Radar target automatic labeling method and system based on deep learning
Through the deep learning-based automatic radar target labeling method, the traditional manual labeling is solved, and efficient and accurate automatic radar target labeling is achieved, reducing costs and improving label consistency.
Patent Information
- Application Number
- CN202411478360.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-22
- Publication Date
- 2025-02-14
Smart Images
Figure CN119445051A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of radar target labeling, and in particular to a method and system for automatic labeling of radar targets based on deep learning. Background Art
[0002] With the continuous development of science and technology, millimeter wave radar plays an increasingly important role in many fields such as autonomous driving, intelligent security, industrial detection, etc. Millimeter wave radar can obtain the location, speed, category and other information of the target object by transmitting millimeter wave signals and receiving echoes.
[0003] In tasks involving target detection, accurate labeling of target detection frames, such as cars, people, two-wheelers, etc., is crucial for subsequent target recognition and analysis. Currently, there are some semi-automatic labeling solutions for millimeter-wave radar point cloud target detection frames:
[0004] The first is semi-automatic labeling based on geometric features. This method uses the geometric features of the target object for semi-automatic labeling. For example, for vehicle targets, some rules can be set to preliminarily determine the range of the detection frame based on its approximate length, width and height ratio and the distribution characteristics of the reflection intensity. Then, the framed range is manually adjusted and confirmed. However, this method has poor adaptability to targets with irregular shapes or special postures, and is prone to inaccurate labeling.
[0005] The second method is to use a target tracking algorithm to automatically track and label the target. The accuracy of this method is very low because the millimeter-wave radar point cloud is not as dense as the lidar point cloud, resulting in very poor tracking results. In the end, a lot of manual adjustment work is still required to complete semi-automatic labeling.
[0006] In general, the traditional method of manually labeling target detection boxes is not only time-consuming and labor-intensive, but the labeling results are easily affected by human factors, and the consistency and accuracy are difficult to guarantee. As the amount of data continues to increase, the efficiency of manual labeling is increasingly difficult to meet actual needs. At the same time, the cost of manual labeling is high, which is not economically feasible for large-scale labeling tasks.
[0007] Regarding the above-mentioned related technologies, it is believed that the traditional method of manually annotating target detection boxes requires a lot of human resources and is not economically and accurately feasible for large-scale labeling tasks. Summary of the invention
[0008] In order to improve the traditional method of manually labeling target detection frames, which requires a lot of human resources and is not economically and accurately feasible for large-scale labeling tasks, the present application provides a radar target automatic labeling method and system based on deep learning.
[0009] In the first aspect, the present application provides a radar target automatic labeling method based on deep learning, which adopts the following technical solutions: including:
[0010] Synchronously collect millimeter-wave radar point cloud data and camera image data;
[0011] Processing the millimeter-wave radar point cloud data according to a radar data processing algorithm to obtain an initial target detection frame;
[0012] Importing the initial target detection frame into a semi-automatic annotation tool, and annotating the initial target detection frame according to the camera image data to obtain a model training data set;
[0013] Training a preset target automatic labeling model according to the model training data set to obtain the trained target automatic labeling model;
[0014] When the millimeter wave radar data is received, one-click labeling is achieved according to the trained target automatic labeling model.
[0015] Optionally, the model training data set includes: the millimeter wave radar point cloud data and corresponding correct target annotations;
[0016] The step of training a preset target automatic labeling model according to the model training data set to obtain the trained target automatic labeling model includes:
[0017] Initializing weights and bias parameters of the target automatic labeling model;
[0018] Importing the millimeter-wave radar point cloud data into the target automatic labeling model in batches, and outputting predicted target labels;
[0019] Calculating a value of a loss function according to the correct target labeling and the predicted target labeling;
[0020] Adjusting the parameters of the target automatic labeling model according to the gradient of the loss function;
[0021] The target automatic labeling model is repeatedly trained until the target automatic labeling model converges or reaches a preset number of training times, thereby obtaining a trained target automatic labeling model.
[0022] Optionally, importing the millimeter wave radar point cloud data into the target automatic labeling model in batches and outputting predicted target labels includes:
[0023] Importing the millimeter-wave radar point cloud data into the target automatic annotation model in batches;
[0024] Columnarizing the millimeter-wave radar point cloud data to generate a point cloud column;
[0025] superimposing the point cloud columns to generate a pseudo image;
[0026] The pseudo image is input into the backbone network for calculation, and the predicted target annotation is output. The predicted target annotation is: the center position, length, width, height and category of the target detection box. The backbone network includes: 3X3 convolution, 5X5 convolution, maximum / average pooling, Batch Normalization, ReLU activation function, fc fully connected layer, and softmax classification function.
[0027] Optionally, the loss function includes: a classification loss function and an intersection-over-union loss function;
[0028] L all =L class +L IoU ;
[0029]
[0030] Among them, L all is the loss function, L class is the classification loss function, L IoU is the intersection-over-union loss function, N represents the total number of categories, x i Predict the target label for the current category, y i is the probability of the current category after the activation function, is the correct target label for the current category.
[0031] Optionally, the intersection-over-union loss function includes:
[0032]
[0033] L IoU =1-IoU(B,B gt );
[0034] Among them, B, B1, B2 are target detection boxes, B gt is the target detection box label, L IoU is the intersection loss and ratio loss.
[0035] Optionally, when the millimeter wave radar data is received, one-click labeling is implemented according to the trained target automatic labeling model, including:
[0036] When the millimeter-wave radar data is received, obtaining the millimeter-wave radar data required to be marked for each frame;
[0037] Each frame of the millimeter-wave radar data is sequentially input into the trained target automatic labeling model to obtain a target detection frame as a target label, and a labeling file is automatically generated.
[0038] In the second aspect, the present application provides a radar target automatic labeling device based on deep learning, which adopts the following technical solutions, including:
[0039] Data acquisition module, used to synchronously collect millimeter-wave radar point cloud data and camera image data;
[0040] An initial detection module, used to process the millimeter-wave radar point cloud data according to a radar data processing algorithm to obtain an initial target detection frame;
[0041] A data generation module, used for importing the initial target detection frame into a semi-automatic annotation tool, and annotating the initial target detection frame according to the camera image data to obtain a model training data set;
[0042] A model training module is used to train a preset target automatic labeling model according to the model training data set to obtain the trained target automatic labeling model;
[0043] The automatic labeling module is used to implement one-click labeling according to the trained target automatic labeling model when receiving millimeter wave radar data.
[0044] In a third aspect, the present application further provides a control device, the device comprising:
[0045] It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executes the above-mentioned deep learning-based radar target automatic labeling method.
[0046] In a fourth aspect, the present application also provides a computer-readable storage medium storing a computer program that can be loaded by a processor and execute the above-mentioned deep learning-based automatic radar target labeling method.
[0047] In summary, the system in this application uses pre-calibrated and calibrated equipment to synchronously collect millimeter-wave radar point cloud data and camera image data, and generates an initial target detection frame according to a traditional radar data processing algorithm. Then the system adjusts the position, length, width, height, and category of all target detection frames according to a semi-automatic annotation tool, and performs manual adjustment to complete the data annotation. Then the system inputs the model training data set into the automatic annotation model, and continuously adjusts the model parameters through feedback from the loss function until the model converges or reaches a preset number of training times, and obtains the trained automatic annotation model. When the system receives the millimeter-wave radar data that needs to be annotated, it generates an annotation file for each frame of millimeter-wave radar data, thereby achieving the effect of automatically annotating the millimeter-wave radar point cloud data, thereby reducing the manual intervention in the model annotation process to increase the accuracy and consistency of the model annotation, while saving the use of human resources and reducing the cost of automatic target annotation. In addition, when making the model training data set, the traditional radar data processing algorithm is combined to annotate the data set used for training, thereby improving the annotation capabilities of obscured targets, distant targets, and ghost targets. It can also adapt to different types of radar point cloud data through continuous training and be flexibly applied to various scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a flowchart of a radar target automatic labeling method based on deep learning.
[0049] Figure 2 It is the structural block diagram of the target automatic labeling model.
[0050] Figure 3 It is a structural block diagram of a radar target automatic labeling device based on deep learning.
[0051] Explanation of the accompanying drawings: 310, data acquisition module; 320, initial detection module; 330, data generation module; 340, model training module; 350, automatic labeling module. DETAILED DESCRIPTION
[0052] The following combination Figure 1 - Figure 3 This application is described in further detail.
[0053] Point cloud data refers to a set of vectors in a three-dimensional coordinate system. The scanned data is recorded in the form of points, each point contains three-dimensional coordinates, and some may contain color information (RGB) or reflection intensity information. The acquisition of reflection intensity information is the echo intensity collected by the receiving device. This intensity information is related to the surface material, roughness, incident angle direction of the target, and the emission energy and laser wavelength of the instrument. The millimeter wave radar point cloud data in this application includes: point distance (R), speed (V), horizontal angle (Ag), pitch angle (Ah), signal strength (M). The camera image data is an RGB image with a resolution of 1280*720.
[0054] At present, there are some semi-automatic annotation solutions for millimeter wave radar point cloud target detection boxes:
[0055] The first is semi-automatic labeling based on geometric features. This method uses the geometric features of the target object for semi-automatic labeling. For example, for vehicle targets, some rules can be set to preliminarily determine the range of the detection frame based on its approximate length, width and height ratio and the distribution characteristics of the reflection intensity. Then, the framed range is manually adjusted and confirmed. However, this method has poor adaptability to targets with irregular shapes or special postures, and is prone to inaccurate labeling.
[0056] The second method is to use a target tracking algorithm to automatically track and label the target. The accuracy of this method is very low because the millimeter-wave radar point cloud is not as dense as the lidar point cloud, resulting in very poor tracking results. In the end, a lot of manual adjustment work is still required to complete semi-automatic labeling.
[0057] The third is semi-automatic annotation that combines multi-sensor information, such as combining information from sensors such as cameras and lidars to annotate target detection frames. Cameras can provide the appearance features of target objects, and lidars can provide more accurate three-dimensional spatial information. By fusing this information, the accuracy of annotation can be improved. However, this method requires processing data from multiple sensors, the system is highly complex, and data fusion between different sensors is also difficult.
[0058] It is not economically feasible for large-scale labeling tasks.
[0059] This application aims to solve the problems of low efficiency of manual labeling of millimeter-wave radar, low accuracy of semi-automatic labeling, and difficulty in labeling special targets (objects out of sight at a distance, obscured objects, and ghost targets). A deep learning model is introduced to infer the input radar point cloud data, obtain the target detection box, and realize automatic labeling of the radar.
[0060] Based on this, the embodiment of the present application discloses a method for automatic labeling of radar targets based on deep learning. The execution subject is a control system. The input is the point cloud data of the 4D millimeter wave radar, and the output is the target detection box (including the center position, length, width, height, and category). Through the deep learning algorithm, automatic labeling can be achieved with one click, which greatly improves the efficiency and accuracy of radar target labeling, especially the labeling rate of obscured objects and distant objects.
[0061] Reference Figure 1 , the embodiment of the present application at least includes step S10 to step S50.
[0062] S10, synchronously collect millimeter wave radar point cloud data and camera image data.
[0063] S20, processing the millimeter wave radar point cloud data according to a radar data processing algorithm to obtain an initial target detection frame.
[0064] S30, importing the initial target detection frame into a semi-automatic labeling tool, and labeling the initial target detection frame according to the camera image data to obtain a model training data set.
[0065] S40, training a preset target automatic labeling model according to a model training data set to obtain a trained target automatic labeling model.
[0066] S50, when the millimeter wave radar data is received, one-click labeling is implemented according to the trained target automatic labeling model.
[0067] Specifically, the system uses pre-calibrated and calibrated equipment to synchronously collect millimeter-wave radar point cloud data and camera image data, and generates an initial target detection frame according to the traditional radar data processing algorithm. Then the system adjusts the position, length, width, height, and category of all target detection frames according to the semi-automatic annotation tool, and performs manual adjustment to complete the data annotation. Then the system inputs the model training data set into the automatic annotation model, and continuously adjusts the model parameters through the feedback of the loss function until the model converges or reaches the preset number of training times, and obtains the trained automatic annotation model. When the system receives the millimeter-wave radar data that needs to be annotated, it generates an annotation file for each frame of millimeter-wave radar data, so as to achieve the effect of automatic annotation of millimeter-wave radar point cloud data, thereby reducing the manual intervention in the model annotation process to increase the accuracy and consistency of model annotation, while saving the use of human resources and reducing the cost of automatic target annotation. In addition, when making the model training data set, the traditional radar data processing algorithm is combined to annotate the data set used for training, which improves the annotation ability of obscured targets, distant targets, and ghost targets. It can also adapt to different types of radar point cloud data through continuous training and be flexibly applied to various scenarios.
[0068] Among them, the traditional radar data processing algorithm is as follows: first, the DBSCAN algorithm is used to cluster the millimeter-wave radar point cloud; then the extended Kalman filter target tracking algorithm is used to identify and track the clustered targets to obtain the target detection frame; finally, the target type is estimated based on the target's speed, signal strength (M) and target size.
[0069] In some embodiments, step S40 specifically includes the following steps: initializing the weights and bias parameters of the target automatic labeling model; importing the millimeter-wave radar point cloud data into the target automatic labeling model in batches, and outputting the predicted target labeling; calculating the value of the loss function based on the correct target labeling and the predicted target labeling; adjusting the parameters of the target automatic labeling model according to the gradient of the loss function; repeatedly training the target automatic labeling model until the target automatic labeling model converges or reaches a preset number of training times, thereby obtaining a trained target automatic labeling model.
[0070] Among them, the millimeter wave radar point cloud data and the corresponding correct target annotation. The correct target annotation refers to the adjusted target annotation data.
[0071] Reference Figure 2 , the model input is a point cloud sequence, the point cloud is columnarized, and then the point cloud columns are stacked together to form a pseudo image; the backbone network is mainly composed of 3X3 convolution, 5X5 convolution, maximum / average pooling, Batch Normalization, ReLU activation function, fc fully connected layer, and softmax classification function; the output is the center position, length, width, height, and category of the target detection box. The output is the center position, length, width, height, and category of the target detection box. The model using deep learning algorithms can automatically extract complex features from radar point cloud data, thereby achieving high-precision target detection and labeling, especially for pedestrians and small targets. In addition, it can also adapt to different types of radar point cloud data through continuous training and be flexibly applied to various scenarios.
[0072] Among them, the backbone network can use neural network structures such as ResNet, VGGNet, DenseNet, MobileNet, CSPNet, ResNet, and Transformer.
[0073] Optionally, import the millimeter-wave radar point cloud data into the target automatic annotation model in batches and output the predicted target annotations:
[0074] Import millimeter-wave radar point cloud data into the target automatic annotation model in batches;
[0075] Columnarize the millimeter-wave radar point cloud data to generate a point cloud column;
[0076] Overlay the point cloud columns to generate a pseudo image;
[0077] The pseudo image is input into the backbone network for calculation, and the predicted target annotation is output. The predicted target annotation is: the center position, length, width, height and category of the target detection box. The backbone network includes: 3X3 convolution, 5X5 convolution, maximum / average pooling, Batch Normalization, ReLU activation function, fc fully connected layer, and softmax classification function.
[0078] Furthermore, the loss functions include: classification loss function and intersection-over-union loss function;
[0079] L all =L class +L IoU ;
[0080]
[0081] Among them, L all is the loss function, L class is the classification loss function, L IoU is the intersection-over-union loss function, N represents the total number of categories, x i Predict the target label for the current category, y i is the probability of the current category after the activation function, is the correct target label for the current category.
[0082] The intersection-over-union loss function includes:
[0083]
[0084] L IoU =1-IoU(B,B gt );
[0085] Among them, B, B1, B2 are target detection boxes, B gt is the target detection box label, L IoU is the intersection loss and ratio loss.
[0086] In some embodiments, step S50 specifically includes the following steps: when receiving millimeter-wave radar data, obtaining the millimeter-wave radar data that needs to be labeled for each frame; inputting each frame of millimeter-wave radar data into the trained target automatic labeling model in turn, obtaining the target detection box as the target label, and automatically generating a labeling file.
[0087] The implementation principle of a radar target automatic labeling method based on deep learning in an embodiment of the present application is as follows: the system uses pre-calibrated and calibrated equipment to synchronously collect millimeter-wave radar point cloud data and camera image data, and generates an initial target detection frame according to a traditional radar data processing algorithm. Then the system adjusts the position, length, width, height, and category of all target detection frames according to a semi-automatic labeling tool, and performs manual adjustment to complete the data labeling. Then the system inputs the model training data set into the automatic labeling model, and continuously adjusts the model parameters through feedback from the loss function until the model converges or reaches a preset number of training times, thereby obtaining a trained automatic labeling model. When the system When the system receives millimeter-wave radar data that needs to be labeled, it generates a labeling file for each frame of millimeter-wave radar data, thereby achieving the effect of automatically labeling millimeter-wave radar point cloud data, thereby reducing manual intervention in the model labeling process to increase the accuracy and consistency of model labeling, while saving the use of human resources and reducing the cost of automatic target labeling. In addition, when making a model training data set, it combines traditional radar data processing algorithms to label the data set used for training, thereby improving the labeling capabilities of obscured targets, distant targets, and peeking targets. It can also adapt to different types of radar point cloud data through continuous training and be flexibly applied to various scenarios.
[0088] Figure 1 FIG. 1 is a flow chart of a method for automatically labeling radar targets based on deep learning in one embodiment. It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but the steps are not necessarily executed in the order indicated by the arrows; unless otherwise specified in this document, there is no strict order restriction for the execution of the steps, and the steps may be executed in other orders; and Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0089] Based on the same technical concept, Figure 3 The embodiment of the present application also provides a radar target automatic labeling device based on deep learning, which adopts the following technical solution. The device includes:
[0090] The data acquisition module 310 is used to synchronously acquire millimeter-wave radar point cloud data and camera image data;
[0091] An initial detection module 320 is used to process the millimeter wave radar point cloud data according to a radar data processing algorithm to obtain an initial target detection frame;
[0092] A data generation module 330 is used to import the initial target detection frame into the semi-automatic annotation tool, and annotate the initial target detection frame according to the camera image data to obtain a model training data set;
[0093] A model training module 340 is used to train a preset target automatic labeling model according to a model training data set to obtain a trained target automatic labeling model;
[0094] The automatic labeling module 350 is used to implement one-click labeling according to the trained target automatic labeling model when receiving millimeter wave radar data.
[0095] In some embodiments, the model training module 340 is specifically used to initialize the weights and bias parameters of the target automatic labeling model;
[0096] Import millimeter-wave radar point cloud data into the target automatic labeling model in batches, and output predicted target labels;
[0097] Calculate the value of the loss function based on the correct target labeling and the predicted target labeling;
[0098] Adjust the parameters of the target automatic labeling model according to the gradient of the loss function;
[0099] The target automatic labeling model is repeatedly trained until the target automatic labeling model converges or reaches a preset number of training times, thereby obtaining a trained target automatic labeling model.
[0100] In some embodiments, the model training module 340 is further used to import the millimeter wave radar point cloud data into the target automatic annotation model in batches;
[0101] Columnarize the millimeter-wave radar point cloud data to generate a point cloud column;
[0102] Overlay the point cloud columns to generate a pseudo image;
[0103] The pseudo image is input into the backbone network for calculation, and the predicted target annotation is output. The predicted target annotation is: the center position, length, width, height and category of the target detection box. The backbone network includes: 3X3 convolution, 5X5 convolution, maximum / average pooling, Batch Normalization, ReLU activation function, fc fully connected layer, and softmax classification function.
[0104] In some embodiments, the loss function includes: a classification loss function and an intersection-over-union loss function;
[0105] L all =L class +L IoU ;
[0106]
[0107] Among them, L all is the loss function, L class is the classification loss function, L IoU is the intersection-over-union loss function, N represents the total number of categories, x i Predict the target label for the current category, y i is the probability of the current category after the activation function, is the correct target label for the current category.
[0108] In some embodiments, the intersection-over-union loss function includes:
[0109]
[0110] L IoU =1-IoU(B,B gt );
[0111] Among them, B, B1, B2 are target detection boxes, B gt is the target detection box label, L IoU is the intersection loss and ratio loss.
[0112] In some embodiments, the automatic labeling module 350 is specifically used to obtain the millimeter wave radar data that needs to be labeled for each frame when the millimeter wave radar data is received;
[0113] Each frame of millimeter wave radar data is sequentially input into the trained target automatic labeling model to obtain a target detection frame as a target label, and a labeling file is automatically generated. The embodiment of the present application also discloses a control device.
[0114] Specifically, the control device includes a memory and a processor, and the memory stores a computer program that can be loaded by the processor and execute the above-mentioned deep learning-based radar target automatic labeling method.
[0115] The embodiment of the present application also discloses a computer-readable storage medium.
[0116] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executes the above-mentioned deep learning-based radar target automatic labeling method. The computer-readable storage medium includes, for example: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program codes.
[0117] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.
Claims
1. A radar target automatic labeling method based on deep learning, characterized in that: include: Synchronously collect millimeter-wave radar point cloud data and camera image data; Processing the millimeter-wave radar point cloud data according to a radar data processing algorithm to obtain an initial target detection frame; Importing the initial target detection frame into a semi-automatic annotation tool, and annotating the initial target detection frame according to the camera image data to obtain a model training data set; Training a preset target automatic labeling model according to the model training data set to obtain the trained target automatic labeling model; When the millimeter wave radar data is received, one-click labeling is achieved according to the trained target automatic labeling model.
2. The method for automatic labeling of radar targets based on deep learning according to claim 1, characterized in that: The model training data set includes: the millimeter wave radar point cloud data and corresponding correct target annotations; The step of training a preset target automatic labeling model according to the model training data set to obtain the trained target automatic labeling model includes: Initializing weights and bias parameters of the target automatic labeling model; Importing the millimeter-wave radar point cloud data into the target automatic labeling model in batches, and outputting predicted target labels; Calculating a value of a loss function according to the correct target labeling and the predicted target labeling; Adjusting the parameters of the target automatic labeling model according to the gradient of the loss function; The target automatic labeling model is repeatedly trained until the target automatic labeling model converges or reaches a preset number of training times, thereby obtaining a trained target automatic labeling model.
3. The method for automatic labeling of radar targets based on deep learning according to claim 2, characterized in that: The step of importing the millimeter wave radar point cloud data into the target automatic labeling model in batches and outputting predicted target labels includes: Importing the millimeter-wave radar point cloud data into the target automatic annotation model in batches; Columnarizing the millimeter-wave radar point cloud data to generate a point cloud column; superimposing the point cloud columns to generate a pseudo image; The pseudo image is input into the backbone network for calculation, and the predicted target annotation is output. The predicted target annotation is: the center position, length, width, height and category of the target detection box. The backbone network includes: 3X3 convolution, 5X5 convolution, maximum / average pooling, Batch Normalization, ReLU activation function, fc fully connected layer, and softmax classification function.
4. The method for automatic labeling of radar targets based on deep learning according to claim 3 is characterized in that: The loss function includes: a classification loss function and an intersection-over-union loss function; L all =L class +L IoU ; Among them, L all is the loss function, L class is the classification loss function, L IoU is the intersection-over-union loss function, N represents the total number of categories, x i Predict the target label for the current category, y i is the probability of the current category after the activation function, is the correct target label for the current category.
5. The method for automatic labeling of radar targets based on deep learning according to claim 4, characterized in that: The intersection-over-union loss function includes: L IoU =1-IoU(B,B gt ); Among them, B, B1, B2 are target detection boxes, B gt is the target detection box label, L IoU is the intersection loss and ratio loss.
6. The method for automatic labeling of radar targets based on deep learning according to claim 1, characterized in that: When the millimeter wave radar data is received, one-key labeling is implemented according to the trained target automatic labeling model, including: When the millimeter-wave radar data is received, obtaining the millimeter-wave radar data required to be marked for each frame; Each frame of the millimeter-wave radar data is sequentially input into the trained target automatic labeling model to obtain a target detection frame as a target label, and a labeling file is automatically generated.
7. A radar target automatic labeling device based on deep learning, characterized in that: The device comprises: Data acquisition module, used to synchronously collect millimeter-wave radar point cloud data and camera image data; An initial detection module, used to process the millimeter-wave radar point cloud data according to a radar data processing algorithm to obtain an initial target detection frame; A data generation module, used for importing the initial target detection frame into a semi-automatic annotation tool, and annotating the initial target detection frame according to the camera image data to obtain a model training data set; A model training module is used to train a preset target automatic labeling model according to the model training data set to obtain the trained target automatic labeling model; The automatic labeling module is used to implement one-click labeling according to the trained target automatic labeling model when receiving millimeter wave radar data.
8. A control device, characterized in that: The device comprises: The method comprises a memory and a processor, wherein the memory stores a computer program which can be loaded by the processor and execute the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that: A computer program is stored which can be loaded by a processor and execute the method according to any one of claims 1 to 6.