Tracking target dynamic and static state judgment method and system based on neural network

Abstract

The present application relates to a kind of neural network-based tracking target dynamic state determination method and system thereof.The method includes: training step, for each tracking target in tracking target set, feature point is extracted, feature point is converted into feature vector and the true value of the dynamic state of tracking target is labeled, feature vector and corresponding dynamic state true value are used as training sample parameters to train neural network, and the trained neural network classification model is obtained;And test step, the feature vector of tracking target is extracted, input into the neural network classification model, and the dynamic state classification result of tracking target is obtained.According to the present application, the accuracy of determination result can be improved by using neural network method to determine the dynamic state of laser radar tracking target.

Description

Technical Field

[0001] This invention relates to autonomous driving technology, and more specifically to a method and system for judging the dynamic and static states of a tracking target based on a neural network. Background Technology

[0002] LiDAR, as an important sensor for autonomous vehicles, is typically used to detect and track targets such as other vehicles on the road. In the road target tracking problem, classifying the target's dynamic and static states is an important sub-problem.

[0003] Existing technologies for determining the motion or static state of a tracked target typically involve directly tracking the target's speed and judging its motion or stillness based on the speed magnitude. However, this method is susceptible to changes in point cloud morphology and tracking position due to occlusion or other factors, leading to the misclassification of static targets as dynamic ones. Furthermore, directly classifying a target's motion or static state based on its speed can easily result in misclassifying slow-moving pedestrians as stationary targets. Misjudging the motion or static state of a target impacts vehicle decision-making, planning, and control; therefore, sophisticated algorithms are needed to improve the accuracy of target motion or static state determination. Summary of the Invention

[0004] In view of the above problems, the present invention aims to propose a neural network-based method and system for judging the dynamic and static states of a tracking target, which can accurately determine the dynamic and static states of the target.

[0005] The present invention provides a method for determining the dynamic and static states of a tracking target based on a neural network, characterized in that it includes:

[0006] The training steps are as follows: For each tracked target in the target set from the vehicle's LiDAR, feature points are extracted, converted into feature vectors, and the ground truth values ​​of the target's dynamic and static states are labeled. The feature vectors and the corresponding ground truth values ​​of the dynamic and static states are used as training sample parameters to train the neural network, resulting in a trained neural network classification model.

[0007] The testing steps involve extracting the feature vector of the tracked target, inputting it into the neural network classification model, and obtaining the dynamic and static state classification results of the tracked target.

[0008] Optionally, the training steps include:

[0009] The first feature extraction sub-step extracts feature points for each tracking target in the tracking target set;

[0010] The first feature transformation sub-step converts the feature points into feature vectors; and

[0011] The model training sub-step involves labeling the transformed feature vectors with the true values ​​of the target's dynamic and static states, using these as training sample parameters to train the neural network, thus obtaining the neural network classification model.

[0012] Optionally, in the first feature extraction sub-step, the point closest to the vehicle, the center point of the long side, and the centroid of the tracked target are extracted as the feature points.

[0013] Optionally, in the first feature extraction sub-step, the method for extracting the point closest to the vehicle is to traverse all points in the point cloud data of the tracking target to obtain the point closest to the origin of the lidar; the method for extracting the center point of the long side is to use a straight line fitting method on the point cloud data to extract the longer edge of the point cloud and find the center point of the edge; the method for extracting the centroid is to directly sum and average the three-dimensional coordinates of each point in the point cloud data, wherein the point cloud data is composed of points collected by the lidar.

[0014] Optionally, the first feature transformation sub-step includes:

[0015] Let the current time be t, and let Pn,t(X) be the nearest point extracted at time t. n,t , Y n,t The center point of the longer side is Pe,t(X). e,t , Y e,t ), with centroid Pc,t(X) c,t , Y c,t The eigenvector is a one-dimensional expansion of the time series of three feature points, represented as Ft=[X]. n,t , Y n,t , X e,t , Y e,t, X c,t , Y c,t, X n,t-1 , Y n,t-1 , X e,t-1 Ye t-1 , X c,t , Y c,t-1 ,..., X n,t-N , Y n,t-N , X e,t-N , Y e,t-N , X c,t-N , Y c,t-N ], where N takes the value of 5, and the label of the target's dynamic and static state true value is denoted as Mt, where Mt takes the value of 0 or 1. Then a training sample data is represented as Si=[Ft, Mt].

[0016] Optionally, the test steps include:

[0017] The second feature extraction sub-step involves extracting feature points from the tracking results for each tracking target.

[0018] The second feature transformation sub-step converts feature points into feature vectors; and

[0019] The testing sub-step involves inputting the feature vector into the trained neural network classification model to determine the dynamic or static state of the currently tracked target and obtain the dynamic or static state classification result.

[0020] Optionally, in the test sub-step, the obtained classification results of the dynamic and static states of the tracked target are further filtered using Bayesian filtering.

[0021] The present invention provides a neural network-based target motion / static state determination system, characterized in that it comprises:

[0022] The training module is used to extract feature points for each tracked target in the target set from the vehicle's LiDAR, convert the feature points into feature vectors, and label the ground truth values ​​of the target's dynamic and static states. The feature vectors and the corresponding ground truth values ​​of the dynamic and static states are used as training sample parameters to train the neural network, resulting in a trained neural network classification model.

[0023] The testing module is used to extract the feature vector of the tracked target, input it into the neural network classification model, and obtain the dynamic and static state classification result of the tracked target.

[0024] Optionally, the training module includes:

[0025] The first feature extraction submodule is used to extract feature points for each tracking target in the tracking target set;

[0026] A first feature transformation submodule is used to convert the feature points into feature vectors; and

[0027] The model training submodule is used to label the transformed feature vectors with the true values ​​of the dynamic and static states of the tracking target, and use them as training sample parameters to train the neural network to obtain the neural network classification model.

[0028] Optionally, in the first feature extraction submodule, the point closest to the vehicle, the center point of the long side, and the centroid of the tracked target are extracted as the feature points.

[0029] Optionally, in the first feature extraction submodule, in the first feature extraction substep, the method for extracting the point closest to the vehicle is to traverse all points in the point cloud data of the tracking target to obtain the point closest to the origin of the lidar; the method for extracting the center point of the long side is to use a straight line fitting method on the point cloud data to extract the longer edge of the point cloud and find the center point of the edge; the method for extracting the centroid is to directly sum and average the three-dimensional coordinates of each point in the point cloud data, wherein the point cloud data is composed of points collected by the lidar.

[0030] Optionally, the test module includes:

[0031] The second feature extraction submodule is used to extract feature points for each tracking target in the tracking target set;

[0032] The second feature transformation submodule is used to convert feature points into feature vectors; and

[0033] The testing submodule is used to input the feature vector into the trained neural network classification model to determine the dynamic and static state of the currently tracked target and obtain the dynamic and static state classification result.

[0034] Optionally, in the test submodule, the obtained classification results of the dynamic and static states of the tracked target are further filtered using Bayesian filtering.

[0035] The present invention provides a computer-readable medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the aforementioned method for determining the dynamic and static states of a tracking target based on a neural network.

[0036] A computer device according to one aspect of the present invention includes a storage module, a processor, and a computer program stored on the storage module and executable on the processor, characterized in that the processor executes the computer program to implement the neural network-based method for determining the dynamic and static states of a tracking target.

[0037] As described above, the target motion and static state judgment method and system based on neural networks of the present invention can solve the misjudgment problem caused by simply using speed judgment (e.g., identifying a static target that is blocked as a moving target, or judging a slow pedestrian as a stationary target) by using a neural network method to judge the motion and static state of the target tracked by the lidar. This can improve the accuracy of the judgment results. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the method for determining the dynamic and static states of a tracking target based on a neural network according to the present invention.

[0039] Figure 2 This is a flowchart illustrating the training phase of a neural network-based target motion / static state judgment method according to an embodiment of the present invention.

[0040] Figure 3 An example of a tracking time series formed by tracking a target is provided.

[0041] Figure 4 The diagram below illustrates the extraction of three feature points.

[0042] Figure 5 This is a flowchart illustrating the testing phase of a neural network-based target motion / static state determination method according to an embodiment of the present invention.

[0043] Figure 6 This is a block diagram illustrating the structure of the target motion / static state judgment system based on neural networks according to the present invention. Detailed Implementation

[0044] The following are some embodiments of the present invention, intended to provide a basic understanding of the invention. They are not intended to identify key or decisive elements of the invention or to limit the scope of protection sought.

[0045] For the purposes of brevity and illustrativeness, the principles of the invention are described herein primarily with reference to exemplary embodiments thereof. However, those skilled in the art will readily recognize that the same principles can be equivalently applied to all types of neural network-based methods and systems for determining the motion and static states of tracking targets, and that these same principles can be implemented therein, with any such variations not departing from the true spirit and scope of this patent application.

[0046] Furthermore, reference is made in the following description to the accompanying drawings, which illustrate specific exemplary embodiments. Electrical, mechanical, logical, and structural modifications may be made to these embodiments without departing from the spirit and scope of the invention. Moreover, while features of the invention are disclosed in conjunction with only one of several embodiments, this feature may be combined with one or more other features of other embodiments if desired and / or advantageous for any given or identifiable function. Therefore, the following description should not be considered limiting in any sense, and the scope of the invention is defined by the appended claims and their equivalents.

[0047] Terms such as “possessing” and “comprising” indicate that, in addition to having units (modules) and steps that are directly and explicitly stated in the specification and claims, the technical solution of the present invention does not exclude the presence of other units (modules) and steps that are not directly or explicitly stated.

[0048] The main technical concept of the neural network-based target motion / static state judgment method of this invention lies in extracting feature points on the target tracked by LiDAR and accumulating feature points from multiple frames as feature vectors to distinguish motion / static states. Specifically, in the training phase, sequences of LiDAR-tracked targets are collected, feature vectors are extracted, and the ground truth values ​​of the target's motion / static state are labeled as training samples to train the parameters of the neural network. In the testing phase, the feature vector of each tracked target is extracted, input into the trained neural network for discrimination, and the motion / static classification result of the current tracked target is output.

[0049] In this way, by using neural networks to classify the static and dynamic states of targets, we can reduce the situation where static targets are misclassified as dynamic targets due to changes in the target point cloud morphology caused by occlusion, and reduce the situation where slow-moving pedestrians are detected as static simply by using speed determination, thereby improving the accuracy of classifying the static and dynamic states of targets tracked by lidar.

[0050] Figure 1 This is a flowchart illustrating the method for determining the dynamic and static states of a tracking target based on a neural network according to the present invention.

[0051] like Figure 1 As shown, the neural network-based target motion / static state determination method of the present invention includes:

[0052] Training Phase (or "Training Step") S100: For each tracked target in the target set, feature points are extracted, converted into feature vectors, and the ground truth values ​​of the target's dynamic and static states are labeled. The feature vectors and the corresponding ground truth values ​​of the dynamic and static states are used as training sample parameters to train the neural network, resulting in a trained neural network classification model for classifying dynamic and static states; and

[0053] In the testing phase (or "testing phase") S200, feature points of the tracked target are extracted, the feature points are converted into feature vectors, and input into the neural network classification model obtained in step S100 to obtain the dynamic and static state classification results of the tracked target.

[0054] Next, the specific content of the training and testing phases will be explained.

[0055] Figure 2 This is a flowchart illustrating the training phase of a neural network-based target motion / static state judgment method according to an embodiment of the present invention.

[0056] like Figure 2 As shown, the training phase includes:

[0057] Step S101: Obtain a set of tracked targets from the lidar, wherein the set contains a sequence of tracked targets with different tracking IDs. Here, the ID is used to distinguish different targets. For each tracked target, each target has a specific ID. The sequence of tracked targets refers to the same target whose position changes over time, forming a driving trajectory.

[0058] Step S102: For each tracked target, extract feature points for every N consecutive frames of tracking results. Here, N is preferably N=5, that is, for each tracked target sequence, extract feature points in the tracking target point cloud for 5 consecutive frames of tracking results. In this embodiment, N=5 is preferably selected. In practical applications, N can also be selected as other numbers, but too few numbers can easily lead to inaccurate detection, and too many numbers can lead to detection delay.

[0059] Step S103: Convert feature points into feature vectors. The specific algorithm for converting feature points into feature vectors will be described below.

[0060] Step S104: For the feature vector labeled with the true value of the dynamic and static state of the tracked target (i.e., the true value label), store the feature vector and label as a training sample;

[0061] Step S105: Obtain the set of dynamic and static state data of the tracked target, as the parameter set for training the neural network; and

[0062] Step S106: Train the neural network using the above parameter set to obtain a trained neural network classification model for classifying dynamic and static states.

[0063] Regarding the tracking target sequence mentioned in steps S101 and S102, in Figure 3 The example demonstrates a tracking time series formed by tracking targets. Each target has a unique ID, and the tracking target sequence refers to the same target whose position changes over time, such as... Figure 3 As shown, this forms a driving trajectory. Here, the target refers to the obstacle extracted from the lidar point cloud. Therefore, the "target" essentially refers to the obstacle composed of the point cloud, which is different from the lidar point cloud on the ground and outside the road.

[0064] Next, the specific details of extracting feature points in step S102 will be explained.

[0065] In this embodiment, three feature points are extracted. Figure 4 The diagram below illustrates the extraction of three feature points.

[0066] like Figure 4 As shown, Figure 4The figure shows the three feature points A, B, and C extracted in this embodiment. Point A represents the point closest to the vehicle in the point cloud data of the currently tracked target, point B represents the center point of the long side scanned by the lidar, and point C represents the centroid of the target point cloud.

[0067] Current tracking methods typically only use the target's center point. When a LiDAR scans a vehicle target, it can only scan a portion of it, making the center point unstable. In this implementation, three feature points are selected, which can achieve redundancy in the stability of the feature points at different distances and angles. This provides richer features for subsequent machine learning algorithms, resulting in more stable results.

[0068] Here, we will explain the method for obtaining the three feature points.

[0069] Point A, the closest point in the target point cloud, is extracted by traversing all points in the point cloud data of the tracked target and obtaining the point closest to the origin of the lidar (i.e., the origin of the coordinate system of the lidar sensor). Point A is then taken as point A.

[0070] Point B, the center point of the longer side of the target point cloud, is extracted by using a straight-line fitting method on the point cloud data to extract the longer edge of the point cloud and finding the center point of the edge. Point A is then used as point B. The straight-line extraction method can use least squares, RANSAC, or other straight-line fitting methods.

[0071] Point C, the centroid of the target point cloud, is extracted by summing and averaging the three-dimensional coordinates of each point in the point cloud data.

[0072] Next, the specific process of converting feature points into feature vectors in step S103 will be explained. Converting feature points into feature vectors involves expanding the time series of the above three feature points in one dimension, as follows:

[0073] Let the current time be t, and let Pn,t(X) be the nearest point extracted at time t. n,t , Y n,t The center point of the longer side is Pe,t(X). e,t , Y e,t ), with centroid Pc,t(X) c,t , Y c,t The eigenvector is a one-dimensional expansion of the time series of three feature points, represented as Ft=[X]. n,t , Y n,t , X e,t , Y e,t, X c,t , Y c,t, X n,t-1 , Yn,t-1 , X e,t-1 Ye t-1 , X c,t , Y c,t-1 ,..., X n,t-N , Y n,t-N , X e,t-N , Y e,t-N , X c,t-N , Y c,t-N ], where N takes the value of 5 (i.e., take 5 frames), and the label of the target's dynamic and static state true value is denoted as Mt, where Mt takes the value of 0 or 1. Then a training sample data is represented as Si=[Ft, Mt].

[0074] In step S106, the neural network is trained using the aforementioned parameters to obtain a trained neural network classification model for classifying dynamic and static states. In this embodiment, preferably, to reduce parameters and computational load while ensuring accuracy, a 5-layer neural network is used as the classification model. The input layer consists of 1x30 feature vectors, the hidden layers have 50, 20, and 10 nodes respectively, and the output layer is a 2x1 vector. The parameters of the neural network are trained using methods such as stochastic gradient descent to obtain the neural network classification model for classifying dynamic and static states.

[0075] Figure 5 This is a flowchart illustrating the testing phase of a neural network-based target motion / static state determination method according to an embodiment of the present invention.

[0076] refer to Figure 5 During the testing phase, the neural network classification model obtained above is used to determine the dynamic and static states of the tracking target. For the tracking target at the current moment, preferably, for example, the tracking results of the last 5 frames are taken, feature points are extracted and converted into feature vectors, and the feature vectors are input into the neural network classification model obtained above to obtain the classification result of the target's dynamic state.

[0077] like Figure 5 As shown, the specific steps in the testing phase include:

[0078] Step S201: Obtain the set of targets tracked by the lidar;

[0079] Step S202: For each tracked target, extract feature points from the latest 5 frames of tracking results;

[0080] Step S203: Convert feature points into feature vectors; and

[0081] Step S204: Input the feature vector into the trained neural network classification model to determine the dynamic and static state of the current tracked target and obtain the dynamic and static state classification result.

[0082] The method for converting feature points into feature vectors in step S203 is the same as that in step S103.

[0083] In the testing phase, as a further preferred approach, after step S204, Bayesian filtering is applied to the obtained dynamic and static state classification results of the tracked target, thereby obtaining more stable dynamic and static state classification results.

[0084] The above describes the target motion and static state judgment method based on neural network of the present invention. Next, the target motion and static state judgment system based on neural network of the present invention will be described.

[0085] Figure 6 This is a block diagram illustrating the structure of the target motion / static state judgment system based on neural networks according to the present invention.

[0086] like Figure 6 As shown, the neural network-based target motion / static state determination system of the present invention includes:

[0087] Training module 100 is used to extract feature points for each tracked target in the target set, convert the feature points into feature vectors, and label the ground truth values ​​of the target's dynamic and static states as training sample parameters to train the neural network, thereby obtaining a trained neural network classification model; and

[0088] The test module 200 is used to extract the feature vector of the tracked target, input it into the neural network classification model, and obtain the dynamic and static state classification result of the tracked target.

[0089] The training module 100 includes:

[0090] The first feature extraction submodule 110 is used to extract feature points for each tracking target in the tracking target set;

[0091] The first feature transformation submodule 120 is used to convert the feature points into feature vectors; and

[0092] The model training submodule 130 is used to label the transformed feature vector with the true values ​​of the dynamic and static states of the tracking target, and use them as training sample parameters to train the neural network to obtain the neural network classification model.

[0093] Preferably, in the first feature extraction submodule 110, the nearest point to the vehicle, the center point of the long edge, and the centroid are extracted as the feature points. Specifically, in the first feature extraction submodule 110, the method for extracting the nearest point to the vehicle is as follows: the method for extracting the nearest point to the vehicle is to traverse all points in the lidar point cloud to obtain the point closest to the lidar origin; the method for extracting the center point of the long edge is to use a straight line fitting method on the lidar point cloud to extract the longer edge of the point cloud and find the center point of the edge; the method for extracting the centroid is to directly sum and average the three-dimensional coordinates of each point in the lidar point cloud, wherein the lidar point cloud is composed of points collected by the lidar.

[0094] Thus, in the first feature transformation submodule, the transformation from feature points to feature vectors is performed as follows: Let the current time be t, and let the nearest point extracted at time t be Pn,t(X). n,t , Y n,t The center point of the longer side is Pe,t(X). e,t ,Y e,t ), with centroid Pc,t(X) c,t , Y c,t The eigenvector is a one-dimensional expansion of the time series of three feature points, represented as Ft=[X]. n,t , Y n,t , X e,t , Y e,t, X c,t , Y c,t, X n,t-1 , Y n,t-1 , X e,t-1 Ye t-1 , X c,t ,Y c,t-1 ,..., X n,t-N , Y n,t-N , X e,t-N , Y e,t-N , X c,t-N , Y c,t-N ], where N is 5 (i.e., 5 frames), and the label of the target's dynamic and static state truth value is denoted as Mt, where Mt is 0 or 1. A training sample data is then represented as Si = [Ft, Mt]. Further, the test module 200 includes:

[0095] The second feature extraction submodule 210 is used to extract feature points for each tracking target in the tracking target set;

[0096] The second feature transformation submodule 220 is used to convert feature points into feature vectors; and

[0097] The test submodule 230 is used to input the feature vector into the trained neural network classification model to determine the dynamic and static state of the currently tracked target and obtain the dynamic and static state classification result.

[0098] Preferably, in the test submodule 230, the obtained classification results of the dynamic and static states of the tracked target are further filtered using Bayesian filtering.

[0099] As described above, the target motion and static state judgment method and system based on neural networks of the present invention can solve the misjudgment problem caused by simply using speed judgment (e.g., identifying a static target that is blocked as a moving target, or judging a slow pedestrian as a stationary target) by using a neural network method to judge the motion and static state of the target tracked by the lidar. This can improve the accuracy of the judgment results.

[0100] The present invention also provides a computer-readable medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements the aforementioned method for determining the dynamic and static states of a tracking target based on a neural network.

[0101] The present invention also provides a computer device, including a storage module, a processor, and a computer program stored on the storage module and executable on the processor, characterized in that the processor executes the computer program to implement the neural network-based target motion and static state judgment method.

[0102] The above examples primarily illustrate the target motion / static state determination method and system based on neural networks of the present invention. Although only some specific embodiments of the present invention have been described, those skilled in the art should understand that the present invention can be implemented in many other forms without departing from its spirit and scope. Therefore, the examples and embodiments shown are considered illustrative rather than restrictive, and the present invention may cover various modifications and substitutions without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. A method for determining the dynamic and static state of a tracking target based on a neural network, characterized in that, include: The training steps are as follows: for each tracked target in the target set from the vehicle's LiDAR, feature points are extracted, the feature points are converted into feature vectors, and the true values ​​of the dynamic and static states of the tracked targets are labeled. The feature vectors and the corresponding true values ​​of the dynamic and static states are used as training sample parameters to train the neural network, and the trained neural network classification model is obtained. as well as The testing process involves extracting feature points of the tracked target, converting them into feature vectors, and then inputting them into the neural network classification model to obtain the dynamic / static state classification result of the tracked target. The training steps include: The first feature extraction sub-step extracts multiple feature points for each tracking target in the tracking target set; The first feature transformation sub-step converts the plurality of feature points into feature vectors; and The model training sub-step involves labeling the transformed feature vectors with the ground truth values ​​of the target's dynamic and static states. These feature vectors and their corresponding ground truth values ​​are then used as training sample parameters to train the neural network, resulting in a neural network classification model. In the first feature extraction sub-step, the point closest to the vehicle, the center point of the long side, and the centroid of the tracked target are extracted as the feature points. The first feature transformation sub-step includes: Let the current time be t, and let Pn,t(X) be the nearest point extracted at time t. n,t Y n,t The center point of the longer side is Pe,t(X). e,t Y e,t ), with centroid Pc,t(X) c,t Y c,t The eigenvector is a one-dimensional expansion of the time series of three feature points, represented as Ft=[X]. n,t Y n,t , X e,t Y e,t, X c,t Y c,t, X n,t-1 Y n,t-1 , X e,t-1 Ye t-1 , X c,t-1 ,Y c,t-1 ,..., X n,t-N Y n,t-N , X e,t-N Y e,t-N , X c,t-N Y c,t-N ], where N takes the value of 5, and the label of the target's dynamic and static state true value is denoted as Mt, where Mt takes the value of 0 or 1. Then a training sample data is represented as Si=[Ft, Mt].

2. The method for determining the dynamic and static states of a tracking target based on a neural network as described in claim 1, characterized in that, In the first feature extraction sub-step, the method for extracting the point closest to the vehicle is to traverse all points in the point cloud data of the tracking target and obtain the point closest to the origin of the lidar; the method for extracting the center point of the long side is to use a straight line fitting method on the point cloud data to extract the longer edge of the point cloud and find the center point of the edge; the method for extracting the centroid is to directly sum and average the three-dimensional coordinates of each point in the point cloud data, wherein the point cloud data is composed of points collected by the lidar.

3. The method for determining the dynamic and static states of a tracking target based on a neural network as described in claim 1, characterized in that, The testing steps include: The second feature extraction sub-step involves extracting multiple feature points from the tracking results for each tracking target. The second feature transformation sub-step converts the plurality of feature points into feature vectors; and The testing sub-step involves inputting the feature vector into the neural network classification model, which then determines the dynamic or static state of the currently tracked target and obtains the dynamic or static state classification result.

4. The method for determining the dynamic and static states of a tracking target based on a neural network as described in claim 3, characterized in that, In the test sub-step, the obtained classification results of the dynamic and static states of the tracked target are further filtered using Bayesian filtering.

5. A system for determining the dynamic and static states of a tracking target based on a neural network, characterized in that, include: The training module is used to extract feature points for each tracked target in the target set from the vehicle's LiDAR, convert the feature points into feature vectors, and label the true values ​​of the dynamic and static states of the tracked targets. The feature vectors and the corresponding true values ​​of the dynamic and static states are used as training sample parameters to train the neural network and obtain the trained neural network classification model. as well as The testing module is used to extract feature points of the tracked target, convert them into feature vectors, and then input them into the neural network classification model to obtain the dynamic and static state classification results of the tracked target. The training module includes: The first feature extraction submodule is used to extract multiple feature points for each tracking target in the tracking target set; The first feature transformation submodule is used to convert the plurality of feature points into feature vectors; and The model training submodule is used to label the transformed feature vectors with the ground truth values ​​of the target's dynamic and static states. The feature vectors and their corresponding ground truth values ​​are then used as training sample parameters to train the neural network, resulting in a neural network classification model. In the first feature extraction submodule, the point closest to the vehicle, the center point of the long side, and the centroid of the tracked target are extracted as the feature points. In the first feature transformation submodule, let the current time be t, and let the nearest point extracted at time t be Pn,t(X). n,t Y n,t The center point of the longer side is Pe,t(X). e,t Y e,t ), with centroid Pc,t(X) c,t Y c,t The eigenvector is a one-dimensional expansion of the time series of three feature points, represented as Ft=[X]. n,t Y n,t , X e,t Y e,t, X c,t Y c,t, X n,t-1 ,Y n,t-1 , X e,t-1 Ye t-1 , X c,t-1 Y c,t-1 ,..., X n,t-N Y n,t-N , X e,t-N Y e,t-N , X c,t-N Y c,t-N ], where N takes the value of 5, and the label of the target's dynamic and static state true value is denoted as Mt, where Mt takes the value of 0 or 1. Then a training sample data is represented as Si=[Ft, Mt].

6. The target motion / static state judgment system based on neural network as described in claim 5, characterized in that, In the first feature extraction submodule, the method for extracting the point closest to the vehicle is to traverse all points in the point cloud data of the tracking target and obtain the point closest to the origin of the lidar; the method for extracting the center point of the long side is to use a straight line fitting method on the point cloud data to extract the longer edge of the point cloud and find the center point of the edge; the method for extracting the centroid is to directly sum and average the three-dimensional coordinates of each point in the point cloud data, wherein the point cloud data is composed of points collected by the lidar.

7. The target motion / static state judgment system based on neural network as described in claim 5, characterized in that, The testing module includes: The second feature extraction submodule is used to extract multiple feature points for each tracking target in the tracking target set; The second feature transformation submodule is used to convert the plurality of feature points into feature vectors; and The testing submodule is used to input the feature vector into the neural network classification model, which then determines the dynamic or static state of the currently tracked target and obtains the dynamic or static state classification result.

8. The target motion / static state judgment system based on neural network as described in claim 7, characterized in that, In the test submodule, the obtained classification results of the dynamic and static states of the tracked target are further filtered using Bayesian filtering.

9. A computer-readable medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the neural network-based method for determining the dynamic and static states of a tracking target as described in any one of claims 1 to 4.

10. A computer device, comprising a storage module, a processor, and a computer program stored on the storage module and executable on the processor, characterized in that, When the processor executes the computer program, it implements the neural network-based target motion / static state judgment method according to any one of claims 1 to 4.

Images