A multi-roadside unit cross-domain target fusion tracking system and method under a foggy scene

Abstract

The application claims a multi-roadside unit cross-domain target fusion tracking system and method in a foggy scene, which is composed of a target detection module, a fusion tracking module and a communication module. The target detection module pre-processes the original point cloud, including background filtering, ground segmentation to obtain non-ground point cloud containing noise, then uses a point cloud space-time sequence denoising method to filter out fog noise point cloud, and then extracts the target through clustering; the fusion tracking module performs space-time fusion on the road target measurement and the measurement set published by the adjacent roadside unit received by the communication module to obtain an augmented target measurement set, and then updates the target trajectory through a PHD filter; the communication module is used for receiving and sending target measurement sets between roadside units and sending road traffic target information to vehicles. The application realizes continuous and stable tracking of foggy road targets through point cloud space-time sequence denoising and multi-roadside unit cross-domain target fusion tracking, and enhances the safety of vehicle driving in fog.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of automation and computer technology, and particularly relates to a multi-roadside unit cross-domain target fusion tracking system and method in a foggy scene. BACKGROUND

[0002] Roadside perception is an important part of the vehicle-road cooperative system, and target tracking technology is the key to ensuring 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 the performance of roadside environment perception.

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

[0004] Chinese patent application: Target trajectory fusion method based on multi-point roadside perception device (application number: CN202410009145.1) discloses a target trajectory fusion system based on a multi-point roadside perception device, which stores point perception information in a central server buffer. After time alignment, the system performs trajectory matching and data fusion. This system uses a centralized server to process fusion data, and only uses point cloud at a certain moment, without considering the context features of the point cloud.

[0005] Chinese patent application: Roadside unit cooperative target tracking system, method and storage medium (application number: CN202210171778.3) discloses a multi-roadside unit cooperative target tracking system, which uses adjacent roadside units to share target trajectory information to reduce new target tracking initiation judgment. However, it does not use target detection data from adjacent roadside units. SUMMARY

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

[0007] A multi-roadside unit cross-domain target fusion tracking system in a foggy scene comprises:

[0008] The target detection module is used for preprocessing the original point cloud, the preprocessing including background filtering, ground segmentation to obtain a non-ground point cloud containing noise, and then using a point cloud spatiotemporal sequence denoising method to filter out the fog noise point cloud, and then extracting the target through clustering;

[0009] The fusion tracking module is used for spatiotemporal fusion of the road target measurement and the measurement set published by the adjacent road side unit to obtain an augmented target measurement set, and then updating the target trajectory through a probability hypothesis density filter PHD.

[0010] The communication module is used for receiving and sending the target measurement set between the road side units and sending the road target information to the vehicle.

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

[0012] 2.1 Point cloud collection and preprocessing: the road side laser radar collects the surrounding environment point cloud data at the current time, and then the collected original point cloud data is preprocessed, including background filtering and ground segmentation to obtain a non-ground point cloud P containing noise N ;

[0013] 2.2 Point cloud spatiotemporal sequence denoising: the continuous multiple frames of point cloud data are superimposed and distributed to different voxel blocks according to the point cloud index, 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 center voxel and update it to the voxel block and perform normalization, and the normalized smooth intensity value I of the center voxel n is compared with a given threshold T to obtain a denoised point cloud set P' N ;

[0014] 2.3 Target extraction: clustering the denoised point cloud P' N , and extracting the target to the observation set Z k and broadcasting to the adjacent road side unit;

[0015] 2.4 Measurement set sending: the road side unit sends the target measurement set to the adjacent road side unit;

[0016] 2.5 Measurement set receiving: the road side 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 sequence, put the target measurement of other nodes meeting the time sequence into the window, and convert the target measurement in the window to the current road side laser radar coordinate system according to the relative position of the road side device, and fuse with the observation data to obtain an augmented target measurement set Z' k ;

[0018] 2.7 Target trajectory extraction: the roadside unit predicts the existing target intensity according to the state transition model and the target survival probability, introduces the new target intensity, updates the target intensity through the observation data, combines the predicted intensity with the measurement information to generate the posterior intensity, and finally extracts the target state estimation through the updated intensity function;

[0019] 2.8 Target state sending: the roadside unit sends the road target information to the vehicle;

[0020] 2.9 Repeat steps 2.1-2.8;

[0021] Further, the space-time sequence point cloud denoising method of step 2.2 specifically comprises:

[0022] 3.1 Point cloud integration: in the time sequence, the point clouds at times [t k-len -t k-1 ] are counted and superimposed with the original point cloud at the current time t to calculate the new point cloud after the point cloud accumulation

[0023]

[0024] 3.2 Voxelization: taking the roadside laser radar coordinate system as the center, a cubic voxel block with a size of d x ×d y ×d z is segmented, and for each point cloud p, the index coordinates are mapped to the voxel block according to the index coordinates, and the index coordinates (i, j, k) are calculated according to the following formula:

[0025]

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

[0027] I i = I·d 2 ·R·e -2αd #(3)

[0028] Where I is the original intensity of the point cloud, d is the distance from the point cloud to the laser radar, R is the reflectivity of the target object, and a is the atmospheric attenuation coefficient. Under the condition that the atmospheric visibility is V meters, the commonly used a approximate formula is:

[0029]

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

[0031]

[0032] The intensity value I of the center voxel is obtained by weighted average of the intensity values of the neighborhood voxels according to the weight i : s :

[0033]

[0034] 3.5 Normalization: After smoothing, the minimum point cloud intensity I min and the maximum point cloud intensity I max in the voxel block are selected, and the normalized intensity value I s is obtained by normalizing the smoothed intensity value I n :

[0035]

[0036] The smoothed intensity value of the center voxel is updated in the voxel block, and the normalized smoothed intensity value I n is compared with the given threshold T, and the voxels smaller than the given threshold T are removed to obtain the filtered point cloud set P' N which can be used for target detection.

[0037] Further, the step 2.7 target trajectory extraction method specifically includes the following steps:

[0038] 4.1 Establish a motion model: the motion state of the target where x is the x position of the target, y is the y position of the target, is the x velocity of the target, is the y velocity of the target, and the state transition equation is:

[0039] X k =F k X k-1 +Γ k v k #(8)

[0040] where F k is the state transition matrix, Γ k is the noise transition matrix, v k is a Gaussian white noise subject to , and L k represents the variance of the noise.

[0041] The measurement equation is:

[0042] Z k =Hk X k +ω k #(9)

[0043] where H k is the observation matrix, ω k is the Gaussian white noise with

[0044] 4.2 Intensity Prediction: Intensity Function of Multi-target State at Time k-1

[0045]

[0046] where J k-1 denotes the number of Gaussian components in the Gaussian Mixture Model at time k-1, is the weight of the Gaussian density, denotes the Gaussian distribution with mean and covariance

[0047] Combining the state transition equation, the multi-target predicted intensity at time k is:

[0048]

[0049] where denotes the predicted weight of the i-th target at time k, denotes the predicted mean of the i-th target at time k, denotes the predicted covariance of the i-th target at time k, denotes the Gaussian distribution with predicted mean and covariance

[0050] 4.3 New-born Target Intensity Calculation: Based on Measurement Set Z k-1 The measurement set of possible new-born targets at time k-1 is obtained as:

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

[0052] where Z S,k-1 denotes the measurement set of surviving targets at time k-1, then the number of new-born targets at time k-1 is:

[0053]

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

[0055] ​​​

[0056] the predicted target intensity v k|k-1 (x) after the prediction step

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

[0058] 4.3 Intensity update: given the predicted target intensity v k|k-1 (x), the surviving target predicted intensity v S,k|k-1 (x), the newborn target predicted intensity γ k|k-1 (x), the surviving target measurement set Z S,k|k-1 , the newborn target measurement set Z B,k|k-1 , the posterior target intensity v k|k (x) at time k is:

[0059]

[0060] 4.4 Gaussian component pruning and merging: given a pruning threshold D th , a merging threshold M th , for all Gaussian components If , the component is removed from the posterior distribution.

[0061] If the distance between two Gaussian components is less than a given merging threshold M th , the two components are considered to represent the same target and should be merged. The merging can be achieved by computing the weighted average of the means and covariances of the two components:

[0062]

[0063] 4.5 Target state extraction: when the number of Gaussian components J is greater than J max , the J max components with the largest weights are kept. The set of pruned and merged Gaussian components can be denoted as The sum of the weights of all Gaussian components is taken as the estimate of the number of targets N k at time k:

[0064]

[0065] The mean of the corresponding Gaussian component is taken as the estimate of the position and velocity of the target, i.e., the target state estimate

[0066]

[0067] The advantages and beneficial effects of the present application are as follows:

[0068] The present application provides a multi-roadside unit cross-domain target fusion tracking system and method in a foggy scene, superimposes multiple frames of target point clouds, and uses a space-time sequence denoising method to filter out fog noise point clouds around the effective target point clouds to extract effective point cloud clusters; the roadside unit predicts the existing target intensity according to the state transition model and the target survival probability, and updates the target intensity according to the observation data, combines the predicted intensity with the measurement information, generates the posterior intensity to extract the target state estimation. The present application innovatively proposes a space-time sequence denoising method, which superimposes multiple frames of point clouds, and then divides them into different voxel blocks to perform Gaussian convolution in turn to filter out fog noise point clouds; the present application innovatively considers the cooperation between multiple roadside devices, superimposes the space-time fusion of multiple adjacent roadside laser radar point clouds, calculates the state transition model and the target survival probability to predict the existing target intensity, introduces the new target intensity, and then updates the target intensity through the observation data, combines the predicted intensity with the measurement information, generates the posterior intensity, and finally extracts the target state estimation through the updated intensity function;

[0069] The present application uses the filtering and multi-roadside unit cross-domain fusion method to improve the target tracking accuracy and stability of the roadside laser radar in the foggy weather, realizes the continuous and stable tracking of the road target in the foggy scene, and improves the safety of the vehicle automatic driving or assisted driving. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 The present application provides a preferred embodiment of a multi-roadside unit cross-domain target fusion tracking system in a foggy scene.

[0071] Figure 2 The present application provides a preferred embodiment of a multi-roadside unit cross-domain target fusion tracking system in a foggy scene.

[0072] Figure 3 The present application provides a preferred embodiment of a multi-roadside unit cross-domain target fusion tracking system in a foggy scene. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only a part of the embodiments of the present application.

[0074] The technical solutions of the present application to solve the above technical problems are:

[0075] Figure 1 The present application provides a preferred embodiment of a multi-roadside unit cross-domain target fusion tracking system in a foggy scene, which includes a perception module, a fusion tracking module and a communication module three parts:

[0076] The target detection module collects point cloud data of the surrounding environment through a roadside laser radar, obtains non-ground point cloud containing noise after preprocessing of background filtering and ground segmentation on the point cloud data, then screens out effective point cloud clusters using a spatio-temporal sequence denoising method, and finally obtains target point cloud through clustering.

[0077] The fusion tracking module realizes cross-domain target perception fusion based on a sliding time window and relative position of sensors, and then updates the target trajectory through PHD (probability hypothesis density filter) filtering.

[0078] The communication module is responsible for receiving and sending target measurement data between adjacent roadside units and sending road target information to vehicles.

[0079] Figure 2 A flowchart of the spatio-temporal sequence denoising method of the present application is shown, which comprises the following steps:

[0080] The roadside laser radar collects point cloud data of the surrounding environment at the current k time, and then pre-processes the collected original point cloud data, including background filtering and ground segmentation to obtain non-ground point cloud. The spatio-temporal sequence denoising method based on point cloud integration and convolution is used to filter out 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 the spatio-temporal sequence denoising method are as follows:

[0081] 1. Point cloud integration: in the time sequence, count the point cloud at time [t k-len ~t k-1 ] and superimpose it with the original point cloud at the current time to calculate the new point cloud after point cloud accumulation

[0082]

[0083] 2. Voxelization: taking the roadside laser radar coordinate system as the center, a cubic voxel block with a size of d x ×d y ×d z is segmented, and for each point cloud p, it is mapped to the voxel block according to its index coordinates, and the index coordinates (i, j, k) are calculated as follows:

[0084]

[0085] 3. Point cloud intensity compensation: according to the point cloud position and the atmospheric attenuation coefficient, the point cloud intensity at different positions is compensated:

[0086] I i = I·d 2 ·R·e -2αd #(3) ​

[0087] Where 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 of atmospheric visibility of V meters, the commonly used approximate formula for α is:

[0088]

[0089] 4. Point cloud filtering: For each voxel block, first calculate the distance d between each neighboring voxel and the central voxel. i Then based on the distance d i The Gaussian convolution kernel formula can be used to calculate the weight W of each neighborhood voxel. i :

[0090]

[0091] Based on the weighted values ​​of the neighboring voxel strength I i A weighted average is performed to obtain the smoothing intensity value I of the central voxel. s :

[0092]

[0093] 5. Normalization: After smoothing, select the point cloud intensity I with the smallest value in the voxel block. min And the maximum point cloud intensity I max Then, regarding the smoothing intensity value I of the central voxel... s Normalization is performed to obtain the normalized intensity value I. n :

[0094]

[0095] Update the smoothing intensity value of the central voxel to the voxel, and then normalize the smoothing intensity value I. n Compared with a given threshold T, voxel blocks smaller than the given threshold T are filtered out, and the filtered point cloud set P′ is extracted. N Then, for the point cloud set P′ N Clustering is performed to obtain the target measurement set Z. k It is then broadcast to other roadside units via the communication module.

[0096] Figure 3 The flowchart shown is a cross-domain target fusion tracking method for multi-path side-cells in foggy weather according to the present invention. The method includes the following steps:

[0097] 1. Based on the translation and rotation matrices (Ri) between roadside unit n, n∈[1,2,…,i,i+1,…,N] and the current roadside unit i. n ,T n ), and the point cloud Z of the adjacent roadside unit n kTransform to the local roadside unit i coordinate system, get the transformed target point cloud. In the time sequence, the roadside unit receives the target point cloud data sent by the adjacent unit, saves the point cloud data in a time window [t-Δt, t], and then dynamically adjusts the time window, so as to obtain the spatio-temporal alignment target measurement set Z' k .

[0098] 2. The roadside unit predicts the existing target intensity according to the state transition model and the target survival probability, introduces the new target intensity, updates the target intensity through the observation data, combines the predicted intensity with the measurement information to generate the posterior intensity, and finally extracts the target state estimation through the updated intensity function. The process includes the following steps:

[0099] Target state vector Where x is the target x position, y is the target y position, is the target position on the x axis, is the target position on the y axis.

[0100] State equation:

[0101] X k =F k X k-1 +Γ k v k #(8)

[0102] Observation equation:

[0103] Z k =H k X k +ω k #(9)

[0104] Where, F k is the state transition matrix, Γ k is the noise transition matrix, v k is the Gaussian white noise , H k is the observation matrix, ω k is the Gaussian white noise , L k , Q k all represent the variance of noise,

[0105] 2.1 Intensity prediction: intensity function of target state at time k-1:

[0106]

[0107] Where, 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, represents the mean value covariance of a Gaussian distribution.

[0108] Combining the state transition equation, the multi-target prediction intensity at time k is:

[0109]

[0110] where J k|k-1 represents the number of Gaussian components predicting at time k;

[0111] represents the prediction weight of the i-th target at time k:

[0112]

[0113] represents the prediction mean of the i-th target at time k:

[0114]

[0115] represents the prediction covariance of the i-th target at time k:

[0116]

[0117] represents the predicted mean value covariance of a Gaussian distribution.

[0118] 2.2 New target intensity calculation: based on the measurement set Z k-1 The measurement set of new targets possibly contained at time k-1 is:

[0119] Z B,k-1 = Z k-1 - Z S,k-1 #(15)

[0120] where Z S,k-1 represents the measurement set of surviving targets at time k-1, and the number of new targets at time k-1 is:

[0121]

[0122] Based on the measurement set Z k-1 at time k-1, the new target intensity γ k|k-1 (x):

[0123]

[0124] After the prediction step, the target intensity v k|k-1(x) = v

[0125] v k|k-1 (x) = v S,k|k-1 (x) + γ k|k-1 (x) # (18)

[0126] 2.3 Strength update: based on target predicted strength v k|k-1 (x), surviving target predicted strength v S,k|k-1 (x), newborn target predicted strength γ k|k-1 (x), surviving target measurement set Z S,k|k-1 , newborn target measurement set Z B,k|k-1 , then the posterior target strength v k|k (x) is:

[0127]

[0128] 2.4 Gaussian component pruning and merging: given pruning threshold D th , merging threshold M th , for all Gaussian components If , then the component is removed from the posterior distribution.

[0129] If the distance between two Gaussian components is less than a given merging threshold M th , then the two components are considered to represent the same target and should be merged. Merging can be achieved by computing the weighted average of the means and covariances of the two components:

[0130]

[0131]

[0132] 2.5 Target state extraction: when the number of Gaussian components J is greater than J max , then the J max Gaussian components with the largest weights are kept, and the set of pruned and merged Gaussian components can be represented as Finally, the sum of the weights of all Gaussian components is taken as the target number estimate N k :

[0133]

[0134] The mean of the corresponding Gaussian component is taken as the estimate of the target's position and velocity, i.e., the target state estimate

[0135]

[0136] The systems, apparatuses, modules or units disclosed in the above examples can be implemented by computer chips or entities, or by products with certain functions.

[0137] It should also be noted that the terms "comprising", "including", or any other variant are intended to cover a non-exclusive inclusion, such that processes, methods, articles, or apparatuses that comprise a list of elements are not required to include only those elements in the list, but can include other elements not expressly listed or inherent to such processes, methods, articles, or apparatuses.

[0138] The above examples are to be understood only as tools for illustrating the present application and not as limiting the scope of protection of the present application. After reading the specification, the skilled person can make various modifications or alterations to the present application, and these equivalent changes and modifications also fall within the scope of the present application defined by the claims.

Claims

1. A multi-path side-unit cross-domain target fusion tracking system for foggy scenes, characterized in that, include: The system includes a target detection module, a fusion tracking module, and a communication module. The target detection module is used to preprocess the original point cloud. The preprocessing includes background filtering, ground segmentation to obtain noisy non-ground point clouds, then using a point cloud spatiotemporal sequence denoising method to filter out fog noise point clouds, and finally extracting the target through clustering. The fusion tracking module is used to perform spatiotemporal fusion of road target measurements and measurement sets received by the communication module from adjacent roadside units to obtain an augmented target measurement set, and then use probability hypothesis density filtering (PHD) to update the target trajectory. The communication module is used for receiving and sending target measurement sets between roadside units and sending road target information to vehicles; The fusion tracking module extracts the target trajectory by including the following steps: 4.1 Establishing a motion model: Target motion state Where x is the target x position, y is the target y position, For the target speed x, For the target velocity y, the state transition equation is: X k =F k X k-1 +Γ k v k (8) Among them, F k It is the state transition matrix, Γ k It is the noise transition matrix, v k It is 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 It is the observation matrix, ω k It is obedience Gaussian white noise; 4.2 Intensity Prediction: Intensity Function of Multi-Target State at Time k-1: Among them, J k-1 This represents the number of Gaussian components in the Gaussian mixture model at time k-1. The weights are Gaussian density. The mean is covariance is Gaussian distribution; Combining the state transition equation, the multi-target prediction intensity at time k is: in, This represents the prediction weight of the i-th target at time k. This represents the predicted mean of the i-th target at time k. This represents the prediction covariance of the i-th target at time k. The mean of the prediction is covariance is Gaussian distribution; 4.3 Calculation of New Target Intensity: Based on Measurement Set Z k-1 The possible set of newly generated target measurements at time k-1 is as follows: WITH B,k-1 =Z k-1 -WITH S,k-1 (15) Among them, Z S,k-1 Let represent the surviving target measurement set at time k-1. Then, the number of newly generated targets at time k-1 is: Based on the measurement set Z at time k-1 k-1 New target intensity γ k|k-1 (x): The target intensity v after the prediction step 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 Target Predicted Intensity v k|k-1 (x), Survival target prediction intensity v S,k|k-1 (x), Predictive intensity of new target γ k|k-1 (x), Survival target measurement set Z S,k|k-1 Newborn target measurement set Z B,k|k-1 Then the posterior target intensity v at time k k|k (x) is: 4.4 Gaussian Component Trimming and Merging: Given a trimming threshold D th Merging threshold M th For all Gaussian components if Then remove that component from the posterior distribution; If the distance between two Gaussian components is less than a given merging threshold M th If the two components represent the same objective, they should be merged. Merging is achieved by calculating a 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 The set of Gaussian components after trimming and merging is represented as follows: Finally, the sum of the weights of all Gaussian components is taken as the target number estimate N at time k. k : The mean of the corresponding Gaussian components The estimated position and velocity of the target, i.e., the target state estimation.

2. A tracking method based on the system of claim 1, characterized in that, Includes the following steps: 2.1 Point Cloud Acquisition and Preprocessing: The roadside lidar acquires point cloud data of the surrounding environment at the current moment. Then, the acquired raw point cloud data is preprocessed, including background filtering and ground segmentation, to obtain a noisy non-ground point cloud P. N ; 2.2 Point Cloud Spatiotemporal Sequence Denoising: Multiple consecutive frames of point cloud data are superimposed and assigned to different voxel blocks according to the point cloud index. Point cloud intensity compensation is then applied to all point clouds. A Gaussian convolution kernel is used to traverse all voxel blocks to calculate the smoothing intensity value of the central voxel, which is then updated and normalized. The normalized smoothing intensity value I of the central voxel is then calculated. n By comparing with a given threshold T, the denoised point cloud set P′ is obtained. N ; 2.3 Target Extraction: Extracting the denoised point cloud P′ N Perform clustering to extract the target into 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 from adjacent nodes; 2.6 Spatiotemporal Fusion: First, a suitable sliding time window is created based on the time series. Target measurements from other nodes that match the time series are placed into the window. Then, based on the relative position of the roadside equipment, the target measurements within the window are transformed to the current roadside lidar coordinate system and fused with the observed data to obtain the augmented target measurement set Z′. k ; 2.7 Target Trajectory Extraction: The roadside unit predicts the intensity of existing targets based on the state transition model and the survival probability of the targets, while introducing the intensity of newly emerging targets. The target intensity is then updated through observation data. The predicted intensity is combined with the measurement information to generate the posterior intensity. Finally, the target state estimate is extracted through the updated intensity function. 2.8 Target Status Transmission: The roadside unit transmits 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 spatiotemporal sequence point cloud denoising method in step 2.2 specifically includes: 3.1 Point Cloud Integral: In time series, statistical t k-len ~t k-1 Point clouds at each time point, and compared with the original point cloud at the current time point. The points are superimposed to obtain a new point cloud after accumulation. 3.2 Voxelization: Centered on the roadside lidar coordinate system, divide the area into segments of size d. x ×d y ×d z For a cube voxel block, for each point cloud p, its index coordinates are mapped onto the voxel block. The formula for calculating the index coordinates (i,j,k) is as follows: 3.3 Point Cloud Intensity Compensation: Based on the point cloud location and atmospheric attenuation coefficient, compensation is made for the point cloud intensity at different locations. I i =I·d 2 ·R·e -2αd (3) in, I represents the original intensity of the point cloud, d represents the distance from the point cloud to the lidar, R represents the reflectivity of the target object, and α is the atmospheric attenuation coefficient. Under the condition of atmospheric visibility of V meters, the commonly used approximate formula for α is as follows: 3.4 Point Cloud Filtering: For each voxel block, first calculate the distance d between each neighboring voxel and the central voxel. i Then based on the distance d i The weight W of each neighborhood voxel is calculated using the Gaussian convolution kernel formula. i : Based on the weighted values ​​of the neighboring voxel strength I i A weighted average is performed to obtain the smoothing intensity value I of the central voxel. s : 3.5 Normalization: After smoothing, select the point cloud intensity I with the smallest value in the voxel block. min And the maximum point cloud intensity I max Then, regarding the smoothing intensity value I of the central voxel... s Normalization is performed to obtain the normalized intensity value I. n : Update the smoothing intensity value of the central voxel to the voxel block, and then normalize the smoothing intensity value I. n By comparing with a given threshold T, voxels smaller than the given threshold T are removed, resulting in a filtered point cloud set P′. N It can be used for target detection.

Images