be merged; the merger can be achieved by calculating the weighted average of the mean and covariance of the two components:

[0062]

[0063] 4.5 Target state extraction: When the number of Gaussian components J is greater than J max , then retain the J with the largest weight max Gaussian components, the set of Gaussian components after pruning and merging can be expressed as Finally, take the sum of the weights of all Gaussian components as the target number estimate N at the current k moment k :

[0064]

[0065] The mean of the corresponding Gaussian component As the estimated position and velocity of the target, that is, the target state estimation

[0066] The advantages and beneficial effects of the present invention are as follows:

[0067] The present invention proposes a cross-domain target fusion tracking system and method for multiple roadside units in foggy scenes, which superimposes multiple frames of target point clouds and uses a spatiotemporal sequence denoising method to filter out the fog noise point clouds around the effective target to extract the effective point cloud cluster; the roadside unit predicts the existing target strength according to the state transition model and the target survival probability, and updates the target strength according to the observation data, combines the predicted strength with the measurement information, and generates a posteriori strength to extract the target state estimate. In claim 2.2, the present invention innovatively proposes a spatiotemporal sequence denoising method, which superimposes multiple frames of point clouds, divides them into different voxel blocks, and performs Gaussian convolution in sequence to filter the point clouds of fog noise; in claim 2.7, the present invention innovatively considers the coordination between multiple roadside devices, performs spatiotemporal fusion of multiple adjacent roadside lidar point clouds, and then calculates the state transition model and the target survival probability to predict the existing target strength, while introducing the new target strength, and then updates the target strength through the observation data, combines the predicted strength with the measurement information, and generates a posteriori strength. Finally, the target state estimate is extracted through the updated strength function;

[0068] The present invention utilizes the method of filtering and cross-domain fusion of multiple roadside units to improve the tracking accuracy and stability of roadside lidar targets in foggy weather, achieves continuous and stable tracking of road targets in foggy weather scenarios, and improves the safety of vehicle automatic driving or assisted driving. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 The present invention provides a preferred embodiment of an overall architecture of a multi-road side unit cross-domain target fusion tracking system in a foggy scene.

[0070] Figure 2 This is a flow chart of a spatiotemporal series denoising method described in the present invention.

[0071] Figure 3 It is a flow chart of a method for fusion tracking of cross-domain targets by multiple roadside units in a foggy scene described in the present invention. DETAILED DESCRIPTION

[0072] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0073] The technical solution of the present invention to solve the above technical problems is:

[0074] Figure 1 The figure shows the overall architecture of a multi-road side unit cross-domain target fusion tracking system in a foggy scene in a preferred embodiment of the present invention. The system includes three parts: a perception module, a fusion tracking module and a communication module:

[0075] The target detection module collects point cloud data of the surrounding environment through the roadside laser radar, obtains non-ground point cloud containing noise after background filtering and ground segmentation preprocessing of the point cloud data, and then uses the spatiotemporal sequence denoising method to screen out valid point cloud clusters, and finally obtains the target point cloud through clustering.

[0076] The fusion tracking module realizes cross-domain target perception fusion based on the sliding time window and the relative position of the sensor, and then updates the target trajectory through PHD (probability hypothesis density filtering).

[0077] The communication module is responsible for receiving and sending target measurement data between adjacent roadside units and sending road target information to vehicles.

[0078] Figure 2 The flowchart of a spatiotemporal sequence denoising method of the present invention is shown, and the process includes the following steps:

[0079] The roadside laser radar collects the surrounding point cloud data at the current k moment, and then preprocesses the collected raw point cloud data, including background filtering and ground segmentation to obtain non-ground point clouds. The spatiotemporal sequence denoising method based on point cloud integration and convolution is used to filter out the fog noise point cloud in the non-ground point cloud, and then the target point cloud cluster is extracted through clustering and added to the target measurement set Z. k The steps of this spatiotemporal series denoising method are:

[0080] 1 Point cloud integration: In the time series, statistics [t k-len ~t k-1 ] and compare it with the original point cloud at the current moment Superposition is performed to calculate the new point cloud after the point cloud accumulation

[0081]

[0082] 2 Voxelization: Taking the roadside LiDAR coordinate system as the center, divide the voxel into a voxel of size d x ×d y ×d z For each point cloud p, it is mapped to the voxel block according to its index coordinates. The index coordinates (i, j, k) are calculated as follows:

[0083]

[0084] 3. Point cloud intensity compensation: Compensate for point cloud intensity at different locations based on point cloud position and atmospheric attenuation coefficient:

[0085] I i =I·d 2 ·R·e -2αd #(3)

[0086] Among them, I is the original intensity of the point cloud, d is the distance from the point cloud to the lidar, R is the reflectivity of the target object, and α is the atmospheric attenuation coefficient. Under the condition that the atmospheric visibility is V meters, the commonly used α approximate formula is expressed as:

[0087]

[0088] 4 Point cloud filtering: For each voxel block, first calculate the distance d between each neighboring voxel and the central voxel i , and then according to the distance d i The weight W of each neighborhood voxel can be calculated using the Gaussian convolution kernel formula i :

[0089]

[0090] According to the weight, the intensity value I of the neighboring voxels i Perform weighted averaging to obtain the smooth intensity value I of the central voxel s :

[0091]

[0092] 5 Normalization: After smoothing, select the smallest point cloud intensity I in the voxel block min and the maximum point cloud intensity I max , and then the smoothing intensity value I of the central voxel s Normalize and get the normalized intensity value I n :

[0093]

[0094] Update the smooth intensity value of the central voxel to the voxel, and then normalize the smooth intensity value I n Compared with the given threshold T, the voxel blocks smaller than the given threshold T are filtered out, and the filtered point cloud set P′ is extracted N , and then for the point cloud set P′ N Perform clustering to obtain the target measurement set Z k And broadcast to other roadside units through the communication module.

[0095] Figure 3 The flowchart of the method for fusion tracking of multiple road side units across domain targets in a foggy scene of the present invention is shown, and the process includes the following steps:

[0096] 1 According to the translation and rotation matrix (R n ,T n ), the point cloud Z of the adjacent roadside unit n kTransform to the coordinate system of the roadside unit i to obtain the transformed target point cloud. In the time series, after the roadside unit receives the target point cloud data sent by the adjacent unit, it saves the point cloud data of a time window [t-Δt, t], and then dynamically adjusts the time window to obtain the time-space aligned target measurement set Z′ k .

[0097] 2. The roadside unit predicts the existing target strength based on the state transition model and the target survival probability, introduces the new target strength, and then updates the target strength through observation data. The predicted strength is combined with the measurement information to generate the posterior strength. Finally, the target state estimate is extracted through the updated strength function. The process includes the following steps:

[0098] Target state vector Where x is the target x position, y is the target y position, is the target position on the x-axis, The target position on the y axis.

[0099] Equation of state:

[0100] X k =F k X k-1 +Γ k v k #(8)

[0101] Observation equation:

[0102] Z k =H k X k +ω k #(9)

[0103] Among them, F k is the state transfer matrix, Γ k is the noise transfer matrix, v k Yes Obedience Gaussian white noise, H k is the observation matrix, ω k Yes Obedience Gaussian white noise, L k , Q k Both represent the variance of the noise,

[0104] 2.1 Intensity prediction: The intensity function of the target state at time k-1:

[0105]

[0106] Among them, J k-1 represents the number of Gaussian components in the Gaussian mixture model at time k-1, is the weight of the Gaussian density, It means the mean The covariance is Gaussian distribution.

[0107] Combined with the state transfer equation, the multi-target prediction strength at time k is:

[0108]

[0109] Among them, J k|k-1 Represents the number of Gaussian components predicted at time k;

[0110] Represents the prediction weight of the i-th target at time k:

[0111]

[0112] Represents the predicted mean of the i-th target at time k:

[0113]

[0114] Represents the predicted covariance of the i-th target at time k:

[0115]

[0116] The predicted mean is The covariance is Gaussian distribution.

[0117] 2.2 Calculation of new target strength: based on measurement set Z k-1 It can be obtained that the new target measurement set that may be included in the k-1 moment is:

[0118] Z B,k-1 =Z k-1 -Z S,k-1 #(15)

[0119] Among them, Z S,k-1 represents the surviving target measurement set at time k-1, then the number of new targets at time k-1 is:

[0120]

[0121] Based on the measurement set Z at k-1 time k-1 , the new target strength γ k|k-1 (x):

[0122]

[0123] Then the target strength v after the prediction step is k|k-1(x):

[0124] v k|k-1 (x) = v S,k|k-1 (x)+γ k|k-1 (x)#(18)

[0125] 2.3 Intensity update: based on the target predicted intensity v k|k-1 (x), survival target prediction strength v S,k|k-1 (x), prediction strength of new target γ k|k-1 (x), survival target measurement set Z S,k|k-1 , new target measurement set Z B,k|k-1 , then the posterior target strength v at time k k|k (x) is:

[0126]

[0127] 2.4 Gaussian component pruning and merging: Given a pruning threshold D th , merging threshold M th , for all Gaussian components if Then the component is deleted from the posterior distribution.

[0128] If the distance between the two Gaussian components is less than the given merging threshold D th , then the two components are considered to represent the same target and should be merged. The merger can be achieved by calculating the weighted average of the mean and covariance of the two components:

[0129]

[0130]

[0131] 2.5 Target state extraction: When the number of Gaussian components J is greater than J max , then retain the J with the largest weight max Gaussian components, the set of Gaussian components after pruning and merging can be expressed as Finally, take the sum of the weights of all Gaussian components as the target number estimate N at the current k moment k :

[0132]

[0133] The mean of the corresponding Gaussian component As the estimated position and velocity of the target, that is, the target state estimation

[0134]

[0135] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0136] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0137] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A multi-roadside unit cross-domain target fusion tracking system in foggy scenes, characterized in that: include: Target detection module, fusion tracking module and communication module. The target detection module is used to preprocess the original point cloud. The preprocessing includes background filtering and ground segmentation to obtain non-ground point cloud containing noise, and then the point cloud spatiotemporal sequence denoising method is used to filter out the fog noise point cloud, and then the target is extracted by clustering; The fusion tracking module is used to perform spatiotemporal fusion of the road target measurement and the measurement set released by the adjacent roadside unit received by the communication module to obtain an augmented target measurement set, and then update the target trajectory through probability hypothesis density filtering PHD. The communication module is used to receive and send target measurement sets between roadside units and send road target information to vehicles.

2. A tracking method based on the system of claim 1, characterized in that: The following steps are involved: 2.1 Point cloud collection and preprocessing: The roadside laser radar collects the surrounding environment point cloud data at the current moment, and then preprocesses the collected raw point cloud data, including background filtering and ground segmentation, to obtain the non-ground point cloud P containing noise. N ; 2.2 Point cloud spatiotemporal sequence denoising: Continuous multi-frame point cloud data are superimposed and assigned to different voxel blocks according to the point cloud index, and then the point cloud intensity of all point clouds is compensated. The Gaussian convolution kernel is used to traverse all voxel blocks to calculate the smooth intensity value of the central voxel and update it to the voxel block and normalize it. The normalized smooth intensity value of the central voxel I n Compared with the given threshold T, the denoised point cloud set P is obtained ′ N ; 2.3 Target Extraction: De-noised Point Cloud P ′ N Perform clustering and extract the target to the observation set Z k And broadcast to adjacent roadside units; 2.4 Measurement set transmission: The roadside unit transmits the target measurement set to the adjacent roadside unit; 2.5 Measurement set reception: The roadside unit receives the target measurement set of the adjacent node; 2.6 Spatiotemporal fusion: First, create a suitable sliding time window according to the time series, put the target measurements of other nodes that meet the time series into the window, and convert the target measurements in the window to the current roadside lidar coordinate system according to the relative position of the roadside equipment, and fuse them with the own observation data to obtain the augmented target measurement set Z. ′ k ; 2.7 Target trajectory extraction: The roadside unit predicts the existing target strength based on the state transition model and the target survival probability, introduces the new target strength, and then updates the target strength through observation data. The predicted strength is combined with the measurement information to generate the posterior strength. Finally, the target state estimate is extracted through the updated strength function. 2.8 Target status transmission: The roadside unit sends road target information to the vehicle; 2.9 Repeat steps 2.1 to 2.

8.

3. The tracking method according to claim 2, characterized in that the step 2.2 of the spatiotemporal sequence point cloud denoising method specifically comprises: 3.1 Point cloud integration: In the time series, statistics [t k-len ~t k-1 ] and compare it with the original point cloud at the current moment Superposition is performed to calculate the new point cloud after the point cloud accumulation 3.2 Voxelization: Taking the roadside LiDAR coordinate system as the center, divide the voxel into a voxel of size d x ×d y ×d z For each point cloud p, it is mapped to the voxel block according to its index coordinates. The index coordinates (i, j, k) are calculated as follows: 3.3 Point cloud intensity compensation: According to the point cloud position and atmospheric attenuation coefficient, compensation is made for the point cloud intensity at different positions: I i =I·d 2 ·R·e -2αd #(3) Among them, I is the original intensity of the point cloud, d is the distance from the point cloud to the lidar, R is the reflectivity of the target object, and α is the atmospheric attenuation coefficient. Under the condition that the atmospheric visibility is V meters, the commonly used α approximate formula is expressed as: 3.4 Point cloud filtering: For each voxel block, first calculate the distance d between each neighborhood voxel and the central voxel i , and then according to the distance d i The weight W of each neighborhood voxel is calculated using the Gaussian convolution kernel formula i : According to the weight, the intensity value I of the neighboring voxels i Perform weighted averaging to obtain the smooth intensity value I of the central voxel s : 3.5 Normalization: After smoothing, select the smallest point cloud intensity I in the voxel block min and the maximum point cloud intensity I max , and then the smoothing intensity value I of the central voxel s Normalize and get the normalized intensity value I n : Update the smooth intensity value of the central voxel to the voxel block, and then normalize the smooth intensity value I n Compared with the given threshold T, the voxels smaller than the given threshold T are removed to obtain the filtered point cloud set P ′ N It can be used for target detection.

4. The tracking method according to claim 2, characterized in that: The target trajectory extraction method in step 2.7 is specifically as follows: 4.1 Establishing motion model: target motion state Where x is the target x position, y is the target y position, is the target x speed, is the target y velocity, and the state transfer equation is: X k =F k X k-1 +Γ k v k #(8) Among them, F k is the state transfer matrix, Γ k is the noise transfer matrix, v k Yes Obedience Gaussian white noise, L k represents the variance of the noise; Its measurement equation: Z k =H k X k +ω k #(9) Among them, H k is the observation matrix, ω k Yes Obedience Gaussian white noise; 4.2 Intensity prediction: Intensity function of multi-objective state at time k-1: Among them, J k-1 represents the number of Gaussian components in the Gaussian mixture model at time k-1, is the weight of the Gaussian density, It means the mean The covariance is Gaussian distribution of Combined with the state transfer equation, the multi-target prediction strength at time k is: in, represents the prediction weight of the i-th target at time k, represents the predicted mean value of the i-th target at time k, represents the predicted covariance of the i-th target at time k, The predicted mean is The covariance is Gaussian distribution of 4.3 Calculation of new target strength: based on measurement set Z k-1 The new target measurement set that may be included in the k-1 moment is obtained as follows: WITH B,k-1 =Z k-1 -WITH S,k-1 #(15) Among them, Z S,k-1 represents the surviving target measurement set at time k-1, then the number of new targets at time k-1 is: Based on the measurement set Z at k-1 time k-1 , the new target strength γ k|k-1 (x): Then the target strength v after the prediction step is k|k-1 (x): v k|k-1 (x)=v S,k|k-1 (x)+γ k|k-1 (x)#(18) 4.3 Intensity update: based on the target predicted intensity v k|k-1 (x), survival target prediction strength v S,k|k-1 (x), prediction strength of new target γ k|k-1 (x), survival target measurement set Z S,k|k-1 , new target measurement set Z B,k|k-1 , then the posterior target strength v at time k k|k (x) is: 4.4 Gaussian component pruning and merging: Given a pruning threshold D th , merging threshold M th , for all Gaussian components if Then delete the component from the posterior distribution; If the distance between the two Gaussian components is less than the given merging threshold D th , then the two components are considered to represent the same target and should be merged; the merger can be achieved by calculating the weighted average of the mean and covariance of the two components: 4.5 Target state extraction: When the number of Gaussian components J is greater than J max , then retain the J with the largest weight max Gaussian components, the set of Gaussian components after pruning and merging can be expressed as Finally, take the sum of the weights of all Gaussian components as the target number estimate N at the current k moment k : The mean of the corresponding Gaussian component As the estimated position and velocity of the target, that is, the target state estimation

Images