Adaptive threshold association method and device based on time domain convolutional neural network

Abstract

The present application relates to the field of data processing technology, and in particular to an adaptive threshold association method and device based on a time-domain convolutional neural network, including: taking a reference scattering point as the center, screening out potential homologous scattering points within a preset association gate; calculating the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point, and forming time series data with the weighted Euclidean distance and the association threshold of the previous adjacent period; inputting the time series data into an association model based on a time-domain convolutional neural network, and obtaining a set of homologous scattering points within a preset association gate according to the output result of the association model; calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, taking the maximum weighted Euclidean distance as the association threshold of the current period, and saving the association threshold. Thus, the problem of false association or missed association caused by the use of a fixed association threshold is solved, and the association threshold can be adaptively adjusted through a method based on TCN.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an adaptive threshold association method and device based on a time-domain convolutional neural network. Background Art

[0002] In order to determine whether the currently detected target can be associated with the existing trajectory, association is often used. In order to facilitate target tracking, when processing multiple echo points generated by larger targets such as cars within the irradiation range of the vehicle-mounted millimeter-wave radar, clustering processing is usually performed first to associate and merge multiple echo points originating from the same target (homologous echo points) into one echo point.

[0003] In the target tracking process, especially in application scenarios such as intelligent driving and assisted driving, in order to accurately determine whether the currently detected target can be associated with the existing trajectory, a common method is to associate the target's echo points. When processing large-sized targets (such as cars) within the irradiation range of the vehicle-mounted millimeter-wave radar, multiple echo points will be generated. Usually, the echo points are clustered, that is, multiple echo points from the same target (homologous echo points) are associated and merged into one echo point to facilitate subsequent target tracking.

[0004] In the related art, a commonly used association algorithm is the Nearest Neighbor Data Association (NNDA) algorithm.

[0005] However, the traditional NNDA algorithm uses a fixed association threshold, which cannot adapt to the changes in the distribution of target scattering points. Therefore, it is easy to cause problems of false association or missed association, which needs to be solved urgently. Summary of the invention

[0006] The present application provides an adaptive threshold association method and device based on a time-domain convolutional neural network to solve the problems of false association or missed association caused by the use of a fixed association threshold in the background technology. Through a method based on TCN (Temporal Convolutional Networks, time-domain convolutional neural network), the association threshold can be adaptively adjusted to reduce the problems of false association or missed association and improve the accuracy of target association.

[0007] The first aspect of the present application provides an adaptive threshold association method based on a time-domain convolutional neural network, comprising the following steps:

[0008] Determine a reference scattering point, and screen out a plurality of potential homologous scattering points within a preset association gate with the reference scattering point as the center;

[0009] Calculating the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point, and combining the weighted Euclidean distance with the correlation threshold of the previous adjacent cycles to form time series data;

[0010] Inputting the time series data into a designed association model based on a time-domain convolutional neural network, and obtaining a set of homologous scattering points within the preset association gate according to an output result of the association model;

[0011] The weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, the maximum weighted Euclidean distance is used as the association threshold of the current cycle, and the association threshold is saved.

[0012] According to an embodiment of the present application, obtaining the set of homologous scattering points within the association gate according to the output result of the association model includes:

[0013] Based on the output result, selecting target potential homologous scattering points that meet preset homology conditions from the multiple potential homologous scattering points;

[0014] The target potential homologous scattering point is added to the homologous scattering point set in the preset association gate.

[0015] According to an embodiment of the present application, the calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, and taking the maximum weighted Euclidean distance as the association threshold of the current cycle, includes:

[0016] sequentially calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point until all homologous scattering points in the homologous scattering point set are traversed to obtain a plurality of weighted Euclidean distances;

[0017] Comparing the multiple weighted Euclidean distances, and selecting the largest weighted Euclidean distance;

[0018] The maximum weighted Euclidean distance is used as the association threshold of the current cycle.

[0019] According to one embodiment of the present application,

[0020] Before inputting the time series data into the designed association model based on the time domain convolutional neural network, it also includes:

[0021] Acquire vehicle-mounted radar scattering point data of multiple targets to obtain a vehicle-mounted radar scattering point data set, and calibrate the scattering points of the same target in the scattering point data set;

[0022] The calibrated scattering point data set is divided into a training set, a validation set and a test set according to a preset ratio, and the training set is used to train an initial association model based on a time-domain convolutional neural network. After the training is completed, the validation set is used to verify the initial association model based on the time-domain convolutional neural network. After the verification is completed, the test set is used to test the initial association model based on the time-domain convolutional neural network to obtain the designed association model based on the time-domain convolutional neural network.

[0023] According to the adaptive threshold association method based on time-domain convolutional neural network provided by the embodiment of the present application, with the reference scattering point as the center, multiple potential homologous scattering points within the preset association gate are screened out, the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point is calculated, and the weighted Euclidean distance and the association threshold of the previous adjacent period are combined into time series data, and the designed association model based on time-domain convolutional neural network is input to obtain the homologous scattering point set within the preset association gate; the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, the maximum weighted Euclidean distance is used as the association threshold of the current period, and the association threshold is saved. Thus, the problem of false association or missed association caused by the fixed association threshold threshold in the background technology is solved. Through the method based on TCN, the association threshold can be adaptively adjusted to reduce the problem of false association or missed association, and the accuracy of target association is improved.

[0024] The second aspect of the present application provides an adaptive threshold association device based on a time-domain convolutional neural network, including:

[0025] A screening module, used to determine a reference scattering point, and screen out a plurality of potential homologous scattering points within a preset association gate with the reference scattering point as the center;

[0026] A calculation module, used for calculating the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point, and combining the weighted Euclidean distance with the correlation threshold of the previous adjacent cycles to form time series data;

[0027] A model output module, used for inputting the time series data into a designed association model based on a time-domain convolutional neural network, and obtaining a set of homologous scattering points within the preset association gate according to an output result of the association model;

[0028] The association threshold determination module is used to calculate the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, take the maximum weighted Euclidean distance as the association threshold of the current cycle, and save the association threshold.

[0029] According to one embodiment of the present application, the model output module is used to:

[0030] Based on the output result, selecting target potential homologous scattering points that meet preset homology conditions from the multiple potential homologous scattering points;

[0031] The target potential homologous scattering point is added to the homologous scattering point set in the preset association gate.

[0032] According to one embodiment of the present application, the association threshold determination module is used to:

[0033] sequentially calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point until all homologous scattering points in the homologous scattering point set are traversed to obtain a plurality of weighted Euclidean distances;

[0034] Comparing the multiple weighted Euclidean distances, and selecting the largest weighted Euclidean distance;

[0035] The maximum weighted Euclidean distance is used as the association threshold of the current cycle.

[0036] According to one embodiment of the present application, before the time series data is input into the designed association model based on the time domain convolutional neural network, the model output module is further used to:

[0037] Acquire vehicle-mounted radar scattering point data of multiple targets to obtain a vehicle-mounted radar scattering point data set, and calibrate the scattering points of the same target in the scattering point data set;

[0038] The calibrated scattering point data set is divided into a training set, a validation set and a test set according to a preset ratio, and the training set is used to train an initial association model based on a time-domain convolutional neural network. After the training is completed, the validation set is used to verify the initial association model based on the time-domain convolutional neural network. After the verification is completed, the test set is used to test the initial association model based on the time-domain convolutional neural network to obtain the designed association model based on the time-domain convolutional neural network.

[0039] According to the adaptive threshold association device based on time-domain convolutional neural network provided by the embodiment of the present application, with the reference scattering point as the center, multiple potential homologous scattering points within the preset association gate are screened out, the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point is calculated, and the weighted Euclidean distance and the association threshold of the previous adjacent period are combined into time series data, and the designed association model based on time-domain convolutional neural network is input to obtain the homologous scattering point set within the preset association gate; the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, the maximum weighted Euclidean distance is used as the association threshold of the current period, and the association threshold is saved. Thus, the problem of false association or missed association caused by the fixed association threshold threshold in the background technology is solved. Through the method based on TCN, the association threshold can be adaptively adjusted to reduce the problem of false association or missed association, and the accuracy of target association is improved.

[0040] The third aspect of the present application provides a vehicle, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the adaptive threshold association method based on a time-domain convolutional neural network as described in the above embodiment.

[0041] The fourth aspect of the present application provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable the computer to execute the adaptive threshold association method based on the time-domain convolutional neural network as described in the above embodiment.

[0042] Additional aspects and advantages of the present application will be given in part in the description below, and in part will become apparent from the description below, or will be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0044] Figure 1 A flowchart of an adaptive threshold association method based on a time-domain convolutional neural network provided according to an embodiment of the present application;

[0045] Figure 2 is an architecture diagram of a TCN-based data association model according to an embodiment of the present application;

[0046] Figure 3 It is a flowchart of an adaptive threshold association method based on a time-domain convolutional neural network according to an embodiment of the present application;

[0047] Figure 4Schematic diagram of a block diagram of an adaptive threshold association device based on a time-domain convolutional neural network according to an embodiment of the present application;

[0048] Figure 5 A schematic diagram of the structure of a vehicle provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] Embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.

[0050] The following describes an adaptive threshold association method and device based on a time-domain convolutional neural network according to an embodiment of the present application with reference to the accompanying drawings.

[0051] Before introducing the adaptive threshold association method based on a time-domain convolutional neural network in an embodiment of the present application, the nearest neighbor data association algorithm in the related art and the time-domain convolutional neural network involved in the adaptive threshold association method based on a time-domain convolutional neural network in the present application are first introduced.

[0052] The specific implementation steps of the nearest neighbor data association algorithm are as follows:

[0053] (1) Select a center point;

[0054] (2) Set a correlation gate, take the center point as the reference point, and use the correlation gate to screen out potential homologous echo points;

[0055] (3) Calculate the Euclidean distance (or Mahalanobis distance) between the potential homologous echo point in the association gate and the reference point, using Representative The distance between the potential homologous echo points and the reference point;

[0056] (4) Setting the association threshold ,judge and If , determine the The potential homologous echo points and the reference point are scattering points of the same target; otherwise, they are scattering points that do not belong to the same target;

[0057] (5) The scattering points determined as the same target within the association gate are associated and merged into one target point.

[0058] It is understandable that since traffic targets include cars, trucks, buses, vans, motorcycles and other types, their target sizes vary greatly, resulting in different distributions of scattering points for different targets. At the same time, the scattering characteristics of the target are also closely related to the incident angle of the radar radiation wave. The distribution of scattering points for the same target will change with changes in factors such as distance and irradiation angle. These factors together cause differences in the dynamic and static distribution of traffic target echo points.

[0059] However, the aforementioned nearest neighbor data association algorithm uses a fixed association threshold value, which cannot adapt to the changes in the distribution of target scattering points. Therefore, it is easy to cause the problem of false association or missed association.

[0060] In response to the problem that the threshold set in the nearest neighbor association process mentioned above is fixed and cannot adapt to changes in the scene, the present application provides an adaptive threshold association method based on a time domain convolutional neural network, which solves the problems of false association or missed association caused by the use of a fixed association threshold. Through the TCN-based method, the association threshold can be adaptively adjusted to reduce the problem of false association or missed association and improve the accuracy of target association.

[0061] The following is an introduction to the time domain convolutional neural network involved in the adaptive threshold association method based on the time domain convolutional neural network of this application.

[0062] The time-domain convolutional neural network (TCN) is a new network architecture composed of hollow and causal convolutional layers with the same input and output lengths. It can process time series data and only rely on past information to predict future output results. In traffic scenarios, the scanning cycle of the vehicle-mounted millimeter-wave radar is within 100ms. Compared with the scanning speed of the millimeter-wave radar, the movement of the target is slow. The millimeter-wave radar can obtain multiple frames of scanning data within a time when the radar electromagnetic wave incident angle and the target state change are small. During this time, the change in the distribution characteristics of the target's scattering points is negligible. The multiple frames of vehicle-mounted millimeter-wave radar scanning data obtained during this time are used to form time series data as the input of TCN. TCN predicts the association threshold of the next stage, which can achieve the purpose of adaptively adjusting the association threshold by predicting the association threshold in the next time period in real time based on the statistics of the data in a short period of time.

[0063] The following introduces the adaptive threshold association method based on time-domain convolutional neural network proposed in this application.

[0064] Specifically, Figure 1 A schematic flow chart of an adaptive threshold association method based on a time-domain convolutional neural network provided in an embodiment of the present application.

[0065] like Figure 1As shown, the adaptive threshold association method based on time domain convolutional neural network includes the following steps:

[0066] In step S101, a reference scattering point is determined, and a plurality of potential homologous scattering points within a preset association gate are screened out with the reference scattering point as the center.

[0067] Among them, the preset association gate is a spatial area centered on the reference scattering point and set according to factors such as the possible moving speed and direction of the target and the resolution of the radar scan. The shape and size of the association gate can be adjusted according to actual conditions to cover scattering points that may belong to the same target.

[0068] Specifically, the embodiment of the present application can select the center point of the target from the data scanned by the vehicle-mounted millimeter-wave radar as a reference scattering point, or select a scattering point with significant characteristics (such as high intensity and stability) as a reference scattering point.

[0069] Furthermore, in determining the reference scattering point Afterwards, the reference scattering point As the center, the preset association gate is used as the boundary. All possible scattering points are searched within the preset association gate as potential homologous scattering points according to the spatial position, velocity, acceleration and other characteristics of the scattering points. That is, the preset association gate is used to screen out the potential homologous scattering points. Potential homologous scattering points .

[0070] In step S102, the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point is calculated, and the weighted Euclidean distance and the correlation threshold of the previous adjacent cycles are combined into time series data.

[0071] It is understandable that the Euclidean distance is mainly used to measure the spatial geometric distance between two points. However, in dynamic scenarios, it is unreasonable to use this distance to distinguish whether the radar scattering points are the same target. For example, at a certain moment, the distance between the car and the motorcycle is close, that is, the spatial geometric distance is very close. At this time, if the original Euclidean distance is used for judgment, it is easy to merge the motorcycle and the car into one target, resulting in misassociation.

[0072] Therefore, this application improves the Euclidean distance by taking the target's speed and acceleration factors into account under the premise of using a constant acceleration model to unify the dimensions. Set as dimensionless weight value, Set the unit to The weight value of (seconds squared), Set the unit to (seconds to the fourth power) weight, and numerically , the improved Euclidean distance formula is as follows:

[0073]

[0074] in, is the weighted Euclidean distance between the potential homologous scattering point and the reference scattering point, are the geometric coordinates of the reference scattering point in the Cartesian coordinate system, For the preset associated gate The geometric coordinates of potential scattering points in the Cartesian coordinate system, is the velocity of the reference scattering point, is the acceleration of the reference scattering point, For the preset associated gate The velocity of a potential scattering point, For the preset associated gate The acceleration of a potential scattering point, , and The value of is obtained through data experiments in multiple scenarios.

[0075] Therefore, based on the above-mentioned improved Euclidean distance formula, the embodiment of the present application can calculate each potential homologous scattering point in the preset association gate. With reference scattering point The weighted Euclidean distance .

[0076] Furthermore, the weighted Euclidean distance calculated in this embodiment can be Compared with the past The correlation threshold of the period constitutes the time series data ,in .

[0077] For example, suppose you are currently in cycles, and assuming , then we need to use the weighted Euclidean distance And the correlation thresholds of the four cycles before the current cycle are used to construct time series data. The correlation threshold of the cycle is , No. The correlation threshold of the cycle is , No. The correlation threshold of the cycle is , No. The correlation threshold of the cycle is , then according to the weighted Euclidean distance And the time series data composed of the correlation thresholds of the four cycles before the current cycle is ,in, .

[0078] In step S103, the time series data is input into the designed association model based on the time domain convolutional neural network, and a set of homologous scattering points within a preset association gate is obtained according to the output result of the association model.

[0079] Among them, the architecture of the designed association model based on time domain convolutional neural network is as follows Figure 2 As shown in the figure, the designed association model based on time-domain convolutional neural network consists of TCN module, reshaping unit (Reshape) and fully connected layer module. The TCN module consists of 4 layers of one-dimensional hole causal convolutional layers. The input sequence data length of each layer is padded as needed. The feature vector output by the TCN module is sent to the fully connected layer module after the reshaping unit (Reshape). The fully connected layer module consists of 3 fully connected layers and ReLU module.

[0080] Specifically, for potential homologous scattering points within the preset association gate , which is different from the reference scattering point The weighted Euclidean distance The correlation thresholds with the previous adjacent consecutive cycles constitute the time series data , this time series data is used as the input of the TCN module. The TCN module infers the potential homologous scattering points within the preset association gate based on the association threshold of the previous time and the weighted Euclidean distance of the current time. With reference scattering point The relationship between the two and outputs potential homologous scatter points With reference scattering point This feature vector is input into the fully connected layer module, and the fully connected layer module infers the potential homologous scattering points based on the feature vector. With reference scattering point If the fully connected layer module outputs "0", it represents a potential homologous scattering point. With reference scattering point Non-homologous scattering points, i.e. potential homologous scattering points With reference scattering point Belong to different targets; if the fully connected layer module outputs all "1", it represents a potential homologous scattering point With reference scattering point Homologous scattering points, i.e. potential homologous scattering points With reference scattering point Belong to the same target.

[0081] Further, in some embodiments, obtaining a set of homologous scattering points within an association gate according to an output result of the association model includes: based on the output result, screening out target potential homologous scattering points that meet preset homology conditions from multiple potential homologous scattering points; and adding the target potential homologous scattering points to the set of homologous scattering points within the preset association gate.

[0082] Specifically, according to the output results of the association model, the potential homologous scattering points with an output result of "0" (i.e., non-homologous scattering points) are screened out, and the target potential homologous scattering points with an output result of "1" (i.e., homologous scattering points) are screened out, and the target potential homologous scattering points are added to the homologous scattering point set in the preset association gate. .

[0083] In step S104, the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, the maximum weighted Euclidean distance is used as the association threshold of the current cycle, and the association threshold is saved.

[0084] Further, in some embodiments, the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, and the maximum weighted Euclidean distance is used as the association threshold of the current cycle, including: sequentially calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point until all homologous scattering points in the homologous scattering point set are traversed to obtain multiple weighted Euclidean distances; comparing the multiple weighted Euclidean distances to screen out the maximum weighted Euclidean distance; and using the maximum weighted Euclidean distance as the association threshold of the current cycle.

[0085] Specifically, the embodiment of the present application can calculate the homologous scattering point set in sequence based on the above improved Euclidean distance formula: Each homologous scattering point in The weighted Euclidean distance is calculated until all homologous scattering points in the homologous scattering point set are traversed to obtain multiple weighted Euclidean distances, and the largest weighted Euclidean distance is selected from the multiple weighted Euclidean distances. , and the maximum weighted Euclidean distance The correlation threshold is used as the correlation threshold of the current period and saved to construct the time series data of the next period.

[0086] Therefore, the embodiment of the present application combines the newly saved association threshold with other historical data to form new time series data, and uses the new time series data as the input of the TCN association module of the next period to continue to make association judgments on the target scattering points, thereby achieving continuous updating and optimization of the association threshold, making it more adaptable to the dynamic changes of target scattering points in different traffic scenarios.

[0087] Furthermore, in some embodiments, before inputting the time series data into the designed association model based on the time-domain convolutional neural network, it also includes: acquiring vehicle-mounted radar scattering point data of multiple targets to obtain a vehicle-mounted radar scattering point data set, and calibrating the scattering points of the same target in the scattering point data set; dividing the calibrated scattering point data set into a training set, a validation set and a test set according to a preset ratio, and using the training set to train the initial association model based on the time-domain convolutional neural network, and after the training is completed, using the validation set to verify the initial association model based on the time-domain convolutional neural network, and after the verification is completed, using the test set to test the initial association model based on the time-domain convolutional neural network to obtain the designed association model based on the time-domain convolutional neural network.

[0088] The preset ratio may be 6:2:2, which is not specifically limited here.

[0089] Specifically, first, the vehicle-borne millimeter-wave radar scattering point data of common targets in multiple traffic scenes are collected to obtain a vehicle-borne millimeter-wave radar scattering point data set, where common targets can be trucks, vans, cars, buses, motorcycles, bicycles, pedestrians, etc. For specific targets in each scene, it is necessary to ensure that the millimeter-wave radar scattering point data of multiple continuous scanning cycles are collected during the collection process.

[0090] Furthermore, the scattering points from the same target (i.e., homologous scattering points) in the collected scattering point data set are calibrated, and the weighted Euclidean distance of the homologous target scattering points in each scanning cycle is calculated. The maximum weighted Euclidean distance of the homologous target scattering points in each scanning cycle is used as the association threshold of the target in that cycle, and the corresponding scenes and cycles are numbered.

[0091] Furthermore, the calibrated vehicle-mounted millimeter-wave radar scattering point data of each target in a continuous scanning cycle under a specific scenario is divided into a training set, a validation set, and a test set according to a preset ratio (such as 6:2:2). The cycle order needs to be maintained during the division process and cannot be disrupted, and the scene number must be retained.

[0092] For example, assuming that there are a total of 100 scanning cycles of vehicle-mounted millimeter-wave radar scattering point data, the calibrated scattering point data set is divided into a training set, a validation set, and a test set according to a ratio of 6:2:2, that is, the scattering point data of 60 scanning cycles are divided into a training set, the scattering point data of 20 scanning cycles are divided into a validation set, and the scattering point data of 20 scanning cycles are divided into a test set. The order of the scanning cycles in the training set, the validation set, and the test set cannot be disrupted. For example, taking the order of every 8 scanning cycles as an example, the scattering point data of 60 scanning cycles in the training set is divided into a validation set, and the scattering point data of 20 scanning cycles is divided into a test set. The scattering point data of the 1st to 8th scanning cycles of the scattering point data of the training set are taken as the first group of time series data (the order of the 1st to 8th cycles cannot be disrupted), and the scattering point data of the 9th to 16th scanning cycles of the scattering point data of the 60 scanning cycles in the training set are taken as the second group of time series data (the order of the 9th to 16th cycles cannot be disrupted), and so on. Since the last group of data is the scattering point data of the 57th to 60th scanning cycles, which does not meet 8 scanning cycles, the scattering point data of the last 57th to 60th scanning cycles can be discarded.

[0093] Furthermore, the data on the training set is used to train the initial association model based on the time-domain convolutional neural network, and after the training is completed, the validation set is used to verify the initial association model based on the time-domain convolutional neural network, and the test set is used to test the initial association model based on the time-domain convolutional neural network, and finally the designed association model based on the time-domain convolutional neural network is obtained.

[0094] Therefore, by training and testing the association model based on the time-domain convolutional neural network, it is ensured that the association model can effectively handle millimeter-wave radar scattering points in different traffic scenarios, thereby improving the accuracy of target association.

[0095] In order to facilitate those skilled in the art to more clearly and intuitively understand the adaptive threshold association method based on the time domain convolutional neural network in the embodiment of the present application, the following is combined with Figure 3 Provide detailed explanation.

[0096] like Figure 3 As shown, the adaptive threshold association method based on time domain convolutional neural network includes the following steps:

[0097] S301, start.

[0098] S302, selecting a reference scattering point.

[0099] S303, defining an associated area (associated gate).

[0100] S304, calculating the weighted Euclidean distance between the potential homologous scattering point and the reference scattering point.

[0101] S305, constructing time series data.

[0102] S306: Use the TCN model to perform association inference.

[0103] S307, determining whether the potential homologous scattering points are associable with the reference scattering points, if so, executing S309, otherwise, executing S308.

[0104] S308, discard the potential homologous scattering point.

[0105] S309, determining whether the weighted Euclidean distance between the potential homologous scattering point and the reference scattering point is the largest, if yes, executing S311, otherwise, executing S310.

[0106] S310, do not update the maximum Euclidean distance.

[0107] S311, update the maximum Euclidean distance.

[0108] S312, determining whether the traversal of potential homologous scattering points in the association gate is complete, if yes, executing S313, otherwise, executing S304.

[0109] S313: Use the maximum Euclidean distance as the association threshold of this cycle.

[0110] S314, end.

[0111] According to the adaptive threshold association method based on time-domain convolutional neural network proposed in the embodiment of the present application, with the reference scattering point as the center, multiple potential homologous scattering points within the preset association gate are screened out, the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point is calculated, and the weighted Euclidean distance and the association threshold of the previous adjacent period are combined into time series data, and the designed association model based on time-domain convolutional neural network is input to obtain the homologous scattering point set within the preset association gate; the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, the maximum weighted Euclidean distance is used as the association threshold of the current period, and the association threshold is saved. Thus, the problem of false association or missed association caused by the fixed association threshold threshold in the background technology is solved. Through the method based on TCN, the association threshold can be adaptively adjusted to reduce the problem of false association or missed association, and the accuracy of target association is improved.

[0112] Next, an adaptive threshold association device based on a time-domain convolutional neural network proposed according to an embodiment of the present application is described with reference to the accompanying drawings.

[0113] Figure 4 It is a block diagram of an adaptive threshold association device based on a time-domain convolutional neural network according to an embodiment of the present application.

[0114] like Figure 4As shown, the adaptive threshold association device 10 based on time domain convolutional neural network includes: a screening module 100, a calculation module 200, a model output module 300 and an association threshold determination module 400.

[0115] Among them, the screening module 100 is used to determine the reference scattering point, and with the reference scattering point as the center, screen out multiple potential homologous scattering points within a preset association gate; the calculation module 200 is used to calculate the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point, and the weighted Euclidean distance and the association threshold of the previous adjacent period form time series data; the model output module 300 is used to input the time series data into the designed association model based on the time domain convolutional neural network, and obtain the homologous scattering point set within the preset association gate according to the output result of the association model; the association threshold determination module 400 is used to calculate the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, use the largest weighted Euclidean distance as the association threshold of the current period, and save the association threshold.

[0116] Furthermore, in some embodiments, the model output module 300 is used to: based on the output result, screen out target potential homologous scattering points that meet preset homology conditions from multiple potential homologous scattering points; and add the target potential homologous scattering points to the set of homologous scattering points in the preset association gate.

[0117] Furthermore, in some embodiments, the association threshold determination module 400 is used to: sequentially calculate the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, until all homologous scattering points in the homologous scattering point set are traversed to obtain multiple weighted Euclidean distances; compare the multiple weighted Euclidean distances to screen out the largest weighted Euclidean distance; and use the largest weighted Euclidean distance as the association threshold of the current period.

[0118] Furthermore, in some embodiments, before inputting the time series data into the designed association model based on the time-domain convolutional neural network, the model output module 300 is also used to: obtain vehicle-mounted radar scattering point data of multiple targets to obtain a vehicle-mounted radar scattering point data set, and calibrate the scattering points of the same target in the scattering point data set; divide the calibrated scattering point data set into a training set, a verification set and a test set according to a preset ratio, and use the training set to train the initial association model based on the time-domain convolutional neural network, and after the training is completed, use the verification set to verify the initial association model based on the time-domain convolutional neural network, and after the verification is completed, use the test set to test the initial association model based on the time-domain convolutional neural network to obtain the designed association model based on the time-domain convolutional neural network.

[0119] It should be noted that the aforementioned explanation of the embodiment of the adaptive threshold association method based on the time-domain convolutional neural network is also applicable to the adaptive threshold association device based on the time-domain convolutional neural network of this embodiment, and will not be repeated here.

[0120] According to the adaptive threshold association device based on time-domain convolutional neural network proposed in the embodiment of the present application, with the reference scattering point as the center, multiple potential homologous scattering points within the preset association gate are screened out, the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point is calculated, and the weighted Euclidean distance and the association threshold of the previous adjacent period are combined into time series data, and the designed association model based on time-domain convolutional neural network is input to obtain the homologous scattering point set within the preset association gate; the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point is calculated, the maximum weighted Euclidean distance is used as the association threshold of the current period, and the association threshold is saved. Thus, the problem of false association or missed association caused by the fixed association threshold threshold in the background technology is solved. Through the method based on TCN, the association threshold can be adaptively adjusted to reduce the problem of false association or missed association, and the accuracy of target association is improved.

[0121] Figure 5 A schematic diagram of the structure of a vehicle provided in an embodiment of the present application. The vehicle may include:

[0122] A memory 501 , a processor 502 , and a computer program stored in the memory 501 and executable on the processor 502 .

[0123] When the processor 502 executes the program, the adaptive threshold association method based on the time-domain convolutional neural network provided in the above embodiment is implemented.

[0124] Furthermore, the vehicle also includes:

[0125] The communication interface 503 is used for communication between the memory 501 and the processor 502 .

[0126] The memory 501 is used to store computer programs that can be executed on the processor 502 .

[0127] The memory 501 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0128] If the memory 501, the processor 502 and the communication interface 503 are implemented independently, the communication interface 503, the memory 501 and the processor 502 can be connected to each other through a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0129] Optionally, in a specific implementation, if the memory 501, the processor 502 and the communication interface 503 are integrated on a chip, the memory 501, the processor 502 and the communication interface 503 can communicate with each other through an internal interface.

[0130] The processor 502 may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0131] An embodiment of the present application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned adaptive threshold association method based on a time-domain convolutional neural network.

[0132] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0133] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0134] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, fragment or portion of code comprising one or more executable instructions for implementing the steps of a custom logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.

[0135] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection with one or N wirings (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0136] It should be understood that the various parts of the present application can be implemented by hardware, software, firmware or a combination thereof. In the above embodiment, N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0137] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0138] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0139] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limiting the present application. A person of ordinary skill in the art may change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. An adaptive threshold association method based on time-domain convolutional neural network, characterized in that: The following steps are involved: Determine a reference scattering point, and screen out a plurality of potential homologous scattering points within a preset association gate with the reference scattering point as the center; Calculating the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point, and combining the weighted Euclidean distance with the correlation threshold of the previous adjacent cycles to form time series data; Inputting the time series data into a designed association model based on a time-domain convolutional neural network, and obtaining a set of homologous scattering points within the preset association gate according to an output result of the association model; Calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, taking the maximum weighted Euclidean distance as the association threshold of the current cycle, and saving the association threshold; Wherein, the determining of the reference scattering point comprises: selecting the center point of the target as the reference scattering point from the data scanned by the vehicle-mounted millimeter-wave radar, or selecting a scattering point with significant features as the reference scattering point; The calculating of the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point includes: taking into account the velocity and acceleration factors of the target under the premise of using a constant acceleration model, and making them uniform in dimension, Set as dimensionless weight value, Set the unit to The weight value of Set the unit to The weight value, and numerically , the improved Euclidean distance formula is as follows: ,in, is the weighted Euclidean distance between the potential homologous scattering point and the reference scattering point, are the geometric coordinates of the reference scattering point in the Cartesian coordinate system, For the preset associated gate The geometric coordinates of potential scattering points in the Cartesian coordinate system, is the velocity of the reference scattering point, is the acceleration of the reference scattering point, For the preset associated gate The velocity of a potential scattering point, For the preset associated gate The acceleration of a potential scattering point, , and The value of is obtained through data experiments in multiple scenarios.

2. The adaptive threshold association method based on time domain convolutional neural network according to claim 1 is characterized in that: The step of obtaining a set of homologous scattering points within the association gate according to the output result of the association model includes: Based on the output result, selecting target potential homologous scattering points that meet preset homology conditions from the multiple potential homologous scattering points; The target potential homologous scattering point is added to the homologous scattering point set in the preset association gate.

3. The adaptive threshold association method based on time domain convolutional neural network according to claim 1 is characterized in that: The calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, and taking the maximum weighted Euclidean distance as the association threshold of the current cycle, includes: sequentially calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point until all homologous scattering points in the homologous scattering point set are traversed to obtain a plurality of weighted Euclidean distances; Comparing the multiple weighted Euclidean distances, and selecting the largest weighted Euclidean distance; The maximum weighted Euclidean distance is used as the association threshold of the current cycle.

4. The adaptive threshold association method based on time-domain convolutional neural network according to claim 1 is characterized in that: Before inputting the time series data into the designed association model based on the time domain convolutional neural network, it also includes: Acquire vehicle-mounted radar scattering point data of multiple targets to obtain a vehicle-mounted radar scattering point data set, and calibrate the scattering points of the same target in the scattering point data set; The calibrated scattering point data set is divided into a training set, a validation set and a test set according to a preset ratio, and the training set is used to train an initial association model based on a time-domain convolutional neural network. After the training is completed, the validation set is used to verify the initial association model based on the time-domain convolutional neural network. After the verification is completed, the test set is used to test the initial association model based on the time-domain convolutional neural network to obtain the designed association model based on the time-domain convolutional neural network.

5. An adaptive threshold association device based on a time-domain convolutional neural network, characterized in that: include: A screening module, used to determine a reference scattering point, and screen out a plurality of potential homologous scattering points within a preset association gate with the reference scattering point as the center; A calculation module, used for calculating the weighted Euclidean distance between each potential homologous scattering point and the reference scattering point, and combining the weighted Euclidean distance with the correlation threshold of the previous adjacent cycles to form time series data; A model output module, used for inputting the time series data into a designed association model based on a time-domain convolutional neural network, and obtaining a set of homologous scattering points within the preset association gate according to an output result of the association model; an association threshold determination module, used to calculate the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point, take the maximum weighted Euclidean distance as the association threshold of the current cycle, and save the association threshold; The screening module is used to select the center point of the target as the reference scattering point from the data scanned by the vehicle-mounted millimeter-wave radar, or to select a scattering point with significant features as the reference scattering point; The calculation module is used to: take into account the speed and acceleration factors of the target under the premise of using a constant acceleration model, and unify the dimensions. Set as dimensionless weight value, Set the unit to The weight value of Set the unit to The weight value, and numerically , the improved Euclidean distance formula is as follows: ,in, is the weighted Euclidean distance between the potential homologous scattering point and the reference scattering point, are the geometric coordinates of the reference scattering point in the Cartesian coordinate system, For the preset associated gate The geometric coordinates of potential scattering points in the Cartesian coordinate system, is the velocity of the reference scattering point, is the acceleration of the reference scattering point, For the preset associated gate The velocity of a potential scattering point, For the preset associated gate The acceleration of a potential scattering point, , and The value of is obtained through data experiments in multiple scenarios.

6. The adaptive threshold association device based on time-domain convolutional neural network according to claim 5 is characterized in that: The model output module is used to: Based on the output result, selecting target potential homologous scattering points that meet preset homology conditions from the multiple potential homologous scattering points; The target potential homologous scattering point is added to the homologous scattering point set in the preset association gate.

7. The adaptive threshold association device based on time-domain convolutional neural network according to claim 5, characterized in that: The association threshold determination module is used to: sequentially calculating the weighted Euclidean distance between each homologous scattering point in the homologous scattering point set and the reference scattering point until all homologous scattering points in the homologous scattering point set are traversed to obtain a plurality of weighted Euclidean distances; Comparing the multiple weighted Euclidean distances, and selecting the largest weighted Euclidean distance; The maximum weighted Euclidean distance is used as the association threshold of the current cycle.

8. The adaptive threshold association device based on time-domain convolutional neural network according to claim 5, characterized in that: Before inputting the time series data into the designed association model based on the time domain convolutional neural network, the model output module is further used to: Acquire vehicle-mounted radar scattering point data of multiple targets to obtain a vehicle-mounted radar scattering point data set, and calibrate the scattering points of the same target in the scattering point data set; The calibrated scattering point data set is divided into a training set, a validation set and a test set according to a preset ratio, and the training set is used to train an initial association model based on a time-domain convolutional neural network. After the training is completed, the validation set is used to verify the initial association model based on the time-domain convolutional neural network. After the verification is completed, the test set is used to test the initial association model based on the time-domain convolutional neural network to obtain the designed association model based on the time-domain convolutional neural network.

9. A vehicle, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the adaptive threshold association method based on a time-domain convolutional neural network as described in any one of claims 1 to 4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The program is executed by a processor to implement the adaptive threshold association method based on a time-domain convolutional neural network as described in any one of claims 1 to 4.

Images