Millimeter wave radar external parameter rapid dynamic automatic calibration method

Abstract

The invention relates to the field of radar external parameter calibration, and discloses a millimeter wave radar external parameter rapid dynamic automatic calibration method, which comprises the following steps that S1, a to-be-tested vehicle runs in a test range of a millimeter wave radar, the millimeter wave radar collects original target data of each target in a surrounding environment, and real-time dynamic positioning equipment synchronously collects vehicle track data; s2, performing target identification and track splicing on the original target data to obtain corrected radar sensing track data; s3, based on the dynamic time error, performing time synchronization on the radar sensing trajectory data and the vehicle trajectory data, and matching the radar sensing trajectory data with the vehicle trajectory data through a timestamp; s4, generating a homography matrix by using the matching point pairs, and optimizing the homography matrix by using a deep reinforcement learning algorithm to reduce calibration errors; and S5, performing external parameter calibration on the millimeter wave radar by using the optimized homography matrix. According to the invention, the error tolerance can be automatically adjusted, and the calibration precision is improved.

Description

Technical Field

[0001] The present invention relates to the field of radar external parameter calibration, and particularly to a method for rapid dynamic automatic calibration of millimeter-wave radar external parameters. Background Art

[0002] With the continuous progress of intelligent network connection technology, millimeter-wave radar is increasingly widely used in the field of road-end perception. Especially in the deployment of roadside units for intelligent network connection, millimeter-wave radar, with its excellent capabilities, can accurately detect and track various targets on the road, including vehicles and pedestrians, providing high-precision and real-time perception data support for the road-end system. Compared with vision sensors, millimeter-wave radar exhibits strong anti-interference characteristics and the advantage of being unaffected by adverse weather conditions. Therefore, it plays an indispensable perception role in the intelligent network connection road-end scenario.

[0003] However, the measurement accuracy and data reliability of millimeter-wave radar largely depend on the accuracy of its external parameter calibration, that is, ensuring the precise alignment between the radar coordinate system and the UTM (Universal Transverse Mercator Grid System) global coordinate system. Currently, most traditional calibration methods rely on cumbersome manual operations and complex static environment debugging processes, which not only take a long time but are also easily affected by the skill level of operators and changes in environmental conditions, resulting in unstable calibration results. For intelligent network connection road-end applications, this inefficient and unstable calibration method clearly cannot meet the system's urgent needs for high precision and high automation. Therefore, there is an urgent need for a technology that can achieve automatic and high-precision external parameter calibration of millimeter-wave radar in complex and variable dynamic environments. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for rapid dynamic automatic calibration of millimeter-wave radar external parameters to solve the problem that most traditional calibration methods rely on cumbersome manual operations and complex static environment debugging processes.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A method for rapid dynamic automatic calibration of millimeter-wave radar external parameters includes the following steps:

[0007] S1: A vehicle to be tested travels within the test range of the millimeter-wave radar, and the millimeter-wave radar collects the original target data of each target in the surrounding environment, and a real-time dynamic positioning device synchronously collects vehicle trajectory data;

[0008] S2: Perform target recognition and trajectory splicing on the original target data to obtain corrected radar perception trajectory data;

[0009] S3: Based on the dynamic time error, synchronize the time of the radar perception trajectory data and the vehicle trajectory data, and match the radar perception trajectory data and the vehicle trajectory data through timestamps;

[0010] S4: Generate a homography matrix using the matched point pairs, and optimize the homography matrix using a deep reinforcement learning algorithm to reduce the calibration error;

[0011] S5: Calibrate the extrinsic parameters of the millimeter-wave radar using the optimized homography matrix.

[0012] According to the above technical means, through the collaborative work of a Real Time Kinematic (RTK) device and a millimeter-wave radar, the present invention can automatically collect and process data, thereby avoiding manual adjustment and repeated experiments, such as manually setting calibration objects and adjusting the radar position, significantly reducing human errors, and improving the calibration efficiency. At the same time, the present invention uses the vehicle to be measured to automatically collect calibration points, replacing the traditional manual one-by-one collection method. During the vehicle driving process, the system automatically records a large number of calibration points, and the collection efficiency is increased several times. This not only greatly saves the time cost, but also can significantly increase the number of calibration points, ensuring a wider data coverage, and thus improving the accuracy and representativeness of the calibration results.

[0013] Through precise time synchronization technology and deep learning algorithms, the present invention optimizes the recognition of the ego-vehicle trajectory, significantly increasing the number of calibration points from several in the traditional method to thousands, greatly enhancing the representativeness and diversity of the data. Combined with the dynamic adaptive matching algorithm, the present invention can automatically adjust the error tolerance according to the actual environmental conditions, thereby effectively eliminating the influence of environmental changes on the calibration accuracy, ensuring millimeter-level accuracy of the final calibration results, and meeting the high-precision requirements in autonomous driving and intelligent networked applications.

[0014] Further, the original target data includes a target serial number, target coordinates, and a radar acquisition timestamp;

[0015] R radar (t 1 ) = [ID(t 1 ), x target (t 1 ), y target (t 1 )],

[0016] where R radar (t 1 ) is the original target data, ID(t 1 ) is the target serial number, x target (t 1 ) is the target abscissa, ytarget (t 1 ) The target vertical coordinate, where t 1 is the radar acquisition timestamp;

[0017] The vehicle trajectory data includes vehicle coordinates and positioning timestamps;

[0018] P RTK (t 2 ) = [x RTK (t 2 ), y RTK (t 2 )],

[0019] where P RTK (t 2 ) is the vehicle trajectory data, x RTK (t 2 ) is the vehicle horizontal coordinate, y RTK (t 2 ) is the vehicle vertical coordinate, and t 2 is the positioning timestamp.

[0020] According to the above technical means, information such as the target serial number, target coordinates, and radar acquisition timestamp together constitute a complete radar perception dataset. The target serial number ensures the uniqueness of each target, the target coordinates provide the precise position information of the target, and the radar acquisition timestamp guarantees the timeliness of the data. This complete dataset provides a solid foundation for subsequent target recognition, trajectory stitching, and time synchronization, thereby improving the accuracy and reliability of the calibration process.

[0021] Since both the vehicle trajectory data and the original target data of the radar are collected in real time and can be matched and optimized during vehicle driving, this method realizes real-time dynamic calibration. This means that in practical applications, the external parameters of the millimeter-wave radar can be continuously updated as the vehicle drives to adapt to different road and traffic environments. This real-time and dynamic advantage improves the adaptability and perception ability of the radar.

[0022] Furthermore, S2 specifically includes the following steps:

[0023] S21: Perform target recognition on the data collected by the millimeter-wave radar through training a deep neural network model, and reconstruct the trajectory according to the time series to generate a mapping function;

[0024] S22: Process the original target data through mapping to obtain the calibrated radar perception trajectory data.

[0025] According to the above technical means, in the process of target tracking by traditional millimeter-wave radar, due to environmental factors, radar performance limitations, or limitations of the algorithm itself, the problem of target ID jumping often occurs. The deep neural network model has powerful feature extraction and classification capabilities and can learn and identify targets in various complex scenarios, such as vehicles, pedestrians, etc. In step S21, by training the deep neural network model to perform target recognition on the data collected by the millimeter-wave radar, the accuracy and robustness of target recognition are significantly improved, so as to more accurately identify and distinguish different targets, and the stability of the target ID can be maintained even in complex scenarios.

[0026] Also in step S21, aiming at the problem of discontinuous trajectories caused by target ID jumping of the millimeter-wave radar, by analyzing and comparing radar data at different time points, the present invention can identify and connect the trajectory segments of the same target, thereby generating a continuous trajectory of the target in the time series, which is of great significance for subsequent correction processing and trajectory tracking in practical applications.

[0027] Further, the mapping function is f:

[0028] T car (t 1 ) = f(R radar (t 1 ))

[0029] where T car (t 1 ) is the corrected radar perception trajectory data, R radar (t 1 ) is the original target data, and t 1 is the radar acquisition timestamp;

[0030] T car (t 1 ) = [x car (t 1 ), y car (t 1 )]

[0031] where x car (t 1 ) is the abscissa of the corrected radar perception target, and y car (t 1 ) is the ordinate of the corrected radar perception target.

[0032] Further, S3 specifically includes the following steps:

[0033] S31: For each point in the radar perception trajectory data, find at least one vehicle trajectory data point near the radar acquisition timestamp t 1 nearby;

[0034] S32: Determine the dynamic error tolerance for the time error between the radar acquisition timestamp t 1 of this point and the positioning timestamp t 2 of the vehicle trajectory data point;

[0035] S33: If the time error is within the dynamic error tolerance range, it is considered that the vehicle trajectory data point matches this point;

[0036] If the time error is not within the dynamic error tolerance range, adjust the dynamic error tolerance or re - perform the time synchronization and matching process.

[0037] According to the above technical means, since the traditional static method uses a fixed time error threshold for matching, but in actual applications, the speed and environment changes of the vehicle will cause the error to change dynamically. Therefore, the present invention introduces the dynamic error tolerance determination. The dynamic error tolerance determination mechanism can flexibly adjust the error tolerance according to the actual situation, thereby improving the accuracy and robustness of the matching, and then flexibly handling the dynamic time error between the radar data and the RTK data caused by the vehicle speed and environment changes.

[0038] Further, the time error is represented by the following formula:

[0039] Δt = |t 1 - t 2 |,

[0040] where Δt is the time error, t 1 is the radar acquisition timestamp, and t 2 is the positioning timestamp;

[0041] The dynamic error tolerance is represented by the following formula:

[0042]

[0043] where Δt max is the dynamic error tolerance, k is a constant, and v is the vehicle speed.

[0044] According to the above technical means, the formula of the dynamic error tolerance shows that as the vehicle speed changes, the tolerance of the time error will change accordingly, enabling the present invention to adapt to the data synchronization requirements under different vehicle speeds and improving the robustness of data matching. Even in the case of large vehicle speed changes or complex environments, the system can maintain stable matching performance and reduce the cases of false matching and missed matching.

[0045] Further, the specific steps of S4 are as follows:

[0046] S41: Generate a homography matrix H based on the geometric transformation relationship using the N pairs of matching points obtained in step S3, and solve for a preliminarily optimized homography matrix H by minimizing the matching error. 1 2 ;

[0047] S42: Optimize the action space based on a reward mechanism and adjust the matrix elements of the homography matrix H to reduce the calibration error of the homography matrix H and obtain an optimal homography matrix H. 2 2 optimal

[0048] According to the above technical means, initially generating the homography matrix H provides an accurate initial estimate for subsequent optimization, reducing the complexity and computational amount of the optimization process. Solving for the preliminarily optimized homography matrix H by minimizing the matching error can further reduce the deviation caused by factors such as noise and sensor errors, laying a foundation for subsequent optimization steps. The present invention introduces a Deep Deterministic Policy Gradient (DDPG) algorithm. By optimizing the action space based on a reward mechanism, the matrix elements of the homography matrix can be dynamically adjusted. Through continuous iteration and optimization, the calibration error can be gradually reduced, continuously improving the calibration accuracy of the homography matrix to flexibly handle various noises and sensor errors and achieve the optimal calibration accuracy of the homography matrix. 1 2

[0049] Further, the homography matrix is expressed as:

[0050]

[0051] where H 1 is the preliminary homography matrix, x RTK (t i ) is the abscissa of the vehicle trajectory data point in the i-th pair of matching points, y RTK (t i ) is the ordinate of the vehicle trajectory data point in the i-th pair of matching points, x car (t i ) is the abscissa of the radar perception trajectory data point in the i-th pair of matching points, y car (t i ) is the ordinate of the radar perception trajectory data point in the i-th pair of matching points, and t i is the timestamp after matching;

[0052] The preliminarily optimized homography matrix is expressed as:

[0053]

[0054] Further, the specific steps of S42 are as follows: Design the state space as the homography matrix H 2 The calibration error, design the reward function as the negative value of the calibration error, and design the action in the action space as the matrix adjustment amount ΔH; Update the homography matrix H through the matrix adjustment amount 2 , generate a new homography matrix H new , and calculate the calibration error of the updated homography matrix H new Use the reward mechanism of the reward function to continuously optimize the action, reduce the calibration error until the optimal homography matrix H is obtained optimal .

[0055] According to the above technical means, by designing the state space as the calibration error of the homography matrix, this method can directly optimize the error, ensure that each iteration is to reduce the error, and make the homography matrix closer to the true value. The reward function is designed as the negative value of the calibration error, which means that when the calibration error decreases, the system will obtain a greater reward. This reward mechanism encourages the algorithm to continuously find the adjustment amount that can reduce the calibration error, thus accelerating the optimization process and improving the efficiency. The action in the action space is designed as the matrix adjustment amount, enabling this method to flexibly adjust each element of the homography matrix to cope with various complex calibration scenarios and ensure obtaining the optimal homography matrix under various conditions. By continuously updating the homography matrix and calculating the calibration error after the update, the present invention can continuously adapt to environmental changes and noise interference, enhance the robustness of the system, and enable it to maintain stable calibration performance under various conditions.

[0056] Further, the reward function is expressed as:

[0057]

[0058] where, R(H) is the reward function, x RTK (t i ) is the abscissa of the vehicle trajectory data point in the i-th pair of matching points, y RTK (t i ) is the ordinate of the vehicle trajectory data point in the i-th pair of matching points, x car (t i ) is the abscissa of the radar perception trajectory data point in the i-th pair of matching points, y car (t i ) is the ordinate of the radar perception trajectory data point in the i-th pair of matching points, t i is the timestamp after matching, H new is the new homography matrix after action adjustment;

[0059] H new = H 2 + ΔH,

[0060] Among them, H 2 is the homography matrix optimized initially by solving the minimum matching error, and ΔH is the matrix adjustment amount.

[0061] Beneficial effects achieved by the present invention:

[0062] Through the collaborative work of a real-time kinematic (RTK) device and a millimeter-wave radar, the present invention can automatically collect and process data, thereby avoiding manual adjustment and repeated experiments, such as manually setting calibration objects and adjusting the radar position, significantly reducing human errors and improving the calibration efficiency. At the same time, the present invention uses the vehicle to be measured to automatically collect calibration points, replacing the traditional manual one-by-one collection method. During the vehicle's driving process, the system automatically records a large number of calibration points, and the collection efficiency is increased by several times. This not only greatly saves the time cost but also significantly increases the number of calibration points, ensuring a wider data coverage, and thus improving the accuracy and representativeness of the calibration results.

[0063] The present invention optimizes the recognition of the ego-vehicle trajectory through precise time synchronization technology and deep learning algorithms, significantly increasing the number of calibration points from several in the traditional method to thousands, greatly enhancing the representativeness and diversity of the data. Combined with the dynamic adaptive matching algorithm, the present invention can automatically adjust the error tolerance according to the actual environmental conditions, thereby effectively eliminating the influence of environmental changes on the calibration accuracy and ensuring millimeter-level accuracy of the final calibration results, meeting the high-precision requirements in autonomous driving and intelligent networked applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 is a schematic diagram of the overall process of the present invention;

[0065] Figure 2 is a schematic diagram of the process of step S2 of the present invention;

[0066] Figure 3 is a schematic diagram of the process of step S3 of the present invention;

[0067] Figure 4 is a schematic diagram of the process of step S4 of the present invention.

[0068] The drawings are only for illustrative purposes and should not be construed as a limitation of this patent; for better illustration of this embodiment, some components in the drawings are omitted, enlarged or reduced, and do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted; the same or similar reference numerals correspond to the same or similar components; the terms describing the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation of this patent. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0069] It should be noted that, without conflict, the embodiments in the present application and the technical features in the embodiments can be combined with each other. The detailed descriptions in the specific embodiments should be understood as the explanatory illustrations of the purpose of the present application and should not be regarded as improper restrictions on the present application.

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will further describe the specific technical solutions of the present application in detail in combination with the accompanying drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application but are not used to limit the scope of the present application.

[0071] In the embodiments of the present application, the terms "include", "comprise", or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or further includes elements inherent to such a process, method, article, or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, article, or device including that element.

[0072] The following will introduce and describe the technical solutions of this embodiment in detail in combination with specific drawings.

[0073] As Figure 1 shown, this embodiment proposes a method for rapid dynamic automatic calibration of the extrinsic parameters of a millimeter-wave radar, including the following steps:

[0074] S1: The vehicle to be measured travels within the test range of the millimeter-wave radar, and the millimeter-wave radar collects the original target data of each target in the surrounding environment, and the real-time dynamic positioning device synchronously collects the vehicle trajectory data;

[0075] S2: Perform target recognition and trajectory stitching on the original target data to obtain the corrected radar perception trajectory data;

[0076] S3: Based on the dynamic time error, synchronize the time of the radar perception trajectory data and the vehicle trajectory data, and match the radar perception trajectory data and the vehicle trajectory data through timestamps;

[0077] S4: Use the matching point pairs to generate a homography matrix, and adopt a deep reinforcement learning algorithm to optimize the homography matrix to reduce the calibration error;

[0078] S5: Use the optimized homography matrix to calibrate the extrinsic parameters of the millimeter-wave radar.

[0079] In this embodiment, the extrinsic calibration of the millimeter-wave radar is specifically as follows: taking the optimized homography matrix as the input, and using the principle of matrix transformation to convert the target data collected by the millimeter-wave radar from the radar coordinate system to the global coordinate system (such as the UTM coordinate system).

[0080] In this embodiment, the specific principle of the fast dynamic automatic extrinsic calibration method is as follows: by means of precise time synchronization technology, the millimeter-wave radar data and RTK data are correlated. RTK provides high-precision vehicle position information, and the millimeter-wave radar provides target detection and tracking data. The two are matched through timestamps to ensure that the data at the same moment can be accurately corresponding.

[0081] Due to the limited tracking ability of the millimeter-wave radar in complex environments, ID jumps and target losses may occur. Therefore, in this embodiment, trajectory recognition and stitching are performed on the millimeter-wave radar data. The interference caused by target ID jumps is eliminated, and the trajectories are optimized and stitched to ensure that the final ego-vehicle trajectory is more complete and stable.

[0082] On the basis of time synchronization and trajectory recognition, this embodiment also introduces a dynamic adaptive matching algorithm. According to different environmental factors and vehicle speeds, the threshold of time error (i.e., dynamic time error) is dynamically adjusted to optimize the matching process between the millimeter-wave radar data and RTK data, enabling the system to flexibly adapt to sensor errors and environmental changes at different speeds, improving the matching accuracy, and avoiding the matching errors caused by environmental changes in traditional methods.

[0083] Through the optimized matching point pairs, this embodiment uses computer vision and geometric methods to generate a homography matrix, and further performs coordinate fine-tuning through deep reinforcement learning technology to eliminate errors, so as to achieve fast, automatic, and high-precision extrinsic calibration of the millimeter-wave radar, overcoming the problems of long time and low accuracy in traditional methods.

[0084] Through the collaborative work of the real-time kinematic (RTK) device and the millimeter-wave radar in this embodiment, data can be automatically collected and processed, thus avoiding manual adjustment and repeated experiments, such as manually setting calibration objects and adjusting the radar position, significantly reducing human errors, and improving the calibration efficiency. At the same time, this embodiment uses the vehicle to be measured to automatically collect calibration points, replacing the traditional manual one-by-one collection method. During the vehicle driving process, the system automatically records a large number of calibration points, and the collection efficiency is increased several times. This not only greatly saves the time cost, but also can significantly increase the number of calibration points, ensure a wider data coverage, and thus improve the accuracy and representativeness of the calibration results.

[0085] In this embodiment, the recognition of the ego-vehicle trajectory is optimized through precise time synchronization technology and deep learning algorithms, significantly increasing the number of calibration points from a few in traditional methods to thousands, greatly enhancing the representativeness and diversity of the data. Combined with the dynamic adaptive matching algorithm, this embodiment can automatically adjust the error tolerance according to the actual environmental conditions, thus effectively eliminating the influence of environmental changes on the calibration accuracy and ensuring millimeter-level accuracy of the final calibration result, meeting the high-precision requirements in autonomous driving and intelligent network applications.

[0086] In this embodiment, the original target data includes the target serial number, target coordinates, and radar acquisition timestamp;

[0087] R radar (t 1 ) = [ID(t 1 ), x target (t 1 ), y target (t 1 )],

[0088] where R radar (t 1 ) is the original target data, ID(t 1 ) is the target serial number, x target (t 1 ) is the target abscissa, y target (t 1 ) is the target ordinate, and t 1 is the radar acquisition timestamp;

[0089] The vehicle trajectory data includes vehicle coordinates and positioning timestamp;

[0090] P RTK (t 2 ) = [x RTK (t 2 ), y RTK (t 2 )],

[0091] where P RTK (t 2 ) is the vehicle trajectory data, x RTK (t 2 ) is the vehicle abscissa, y RTK (t 2 ) is the vehicle ordinate, and t 2 is the positioning timestamp.

[0092] In this embodiment, information such as the target serial number, target coordinates, and radar acquisition timestamp together constitute a complete radar perception dataset. The target serial number ensures the uniqueness of each target, the target coordinates provide the precise position information of the target, and the radar acquisition timestamp guarantees the timeliness of the data. This complete dataset provides a solid foundation for subsequent target recognition, trajectory stitching, and time synchronization, thereby improving the accuracy and reliability of the calibration process.

[0093] Since both the vehicle trajectory data and the original target data of the radar are collected in real time and can be matched and optimized during vehicle driving, this method realizes real-time dynamic calibration. This means that in practical applications, the external parameters of the millimeter-wave radar can be continuously updated as the vehicle moves to adapt to different road and traffic environments. This advantage of real-time and dynamics improves the adaptability and perception ability of the radar.

[0094] A higher sampling frequency means that more data points can be collected within the same time period. Preferably in this embodiment, the sampling frequency of the radar is 20 Hz, that is, sampling once every 50 ms. The sampling frequency of RTK data is 100 Hz, that is, sampling once every 10 ms. RTK data is collected at a sampling frequency of 100 Hz. Compared with traditional low-frequency sampling, it can capture the dynamic changes of the vehicle, such as position, speed, and acceleration, more meticulously. Similarly, the radar operates at a sampling frequency of 20 Hz, and can also sample target information multiple times in a short time, improving the accuracy and resolution of target detection, and providing a more reliable basis for the subsequent calibration process.

[0095] As Figure 2 shown, S2 specifically includes the following steps:

[0096] S21: Perform target recognition on the data collected by the millimeter-wave radar through training a deep neural network model, and perform trajectory reconstruction according to the time series to generate a mapping function;

[0097] S22: Perform mapping processing on the original target data to obtain calibrated radar perception trajectory data.

[0098] In the process of target tracking by traditional millimeter-wave radars, due to environmental factors, radar performance limitations, or limitations of the algorithm itself, the problem of target ID jumping often occurs. The deep neural network model has powerful feature extraction and classification capabilities, and can learn and identify targets in various complex scenarios, such as vehicles and pedestrians. In step S21, in this embodiment, a deep neural network model is trained to perform target recognition on the data collected by the millimeter-wave radar, significantly improving the accuracy and robustness of target recognition, so as to more accurately identify and distinguish different targets, and maintain the stability of the target ID even in complex scenarios.

[0099] Also in step S21, for the problem of discontinuous trajectories caused by the jump of the millimeter-wave radar target ID, by analyzing and comparing radar data at different time points, this embodiment can identify and connect the trajectory segments of the same target, thereby generating a continuous trajectory of the target in the time series, which is of great significance for subsequent calibration processing and trajectory tracking in practical applications.

[0100] In this embodiment, the mapping function is f:

[0101] T car (t 1 ) = f(R radar (t 1 ))

[0102] where, T car (t 1 ) is the calibrated radar perception trajectory data, R radar (t 1 ) is the original target data, and t 1 is the radar acquisition timestamp;

[0103] T car (t 1 ) = [x car (t 1 ), y car (t 1 )]

[0104] where, x car (t 1 ) is the abscissa of the calibrated radar perception target, and y car (t 1 ) is the ordinate of the calibrated radar perception target.

[0105] As Figure 3 shown, S3 specifically includes the following steps:

[0106] S31: For each point in the radar perception trajectory data, find at least one vehicle trajectory data point near the radar acquisition timestamp t 1 of this point;

[0107] S32: Make a dynamic error tolerance determination on the time error between the radar acquisition timestamp t 1 of this point and the positioning timestamp t 2 of the vehicle trajectory data point;

[0108] S33: If the time error is within the dynamic error tolerance range, it is considered that the vehicle trajectory data point matches this point; if the time error is not within the dynamic error tolerance range, adjust the dynamic error tolerance or re-perform the time synchronization and matching process.

[0109] Since traditional static methods use a fixed time error threshold for matching, but in practical applications, the speed and environment changes of the vehicle will cause the error to change dynamically. Therefore, this embodiment introduces a dynamic error tolerance determination. The dynamic error tolerance determination mechanism can flexibly adjust the error tolerance according to the actual situation, thereby improving the accuracy and robustness of the matching, and then flexibly handling the dynamic time error caused by the vehicle speed and environment changes between radar data and RTK data.

[0110] In this embodiment, the time error is represented by the following formula:

[0111] Δt = |t 1 - t 2 |,

[0112] where Δt is the time error, t 1 is the radar acquisition timestamp, and t 2 is the positioning timestamp;

[0113] The dynamic error tolerance is represented by the following formula:

[0114]

[0115] where Δt max is the dynamic error tolerance, k is a constant, and v is the vehicle driving speed.

[0116] The formula of the dynamic error tolerance in this embodiment shows that as the vehicle driving speed changes, the tolerance of the time error will change accordingly, enabling this embodiment to adapt to the data synchronization requirements at different vehicle speeds and improving the robustness of data matching. Even in the case of large vehicle speed changes or complex environments, the system can maintain stable matching performance and reduce the situations of false matching and missed matching.

[0117] As Figure 4 shown, S4 specifically includes the following steps:

[0118] S41: For the N pairs of matching points obtained through step S3, generate a homography matrix H 1 through the geometric transformation relationship, and solve the preliminarily optimized homography matrix H 2 by minimizing the matching error;

[0119] S42: Optimize the action space based on the reward mechanism, adjust the matrix elements of the homography matrix H 2 , reduce the calibration error of the homography matrix H 2 , and obtain the optimal homography matrix H optimal .

[0120] In this embodiment, the homography matrix H 1It provides an accurate initial estimate for subsequent optimization, reducing the complexity and computational amount of the optimization process. The initial optimized homography matrix H is solved by minimizing the matching error 2 , which can further reduce the deviation caused by factors such as noise and sensor errors, laying a foundation for subsequent optimization steps. In this embodiment, a deep reinforcement learning algorithm is introduced. By optimizing the action space through a reward mechanism, the matrix elements of the homography matrix can be dynamically adjusted. Through continuous iteration and optimization, the calibration error can be gradually reduced, and the calibration accuracy of the homography matrix can be continuously improved to flexibly cope with various noises and sensor errors, making the calibration accuracy of the homography matrix reach the optimal.

[0121] In this embodiment, the homography matrix is expressed as:

[0122]

[0123] where H 1 is the initial homography matrix, x RTK (t i ) is the abscissa of the vehicle trajectory data point in the i-th pair of matching points, y RTK (t i ) is the ordinate of the vehicle trajectory data point in the i-th pair of matching points, x car (t i ) is the abscissa of the radar perception trajectory data point in the i-th pair of matching points, y car (t i ) is the ordinate of the radar perception trajectory data point in the i-th pair of matching points, t i is the timestamp after matching;

[0124] The initial optimized homography matrix is expressed as:

[0125]

[0126] In this embodiment, the specific content of step S42 is: the state space is designed as the calibration error of the homography matrix H 2 , the reward function is designed as the negative value of the calibration error, and the action in the action space is designed as the matrix adjustment amount ΔH; the homography matrix H 2 is updated through the matrix adjustment amount to generate a new homography matrix H new , and the calibration error of the updated homography matrix H new is calculated. The reward mechanism of the reward function is used to continuously optimize the action to reduce the calibration error until the optimal homography matrix H optimal is obtained.

[0127] In this embodiment, by designing the state space as the calibration error of the homography matrix, this embodiment can directly optimize the error, ensuring that each iteration to reduce the error makes the homography matrix closer to the true value. The reward function is designed as the negative value of the calibration error, which means that when the calibration error decreases, the system will obtain a greater reward. This reward mechanism encourages the algorithm to continuously find the adjustment amount that can reduce the calibration error, thereby accelerating the optimization process and improving the efficiency. The actions in the action space are designed as matrix adjustment amounts, enabling this embodiment to flexibly adjust each element of the homography matrix to cope with various complex calibration scenarios and ensuring that the optimal homography matrix can be obtained under various conditions. By continuously updating the homography matrix and calculating the updated calibration error, this embodiment can continuously adapt to environmental changes and noise interference, enhancing the robustness of the system and enabling it to maintain stable calibration performance under various conditions.

[0128] During the training process, the DDPG algorithm continuously optimizes the homography matrix through the following steps:

[0129] Experience replay: The algorithm first collects a series of experiences (i.e., sequences of states, actions, and rewards) and stores them in an experience replay buffer. During the training process, the algorithm randomly samples a batch of experiences from this buffer to update the network parameters to increase data diversity and reduce variance during training.

[0130] Target network: To stabilize the training process, two independent networks are adopted in the DDPG algorithm: the actor network and the critic network. At the same time, a target network is introduced to provide a stable training target. The target network is a copy of the actor network and the critic network, but its parameters are updated more slowly, which helps to reduce fluctuations during training.

[0131] Network update: During each update process, the actor network generates the optimal adjustment action (i.e., how to adjust the homography matrix) based on the current state, then executes this action and observes the results (i.e., the new state and reward). The critic network then evaluates the quality of the action selected by the actor network based on this experience and gives the corresponding reward value. By continuously repeating this process and updating the parameters of the actor network and the critic network, the algorithm gradually learns how to adjust the homography matrix to minimize the calibration error.

[0132] In this embodiment, the reward function is expressed as:

[0133]

[0134] where \(R(H)\) is the reward function, \(x\) RTK \((t\) i ) is the abscissa of the vehicle trajectory data point in the \(i\)-th pair of matching points, \(y\) RTK \((t\) i) is the ordinate of the vehicle trajectory data point in the i-th pair of matching points, x car (t i ) is the abscissa of the radar perception trajectory data point in the i-th pair of matching points, y car (t i ) is the ordinate of the radar perception trajectory data point in the i-th pair of matching points, t i is the timestamp after matching, H new is the new homography matrix after action adjustment;

[0135] H new = H 2 + ΔH,

[0136] where, H 2 is the homography matrix of preliminary optimization solved by minimizing the matching error, and ΔH is the matrix adjustment amount.

[0137] Preferably, a fast dynamic automatic calibration method for the extrinsic parameters of a millimeter-wave radar in this embodiment further includes: verifying the optimal homography matrix H optimal and comparing it with the actual calibration data. If the calibration error exceeds the set tolerance, the matching parameters are automatically adjusted and calibration is performed again until the accuracy requirement is met.

[0138] The serial numbers of the embodiments of the present application above are only for description and do not represent the advantages or disadvantages of the embodiments. The above are only the preferred embodiments of the present application, and do not limit the patent scope of the present application accordingly. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

Claims

1. A method for rapid dynamic automatic calibration of millimeter wave radar external parameters, characterized in that: The following steps are involved: S1: The vehicle to be tested drives within the test range of the millimeter-wave radar. The millimeter-wave radar collects the original target data of each target in the surrounding environment, and the real-time dynamic positioning device synchronously collects the vehicle trajectory data; S2: performing target recognition and trajectory splicing on the original target data to obtain corrected radar perception trajectory data; S3: Based on the dynamic time error, time-synchronize the radar perception trajectory data with the vehicle trajectory data, and match the radar perception trajectory data with the vehicle trajectory data through a timestamp; S4: Generate a homography matrix using matching point pairs, and use a deep reinforcement learning algorithm to optimize the homography matrix to reduce calibration errors; S5: Use the optimized homography matrix to calibrate the millimeter wave radar external parameters.

2. The method for rapid dynamic automatic calibration of millimeter wave radar external parameters according to claim 1 is characterized in that: The original target data includes a target serial number, target coordinates and radar acquisition timestamp; R radar (t1)=[ID(t1),x target (t1),y target (t1)], Among them, R radar (t1) is the original target data, ID(t1) is the target sequence number, x target (t1) is the target horizontal coordinate, y target (t1) target ordinate, t1 is the radar acquisition timestamp; The vehicle trajectory data includes vehicle coordinates and positioning timestamp; P RTK (t2)=[x RTK (t2),y RTK (t2)], Among them, P RTK (t2) is the vehicle trajectory data, x RTK (t2) is the vehicle horizontal coordinate, y RTK (t2) is the vehicle longitudinal coordinate, and t2 is the positioning timestamp.

3. The method for rapid dynamic automatic calibration of millimeter wave radar external parameters according to claim 2 is characterized in that: The S2 specifically includes the following steps: S21: Perform target recognition on the data collected by the millimeter-wave radar by training a deep neural network model, reconstruct the trajectory according to the time series, and generate a mapping function; S22: The original target data is mapped to obtain corrected radar perception trajectory data.

4. The method for rapid dynamic automatic calibration of millimeter wave radar external parameters according to claim 3 is characterized in that: The mapping function is f: T car (t1)=f(R radar (t1)), Among them, T car (t1) is the corrected radar perception trajectory data, R radar (t1) is the original target data, t1 is the radar acquisition timestamp; T car (t1)=[x car (t1),y car (t1)], Among them, x car (t1) is the horizontal coordinate of the radar perceived target after correction, y car (t1) is the corrected radar-perceived target ordinate.

5. The method for rapid dynamic automatic calibration of millimeter wave radar external parameters according to claim 2 is characterized in that: The S3 specifically includes the following steps: S31: for each point in the radar perception trajectory data, find at least one vehicle trajectory data point near the radar acquisition timestamp t1 of this point; S32: performing dynamic error tolerance determination on the time error between the radar acquisition timestamp t1 of this point and the positioning timestamp t2 of the vehicle trajectory data point; S33: If the time error is within the dynamic error tolerance range, the vehicle trajectory data point is considered to match this point; If the time error is not within the dynamic error tolerance range, the dynamic error tolerance is adjusted or the time synchronization and matching process is performed again.

6. A millimeter wave radar external parameter rapid dynamic automatic calibration method according to claim 5, characterized in that: The time error is expressed by the following formula: Δt=|t1-t2|, Among them, Δt is the time error, t1 is the radar acquisition timestamp, and t2 is the positioning timestamp; The dynamic error tolerance is expressed by the following formula: Among them, Δt max is the dynamic error tolerance, k is a constant, and v is the vehicle speed.

7. The method for rapid dynamic automatic calibration of millimeter wave radar external parameters according to claim 5, characterized in that: The S4 specifically comprises the following steps: S41: Generate a homography matrix H1 through the geometric transformation relationship of the N pairs of matching points obtained in step S3, and solve the preliminarily optimized homography matrix H2 by minimizing the matching error; S42: Optimize the action space based on the reward mechanism, adjust the matrix elements of the homography matrix H2, reduce the calibration error of the homography matrix H2, and obtain the optimal homography matrix H optimal .

8. The method for rapid dynamic automatic calibration of millimeter wave radar external parameters according to claim 7, characterized in that: The homography matrix is ​​expressed as: Among them, H1 is the preliminary homography matrix, x RTK (t i ) is the horizontal coordinate of the vehicle trajectory data point in the i-th pair of matching points, y RTK (t i ) is the ordinate of the vehicle trajectory data point in the i-th pair of matching points, x car (t i ) is the horizontal coordinate of the radar perception trajectory data point in the i-th pair of matching points, y car (t i ) is the ordinate of the radar perception trajectory data point in the i-th pair of matching points, t i The timestamp after matching; The preliminary optimized homography matrix is ​​expressed as:

9. A millimeter wave radar external parameter rapid dynamic automatic calibration method according to claim 8, characterized in that: The step S42 is specifically as follows: designing the state space as the calibration error of the homography matrix H2, designing the reward function as the negative value of the calibration error, and designing the action in the action space as the matrix adjustment amount ΔH; updating the homography matrix H2 by the matrix adjustment amount to generate a new homography matrix H new , and calculate the updated homography matrix H new The calibration error of the reward function is used to continuously optimize the action to reduce the calibration error until the optimal homography matrix H is obtained. optimal .

10. A millimeter wave radar external parameter rapid dynamic automatic calibration method according to claim 9, characterized in that: The reward function is expressed as: Among them, R(H) is the reward function, x RTK (t i ) is the horizontal coordinate of the vehicle trajectory data point in the i-th pair of matching points, y RTK (t i ) is the ordinate of the vehicle trajectory data point in the i-th pair of matching points, x car (t i ) is the horizontal coordinate of the radar perception trajectory data point in the i-th pair of matching points, y car (t i ) is the ordinate of the radar perception trajectory data point in the i-th pair of matching points, t i is the timestamp after matching, H new is the new homography matrix after action adjustment; H new =H2+ΔH, Among them, H2 is the homography matrix that is initially optimized by minimizing the matching error, and ΔH is the matrix adjustment amount.

Images