A physical constraint calibration method and device for visual ADAS equipment

Abstract

The application provides a physical constraint calibration method and device for visual ADAS equipment, the method comprising: collecting ranging output data of multiple visual ADAS equipment to be optimized and performing multidimensional preprocessing to obtain a unified input vector of each visual ADAS equipment; inputting the unified input vector into a preset multi-source multi-stage fusion network to generate a ranging pseudo true value; generating an error feature vector based on the ranging output data of the visual ADAS equipment and the ranging pseudo true value; inputting the error feature vector into a physical constraint multi-branch network for diagnosis to output a corresponding calibration parameter deviation; correcting the camera calibration parameter of the visual ADAS equipment based on the calibration parameter deviation to obtain corrected ranging output data; constructing a corrected ranging error based on the corrected ranging output data and the ranging pseudo true value; and evaluating the effect of the current round of calibration parameter correction based on the corrected ranging error, the corrected ranging output data and the ranging pseudo true value.

Description

Technical Field

[0001] This application relates to the field of advanced driver assistance technology, and more specifically, to a physical constraint calibration method and apparatus for vision-based ADAS devices. Background Technology

[0002] In recent years, visual ADAS, with visual sensors as its core perception component, has become the mainstream technology in the automotive aftermarket due to its advantages such as rich environmental information, relatively low cost, and road sign recognition capabilities. Visual ADAS systems encompass core functions such as Automatic Emergency Braking (AEB) and Adaptive Cruise Control, playing a crucial role in target recognition, ranging, and scene understanding. To ensure the accuracy and reliability of these functions, rigorous performance evaluation and calibration of aftermarket visual ADAS systems are essential.

[0003] Currently, there are several existing methods for evaluating and calibrating the performance of visual ADAS systems: 1. Commissioning professional testing institutions to perform calibration using expensive ground truth equipment (such as high-precision LiDAR and differential GPS) and specialized facilities, or using existing low-cost single ordinary ADAS equipment as a reference benchmark; 2. Offline calibration methods based on calibration boards, such as the checkerboard method; 3. Online calibration methods, such as online calibration based on lane line parallelism or vanishing points; 4. Purely data-driven deep learning calibration methods, such as end-to-end calibration methods based on neural networks.

[0004] However, existing performance evaluation and calibration methods generally suffer from the dilemma of balancing high accuracy and low cost. First, offline calibration methods are highly demanding in terms of environment and lack practicality in post-installation scenarios. They not only struggle to stably reproduce controlled environments but also cannot handle dynamic parameter drift in everyday use. Second, online calibration methods have poor adaptability and incomplete parameter coverage, making it difficult to handle complex biases caused by multi-parameter coupling. Furthermore, purely data-driven deep learning calibration methods lack physical interpretability and reliability, and are prone to physical distortion. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a physical constraint calibration method and apparatus for visual ADAS devices. By organically integrating multi-source fusion evaluation with physical constraint calibration diagnosis, the camera calibration parameters of visual ADAS devices are corrected and evaluated. Without the need for high-precision external reference equipment and professional testing sites, it achieves physically interpretable adaptive correction of calibration parameters and continuous self-improvement of the system, providing a reliable error benchmark for subsequent intelligent calibration. This significantly reduces the evaluation cost and deployment threshold in aftermarket scenarios, improves the practicality of aftermarket scenarios, and can handle dynamic parameter drift in daily use. Simultaneously, it improves adaptability and parameter coverage, and can handle complex deviations caused by multi-parameter coupling. Furthermore, the parameter prediction results have clear physical basis, improving the physical interpretability and reliability of the model and reducing the likelihood of physical distortion problems, thus achieving a performance evaluation and calibration that balances high accuracy and low cost.

[0006] In a first aspect, embodiments of this application provide a physical constraint calibration method for vision ADAS devices, the method comprising: The ranging output data of multiple vision ADAS devices to be optimized are collected and the ranging output data is preprocessed in multiple dimensions to obtain a unified input vector for each vision ADAS device. The unified input vector is input into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value; wherein, the ranging pseudo-true value represents the calibrated error reference benchmark; Based on the ranging output data of the visual ADAS device and the ranging pseudo-true value, an error feature vector for parameter diagnosis is generated, and the error feature vector is input into a preset physical constraint multi-branch network for diagnosis, and the corresponding calibration parameter deviation is output. The camera calibration parameters of the visual ADAS device are corrected based on the calibration parameter deviation to obtain corrected ranging output data, and a corrected ranging error is constructed based on the corrected ranging output data and the ranging pseudo-true value. The effectiveness of this round of calibration parameter correction is evaluated based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value.

[0007] In one possible implementation, ranging output data from multiple vision ADAS devices to be optimized is collected, and the ranging output data undergoes multi-dimensional preprocessing, including: The message signals output by multiple vision ADAS devices to be optimized are collected, and the message signals are parsed to extract the ranging data output by each vision ADAS device. Transform the ranging data of all vision ADAS devices from the local coordinate system to the unified vehicle coordinate system, and add a unified timestamp to each frame of ranging data; Identify and eliminate the dimensional and numerical range differences among different vision ADAS devices.

[0008] In one possible implementation, the multi-source, multi-stage fusion network includes a parallel feature extraction module and a multi-stage interactive fusion module; the step of inputting the unified input vector into the preset multi-source, multi-stage fusion network to generate ranging pseudo-true values ​​includes: In the parallel feature extraction module, the proprietary feature representation of the visual ADAS device is extracted based on the preset feature mapping function and the unified input vector of the visual ADAS device. In the multi-stage interactive fusion module, the weight coefficient of each visual ADAS device in the fusion process is determined, and based on the weight coefficients of all visual ADAS devices and their corresponding proprietary feature representations, a comprehensive feature representation after feature fusion of all visual ADAS devices is obtained, and a ranging pseudo-true value is generated based on the comprehensive feature representation.

[0009] In one possible implementation, the generation of an error feature vector for parameter diagnosis based on the ranging output data of the visual ADAS device and the ranging pseudo-true value includes: The ranging output data of the visual ADAS device is compared with the ranging pseudo-true value to construct a ranging error sequence of the visual ADAS device relative to the ranging pseudo-true value; Multiple features are extracted based on the ranging error sequence, and an error feature vector for parameter diagnosis is generated based on the multiple features.

[0010] In one possible implementation, the calibration parameter deviation includes at least focal length deviation, pitch angle deviation, and installation height deviation; the step of inputting the error feature vector into a preset physical constraint multi-branch network and outputting the corresponding calibration parameter deviation includes: The error feature vector is divided into multiple branches. Based on the error feature vector, the feature representations of different branches and the attention weight coefficients of the branches are obtained. Based on the attention weight coefficients and the feature representations, the feature representations of different branches are adaptively weighted and fused to obtain the comprehensive feature representation after multi-head attention fusion. The geometric prior of the pinhole camera model is embedded into the physical constraint multi-branch network. The geometric prior represented by the pinhole camera model is used to decouple the parameters of the comprehensive feature representation and output the focal length deviation, pitch angle deviation and installation height deviation. In one possible implementation, evaluating the effect of the calibration parameter correction based on the corrected ranging error, the corrected ranging output data, and the true value of the ranging pseudo-value includes: The corrected ranging error, the corrected ranging output data, and the true ranging value are compared to generate a corrected performance index, and the effect of this round of calibration parameter correction is evaluated based on the performance index. When the corrected performance index meets the corresponding preset threshold, the final camera calibration parameters of the visual ADAS device are output. When the corrected performance index does not meet the corresponding preset threshold, the ranging error sequence is reconstructed to correct the camera calibration parameters of the visual ADAS device.

[0011] In one possible implementation, comparing the corrected ranging error, the corrected ranging output data, and the true ranging value to generate a corrected performance index, and evaluating the effect of this round of calibration parameter correction based on the performance index, includes: Calculate the absolute error, relative error, mean square error, and root mean square error between the corrected ranging output data and the true ranging value, and evaluate the ranging accuracy after the calibration parameters are corrected in this round based on the absolute error, the relative error, the mean square error, and the root mean square error. Determine the corrected mean square error and the original mean square error, and calculate the performance improvement rate based on the corrected mean square error and the original mean square error. Quantitatively evaluate the improvement effect of this round of calibration parameter correction on the equipment ranging performance based on the performance improvement rate.

[0012] Secondly, embodiments of this application also provide a physical constraint calibration device for vision ADAS devices, the device comprising: The first acquisition module is used to collect ranging output data from multiple visual ADAS devices to be optimized and to perform multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each visual ADAS device. The generation module is used to input the unified input vector into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value; wherein, the ranging pseudo-true value represents the calibrated error reference benchmark; The second acquisition module is used to generate an error feature vector for parameter diagnosis based on the ranging output data of the visual ADAS device and the ranging pseudo-true value, and input the error feature vector into a preset physical constraint multi-branch network for diagnosis, and output the corresponding calibration parameter deviation. The module is used to correct the camera calibration parameters of the visual ADAS device based on the calibration parameter deviation, obtain the corrected ranging output data, and construct the corrected ranging error based on the corrected ranging output data and the ranging pseudo-true value. The evaluation module is used to evaluate the effect of the calibration parameter correction in this round based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value.

[0013] In one possible implementation, the first acquisition module is specifically used for: The message signals output by multiple vision ADAS devices to be optimized are collected, and the message signals are parsed to extract the ranging data output by each vision ADAS device. Transform the ranging data of all vision ADAS devices from the local coordinate system to the unified vehicle coordinate system, and add a unified timestamp to each frame of ranging data; Identify and eliminate the dimensional and numerical range differences among different vision ADAS devices.

[0014] In one possible implementation, the multi-source, multi-stage fusion network includes a parallel feature extraction module and a multi-stage interactive fusion module; the generation module is specifically used for: In the parallel feature extraction module, the proprietary feature representation of the visual ADAS device is extracted based on the preset feature mapping function and the unified input vector of the visual ADAS device. In the multi-stage interactive fusion module, the weight coefficient of each visual ADAS device in the fusion process is determined, and based on the weight coefficients of all visual ADAS devices and their corresponding proprietary feature representations, a comprehensive feature representation after feature fusion of all visual ADAS devices is obtained, and a ranging pseudo-true value is generated based on the comprehensive feature representation.

[0015] In one possible implementation, the second acquisition module is specifically used for: The ranging output data of the visual ADAS device is compared with the ranging pseudo-true value to construct a ranging error sequence of the visual ADAS device relative to the ranging pseudo-true value; Multiple features are extracted based on the ranging error sequence, and an error feature vector for parameter diagnosis is generated based on the multiple features.

[0016] In one possible implementation, the calibration parameter deviation includes at least focal length deviation, pitch angle deviation, and installation height deviation; the second acquisition module is specifically used for: The error feature vector is divided into multiple branches. Based on the error feature vector, the feature representations of different branches and the attention weight coefficients of the branches are obtained. Based on the attention weight coefficients and the feature representations, the feature representations of different branches are adaptively weighted and fused to obtain the comprehensive feature representation after multi-head attention fusion. The geometric prior of the pinhole camera model is embedded into the physical constraint multi-branch network. The geometric prior represented by the pinhole camera model is used to decouple the parameters of the comprehensive feature representation and output the focal length deviation, pitch angle deviation and installation height deviation. In one possible implementation, the evaluation module is specifically used for: The corrected ranging error, the corrected ranging output data, and the true ranging value are compared to generate a corrected performance index, and the effect of this round of calibration parameter correction is evaluated based on the performance index. When the corrected performance index meets the corresponding preset threshold, the final camera calibration parameters of the visual ADAS device are output. When the corrected performance index does not meet the corresponding preset threshold, the ranging error sequence is reconstructed to correct the camera calibration parameters of the visual ADAS device.

[0017] In one possible implementation, the evaluation module is specifically used for: Calculate the absolute error, relative error, mean square error, and root mean square error between the corrected ranging output data and the true ranging value, and evaluate the ranging accuracy after the calibration parameters are corrected in this round based on the absolute error, the relative error, the mean square error, and the root mean square error. Determine the corrected mean square error and the original mean square error, and calculate the performance improvement rate based on the corrected mean square error and the original mean square error. Quantitatively evaluate the improvement effect of this round of calibration parameter correction on the equipment ranging performance based on the performance improvement rate.

[0018] Thirdly, embodiments of this application provide an electronic device, including: a processor, a storage medium, and a bus. The storage medium stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the storage medium via the bus, and the processor executes the machine-readable instructions to perform the steps of the physical constraint calibration method for vision ADAS devices as described in any of the first aspects.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the physical constraint calibration method for vision ADAS devices as described in any one of the first aspects.

[0020] This application provides a physical constraint calibration method and apparatus for visual ADAS devices. It collects ranging output data from multiple visual ADAS devices to be optimized and performs multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each visual ADAS device. This unified input vector is then input into a preset multi-source, multi-stage fusion network to generate a ranging pseudo-true value. Based on the ranging output data and the ranging pseudo-true value of the visual ADAS devices, an error feature vector for parameter diagnosis is generated. This error feature vector is then input into a preset physical constraint multi-branch network for diagnosis, outputting the corresponding calibration parameter deviation. Based on the calibration parameter deviation, the camera calibration parameters of the visual ADAS devices are corrected to obtain corrected ranging output data. A corrected ranging error is constructed based on the corrected ranging output data and the ranging pseudo-true value. The effect of this round of calibration parameter correction is evaluated based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value. This application organically integrates multi-source fusion evaluation with physical constraint calibration diagnosis to correct and evaluate camera calibration parameters of visual ADAS devices. Without requiring high-precision external reference equipment or professional testing facilities, it achieves physically interpretable adaptive correction of calibration parameters and continuous self-improvement of the system. This provides a reliable error benchmark for subsequent intelligent calibration, significantly reducing evaluation costs and deployment barriers in aftermarket scenarios, improving practicality in these scenarios, and addressing dynamic parameter drift during daily use. Simultaneously, it improves adaptability and parameter coverage, handling complex deviations caused by multi-parameter coupling. Furthermore, the parameter prediction results have clear physical basis, enhancing the physical interpretability and reliability of the model and reducing the likelihood of physical distortion. Thus, it achieves performance evaluation and calibration that balances high accuracy and low cost.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a physical constraint calibration method for vision-based ADAS devices provided according to an embodiment of this application; Figure 2 This is a schematic diagram of the overall process for physical constraint calibration of vision-oriented ADAS devices; Figure 3This is a schematic diagram of the physical constraint calibration device for vision ADAS devices provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0025] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0026] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0027] In recent years, visual ADAS, with visual sensors as its core perception component, has become the mainstream technology in the automotive aftermarket due to its advantages such as rich environmental information, relatively low cost, and road sign recognition capabilities. Visual ADAS systems encompass core functions such as Automatic Emergency Braking (AEB) and Adaptive Cruise Control, playing a crucial role in target recognition, ranging, and scene understanding. To ensure the accuracy and reliability of these functions, rigorous performance evaluation and calibration of aftermarket visual ADAS systems are essential.

[0028] Currently, there are several existing methods for evaluating and calibrating the performance of visual ADAS systems: 1. Commissioning professional testing institutions to perform calibration using expensive ground truth equipment (such as high-precision LiDAR and differential GPS) and specialized facilities, or using existing low-cost single ordinary ADAS equipment as a reference benchmark; 2. Offline calibration methods based on calibration boards, such as the checkerboard method; 3. Online calibration methods, such as online calibration based on lane line parallelism or vanishing points; 4. Purely data-driven deep learning calibration methods, such as end-to-end calibration methods based on neural networks.

[0029] However, existing performance evaluation and calibration methods generally suffer from the dilemma of balancing high accuracy and low cost. First, offline calibration methods are highly demanding in terms of environment and lack practicality in post-installation scenarios. They not only struggle to stably reproduce controlled environments but also cannot handle dynamic parameter drift in everyday use. Second, online calibration methods have poor adaptability and incomplete parameter coverage, making it difficult to handle complex biases caused by multi-parameter coupling. Furthermore, purely data-driven deep learning calibration methods lack physical interpretability and reliability, and are prone to physical distortion.

[0030] To address this issue, this application provides a physical constraint calibration method and apparatus for visual ADAS devices. By organically integrating multi-source fusion evaluation with physical constraint calibration diagnosis, the camera calibration parameters of visual ADAS devices are corrected and evaluated. Without requiring high-precision external reference equipment or professional testing facilities, it achieves physically interpretable adaptive correction of calibration parameters and continuous self-improvement of the system. This provides a reliable error benchmark for subsequent intelligent calibration, significantly reducing evaluation costs and deployment barriers in aftermarket scenarios, improving practicality in aftermarket applications, and handling dynamic parameter drift during daily use. Simultaneously, it improves adaptability and parameter coverage, handling complex deviations caused by multi-parameter coupling. Furthermore, the parameter prediction results have clear physical basis, enhancing the physical interpretability and reliability of the model and reducing the likelihood of physical distortion. Thus, it achieves performance evaluation and calibration that balances high accuracy and low cost.

[0031] Figure 1 This is a flowchart of a physical constraint calibration method for vision-based ADAS devices provided according to embodiments of this application. Figure 1 As shown in the embodiment of this application, the physical constraint calibration method for vision ADAS devices may specifically include: S101. Collect ranging output data from multiple vision ADAS devices to be optimized and perform multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each vision ADAS device.

[0032] S102. Input the unified input vector into the preset multi-source multi-stage fusion network to generate the ranging pseudo-true value.

[0033] S103. Based on the ranging output data and the ranging pseudo-true value of the visual ADAS device, an error feature vector is generated for parameter diagnosis. The error feature vector is then input into a preset physical constraint multi-branch network for diagnosis, and the corresponding calibration parameter deviation is output.

[0034] S104. Based on the calibration parameter deviation, the camera calibration parameters of the visual ADAS device are corrected to obtain the corrected ranging output data, and the corrected ranging error is constructed based on the corrected ranging output data and the ranging pseudo-true value.

[0035] S105. Evaluate the effect of this round of calibration parameter correction based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value.

[0036] The aforementioned physical constraint calibration method for vision ADAS devices organically integrates multi-source fusion evaluation with physical constraint calibration diagnosis to correct and evaluate the camera calibration parameters of vision ADAS devices. Without requiring high-precision external reference equipment or professional testing sites, it achieves physically interpretable adaptive correction of calibration parameters and continuous self-improvement of the system, providing a reliable error benchmark for subsequent intelligent calibration. This significantly reduces evaluation costs and deployment barriers in aftermarket scenarios. Simultaneously, it provides clear physical basis for parameter prediction results, improving the interpretability and reliability of the model and avoiding the physical distortion problems that easily occur in purely data-driven methods. Thus, it achieves a balance between cost, convenience, scenario adaptability, and physical rationality.

[0037] The exemplary steps described above in the embodiments of this application are illustrated below with specific examples: S101: Collect ranging output data from multiple vision ADAS devices to be optimized and perform multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each vision ADAS device.

[0038] In this embodiment, ranging output data from multiple visual ADAS devices to be optimized are collected and preprocessed in multiple dimensions to obtain a unified input vector for each visual ADAS device for subsequent processing. For example, as Figure 2 As shown.

[0039] In some implementations, message signals output from multiple vision ADAS devices to be optimized are collected, and the message signals are parsed to extract the ranging data output by each vision ADAS device. The ranging data from all vision ADAS devices is transformed from a local coordinate system to a unified vehicle coordinate system, and a unified timestamp is added to each frame of ranging data. The dimensional differences and numerical range differences between different vision ADAS devices are identified and eliminated. The ranging data includes at least ranging information and related status information.

[0040] Specifically, the ranging output data of multiple heterogeneous vision ADAS devices is collected through protocol parsing. The data is then synchronized and standardized to form a unified input vector. Specifically, the local coordinate systems of each heterogeneous vision ADAS device are transformed to a unified vehicle coordinate system, and a unified timestamp is added to each frame of data. This yields a preprocessed unified input vector. .

[0041] Among them, protocol parsing is used to extract ranging information and related status information output by each device; time synchronization is used to ensure that the outputs of multiple devices are comparable under a unified time reference; coordinate alignment is used to map the outputs of different devices to a unified reference coordinate system; and standardization processing is used to eliminate the impact of differences in the dimensions and numerical ranges of different devices on the subsequent fusion process.

[0042] Therefore, preprocessing can provide a high-quality, multi-source consistent data foundation for subsequent fusion and intelligent calibration.

[0043] S102, input the unified input vector into the preset multi-source multi-stage fusion network to generate the ranging pseudo-true value.

[0044] In this embodiment, the ranging pseudo-true value represents the calibration error reference benchmark; the multi-source multi-stage fusion network includes a parallel feature extraction module and a multi-stage interactive fusion module; the unified input vector in step S101 is input into the multi-source multi-stage fusion network to generate the ranging pseudo-true value for subsequent processing. For example, as... Figure 2 As shown.

[0045] In some implementations, in the parallel feature extraction module, a proprietary feature representation of the visual ADAS device is extracted based on a preset feature mapping function and a unified input vector of the visual ADAS device. In the multi-stage interactive fusion module, the weight coefficients of each visual ADAS device in the fusion process are determined, and based on the weight coefficients of all visual ADAS devices and their corresponding proprietary feature representations, a comprehensive feature representation of all visual ADAS devices after feature fusion is obtained. A ranging pseudo-true value is then generated based on this comprehensive feature representation. .

[0046] Specifically, in the parallel feature extraction module: features from each device are extracted. ,in, Indicates the first The preprocessed input vectors of a heterogeneous vision ADAS device This represents the feature mapping function corresponding to the parallel feature extraction module. Indicates the first Each device has its own feature representation. In the multi-stage interactive fusion module: features between devices are gradually fused to generate a fused representation. ; in, Indicates the first The weighting coefficients of individual device features in the fusion process. This represents the comprehensive representation after fusing features from multiple devices. Finally, based on this comprehensive representation... Generate high-confidence ranging pseudo-true values .

[0047] S103 generates an error feature vector for parameter diagnosis based on the ranging output data and the ranging pseudo-true value of the visual ADAS device, and inputs the error feature vector into a preset physical constraint multi-branch network for diagnosis, and outputs the corresponding calibration parameter deviation.

[0048] In this embodiment, the physical constraint multi-branch network includes a multi-branch feature extraction module, a multi-head attention fusion module, a physical constraint embedding module, and a decoupling parameter estimation module; the calibration parameter deviation includes at least focal length deviation, pitch angle deviation, and installation height deviation; an error feature vector for parameter diagnosis is generated by using the ranging output data and ranging pseudo-true values ​​of the visual ADAS device in step S1O1, and the error feature vector is input into the physical constraint multi-branch network for diagnosis, outputting the calibration parameter deviation for subsequent processing. For example, as... Figure 2 As shown.

[0049] Optionally, when generating an error feature vector for parameter diagnosis based on the ranging output data and the pseudo-normal ranging value of the visual ADAS device, the ranging output data and the pseudo-normal ranging value of the visual ADAS device are compared to construct a ranging error sequence of the visual ADAS device relative to the pseudo-normal ranging value; multiple features are extracted based on the ranging error sequence, and an error feature vector for parameter diagnosis is generated based on these multiple features. These multiple features include at least basic statistical features, piecewise error features, relative error features, higher-order statistical features, trend features, distance correlation features, and linear fitting features.

[0050] Specifically, the ranging output data of visual ADAS devices (Same as above) ) and pseudo-truth value contrast:

[0051] in, This represents the ranging results output by the vision ADAS device that needs optimization. This represents the high-confidence ranging pseudo-true value generated by the multi-source fusion module. This represents the ranging error sequence of the device to be optimized relative to the pseudo-true value.

[0052] Continuing on, we will extract seven types of features from the error sequence. (Basic statistics, piecewise error, relative error, higher-order statistics, trend, distance correlation, linear fitting), where the linear fitting feature is expressed as:

[0053] in, This represents the actual ranging data from visual ADAS devices. The slope of the fit represents the change in error with distance. Indicates the error fitting intercept; Used to characterize the trend of error as a function of distance. Used to characterize the degree of error shift in the whole.

[0054] Therefore, by constructing error sequences and extracting features, the variation of errors with distance and time can be characterized, providing input for intelligent calibration.

[0055] Optionally, when inputting the error feature vector into a preset physical constraint multi-branch network and outputting the corresponding calibration parameter deviation, the error feature vector is divided into multiple branches. Based on the error feature vector, the feature representations of different branches and the attention weight coefficients of those branches are obtained. Based on the attention weight coefficients and feature representations, the feature representations of different branches are adaptively weighted and fused to obtain the comprehensive feature representation after multi-head attention fusion. The preset geometric prior of the pinhole camera model is embedded into the physical constraint multi-branch network. The geometric prior represented by the pinhole camera model is used to decouple the parameters of the comprehensive feature representation for estimation, and the focal length deviation, pitch angle deviation, and installation height deviation are output. Specifically, for example, the seven error feature vectors extracted above are input into a pre-trained Physically Constrained Multi-Branch Network (PGMBN). In the multi-branch feature extraction module, three branches are modeled respectively: physical features (including basic statistics and linear fitting), statistical features (including higher-order statistics, piecewise error, and relative error), and trend features (including trend and distance correlation). In the multi-head attention fusion module, features from different branches are adaptively weighted and fused.

[0056] in, Indicates the first The feature representation of each branch output. Indicates the first Attention weight coefficients for each branch, This represents the comprehensive feature representation after multi-head attention fusion.

[0057] In the physical constraint embedding module: the geometric prior of the pinhole camera model is embedded into the network; in the decoupled parameter estimation module: the focal length deviation, pitch angle deviation and installation height deviation are output.

[0058] in, Indicates focal length deviation. Indicates pitch angle deviation. Indicates installation height deviation. This represents the mapping function from the comprehensive characteristics to the parameter deviations.

[0059] It should be noted that by introducing the pinhole camera projection model into the neural network structure and loss constraints, the parameter prediction results have a clear physical basis, which improves the interpretability and reliability of the model and avoids the physical distortion problem that is prone to occur in pure data-driven methods.

[0060] Therefore, through intelligent calibration with physical constraints, the joint diagnosis and intelligent correction of key calibration parameters can be achieved, ensuring that the prediction results conform to physical laws.

[0061] S104, based on the calibration parameter deviation, corrects the camera calibration parameters of the visual ADAS device to obtain the corrected ranging output data, and constructs the corrected ranging error based on the corrected ranging output data and the ranging pseudo-true value.

[0062] In this embodiment, the corrected ranging error is used to characterize the degree of deviation between the ranging output data of the visual ADAS device after the camera calibration parameters are corrected and the ranging pseudo-true value. Based on the calibration parameter deviation in step S103, the camera calibration parameters of the visual ADAS device are corrected. The corrected ranging error is constructed based on the corrected ranging output data and the ranging pseudo-true value for subsequent processing. For example, as... Figure 2 As shown.

[0063] Specifically, the calibration parameters (camera parameters, i.e., focal length, pitch angle, and mounting height) of the vision ADAS device to be optimized are updated based on focal length deviation, pitch angle deviation, and mounting height deviation to obtain the corrected ranging output data. and the true value of the ranging pseudo-value Compare and construct the corrected error :

[0064] in, This indicates the ranging results of the vision ADAS device to be optimized after parameter correction. This represents the high-confidence ranging pseudo-true value generated by the multi-source fusion module. This indicates the distance measurement error after parameter correction.

[0065] It should be noted that the corrected error This step is used to characterize the degree of deviation of the device output after parameter correction from the pseudo-true value, and serves as the input basis for subsequent calibration effect evaluation and closed-loop verification. Its function is to correct the calibration parameters of the device to be optimized based on the diagnosed parameter deviation, and generate corrected error observations for subsequent evaluation, providing input for closed-loop verification.

[0066] S105 evaluates the effectiveness of this round of calibration parameter correction based on the corrected ranging error, the corrected ranging output data, and the true value of the ranging.

[0067] In this embodiment of the application, the effect of the calibration parameter correction in this round is evaluated based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value in step S104.

[0068] In some implementations, the corrected ranging error, the corrected ranging output data, and the true ranging value are compared to generate a corrected performance index, and the effect of the current calibration parameter correction is evaluated based on the performance index. When the corrected performance index meets the corresponding preset threshold, the final camera calibration parameters of the visual ADAS device are output. When the corrected performance index does not meet the corresponding preset threshold, the ranging error sequence is reconstructed to correct the camera calibration parameters of the visual ADAS device.

[0069] Optionally, the absolute error, relative error, mean square error, and root mean square error between the corrected ranging output data and the true ranging value are calculated, and the ranging accuracy after this round of calibration parameter correction is evaluated based on the absolute error, relative error, mean square error, and root mean square error; the mean square error after correction and the mean square error before correction are determined, and the performance improvement rate is calculated based on the mean square error after correction and the mean square error before correction; the improvement effect of this round of calibration parameter correction on the ranging performance of the equipment is quantitatively evaluated based on the performance improvement rate.

[0070] For example, using the corrected error and the corrected distance measurement results The calibration correction effect is quantitatively evaluated by comparing the results with the pseudo-true values. Specifically, the following assessments are conducted: Ranging error evaluation: The absolute error, relative error, mean square error, and root mean square error between the corrected ranging result and the pseudo-true value are calculated to characterize the ranging accuracy after calibration correction; Improvement magnitude evaluation: The root mean square error of the ranging before and after correction is compared to calculate the performance improvement rate, which is used to quantify the improvement effect of calibration correction on the ranging performance of the equipment; Physical consistency evaluation: Based on the theoretical error patterns corresponding to the predicted focal length deviation, pitch angle deviation, and installation height deviation, the consistency between these patterns and the actual observed error sequences is evaluated to verify the physical rationality of the parameter estimation results.

[0071] In addition, there is a closed-loop termination judgment: when the corrected ranging error is lower than the preset threshold, or the performance improvement rate reaches the target requirement, the final calibration result is output; when the preset conditions are not met, the process returns to continue executing error sequence construction and feature extraction, parameter diagnosis and number correction, until the closed-loop optimization is completed.

[0072] Therefore, the ranging performance after parameter correction is quantitatively verified, and a criterion is provided for determining whether closed-loop optimization should be terminated.

[0073] The physical constraint calibration method for vision ADAS devices provided in this application collects ranging output data from multiple vision ADAS devices to be optimized and performs multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each vision ADAS device. The unified input vector is input into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value. Based on the ranging output data and the ranging pseudo-true value of the vision ADAS devices, an error feature vector for parameter diagnosis is generated. The error feature vector is input into a preset physical constraint multi-branch network for diagnosis, and the corresponding calibration parameter deviation is output. Based on the calibration parameter deviation, the camera calibration parameters of the vision ADAS devices are corrected to obtain corrected ranging output data. Based on the corrected ranging output data and the ranging pseudo-true value, a corrected ranging error is constructed. Based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value, the effect of this round of calibration parameter correction is evaluated. This application presents a physical constraint calibration method for visual ADAS devices. By organically integrating multi-source fusion evaluation with physical constraint calibration diagnosis, it corrects and evaluates the camera calibration parameters of visual ADAS devices. Without requiring high-precision external reference equipment or professional testing facilities, it achieves physically interpretable adaptive correction of calibration parameters and continuous self-improvement of the system. This provides a reliable error benchmark for subsequent intelligent calibration, significantly reducing evaluation costs and deployment barriers in aftermarket scenarios, improving practicality in these scenarios, and handling dynamic parameter drift during daily use. Simultaneously, it improves adaptability and parameter coverage, handling complex deviations caused by multi-parameter coupling. Furthermore, the parameter prediction results have clear physical basis, enhancing the physical interpretability and reliability of the model and reducing the likelihood of physical distortion. Thus, it achieves performance evaluation and calibration that balances high accuracy and low cost.

[0074] In summary, this application utilizes ranging data collected from multiple heterogeneous vision ADAS devices installed at the front of the vehicle. Combined with a multi-source, multi-stage feature fusion network, data from different devices is transformed into a unified coordinate system for fusion. This method not only simplifies subsequent calibration and error analysis processes but also enables comparison and analysis of the outputs of each device within a unified space, thereby generating a high-confidence pseudo-true value for ranging. This pseudo-true value serves as the error reference basis for intelligent calibration, improving the accuracy of joint diagnosis of focal length, pitch angle, and installation height, while avoiding reliance on expensive external measuring instruments or complex experimental environments. Furthermore, a closed-loop intelligent calibration mechanism is constructed, using the multi-source fusion pseudo-true value as a reference and a Physically Constrained Multi-Branch Network (PGMBN) as its core. Specifically, through error sequence construction, parameter deviation diagnosis, calibration parameter correction, and correction effect verification, continuous diagnosis and adaptive optimization of focal length deviation, pitch angle deviation, and installation height deviation are achieved, ensuring good stability and sustainable improvement capabilities in the calibration process. Finally, through the structured design of error feature extraction, physical constraint embedding, and parameter decoupling estimation process, the closed-loop evaluation, error analysis, PGMBN intelligent calibration, and parameter correction have good processing efficiency and scene adaptability, thus maintaining high calibration correction effect under different vehicle models, different roads, and complex environments, and has engineering deployment potential.

[0075] Figure 3 This is a schematic diagram of the physical constraint calibration device for vision-based ADAS devices provided in the embodiments of this application; as shown below. Figure 3 As shown in the embodiment of this application, the physical constraint calibration device 300 for vision ADAS devices may specifically include: The first acquisition module 301 is used to collect ranging output data from multiple visual ADAS devices to be optimized and to perform multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each visual ADAS device. The generation module 302 is used to input the unified input vector into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value; wherein the ranging pseudo-true value represents the calibrated error reference benchmark; The second acquisition module 303 is used to generate an error feature vector for parameter diagnosis based on the ranging output data of the visual ADAS device and the ranging pseudo-true value, and input the error feature vector into a preset physical constraint multi-branch network for diagnosis, and output the corresponding calibration parameter deviation. The construction module 304 is used to correct the camera calibration parameters of the visual ADAS device based on the calibration parameter deviation, obtain the corrected ranging output data, and construct the corrected ranging error based on the corrected ranging output data and the ranging pseudo-true value. Evaluation module 305 is used to evaluate the effect of the calibration parameter correction in this round based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value.

[0076] In one possible implementation, the first acquisition module is specifically used for: The message signals output by multiple vision ADAS devices to be optimized are collected, and the message signals are parsed according to the protocol to extract the ranging data output by each vision ADAS device. Transform the ranging data of all vision ADAS devices from the local coordinate system to the unified vehicle coordinate system, and add a unified timestamp to each frame of ranging data; Identify and eliminate the dimensional and numerical range differences among different vision ADAS devices.

[0077] In one possible implementation, the multi-source, multi-stage fusion network includes a parallel feature extraction module and a multi-stage interactive fusion module; the generation module is specifically used for: In the parallel feature extraction module, based on the preset feature mapping function and the unified input vector of the visual ADAS device, the proprietary feature representation of the visual ADAS device is extracted. In the multi-stage interactive fusion module, the weight coefficient of each visual ADAS device in the fusion process is determined, and based on the weight coefficients of all visual ADAS devices and their corresponding proprietary feature representations, the comprehensive feature representation after feature fusion of all visual ADAS devices is obtained, and the ranging pseudo-true value is generated based on the comprehensive feature representation.

[0078] In one possible implementation, the second acquisition module is specifically used for: By comparing the ranging output data of the visual ADAS device with the true ranging value, a ranging error sequence of the visual ADAS device relative to the true ranging value is constructed. Multiple features are extracted based on the ranging error sequence, and an error feature vector for parameter diagnosis is generated based on these features.

[0079] In one possible implementation, the calibration parameter deviations include at least focal length deviation, pitch angle deviation, and installation height deviation; the second acquisition module is specifically used for: The error feature vector is divided into multiple branches. Based on the error feature vector, the feature representations of different branches and the attention weight coefficients of the branches are obtained. Based on the attention weight coefficients and feature representations, the feature representations of different branches are adaptively weighted and fused to obtain the comprehensive feature representation after multi-head attention fusion. The geometric prior of the pinhole camera model is embedded into the physical constraint multi-branch network. The geometric prior represented by the pinhole camera model is used to decouple the parameters of the comprehensive feature representation and output the focal length deviation, pitch angle deviation and installation height deviation. In one possible implementation, the evaluation module is specifically used for: The corrected ranging error, the corrected ranging output data, and the true ranging value are compared to generate the corrected performance index, and the effect of this round of calibration parameter correction is evaluated based on the performance index. When the corrected performance indicators meet the corresponding preset thresholds, the final camera calibration parameters of the visual ADAS device are output. If the corrected performance metrics do not meet the corresponding preset thresholds, the ranging error sequence is reconstructed to correct the camera calibration parameters of the visual ADAS device.

[0080] In one possible implementation, the evaluation module is specifically used for: Calculate the absolute error, relative error, mean square error, and root mean square error between the corrected ranging output data and the true ranging value, and evaluate the ranging accuracy after the calibration parameters are corrected based on the absolute error, relative error, mean square error, and root mean square error. Determine the corrected mean square error and the original mean square error, and calculate the performance improvement rate based on the corrected mean square error and the original mean square error. Quantitatively evaluate the effect of this round of calibration parameter correction on the equipment ranging performance based on the performance improvement rate.

[0081] The physical constraint calibration device for visual ADAS devices provided in this application collects ranging output data from multiple visual ADAS devices to be optimized and performs multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each visual ADAS device. The unified input vector is input into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value. Based on the ranging output data and the ranging pseudo-true value of the visual ADAS devices, an error feature vector for parameter diagnosis is generated. The error feature vector is input into a preset physical constraint multi-branch network for diagnosis, and the corresponding calibration parameter deviation is output. Based on the calibration parameter deviation, the camera calibration parameters of the visual ADAS devices are corrected to obtain corrected ranging output data. Based on the corrected ranging output data and the ranging pseudo-true value, a corrected ranging error is constructed. Based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value, the effect of this round of calibration parameter correction is evaluated. This application presents a physical constraint calibration device for visual ADAS devices. By organically integrating multi-source fusion evaluation with physical constraint calibration diagnosis, it corrects and evaluates the camera calibration parameters of visual ADAS devices. Without requiring high-precision external reference equipment or professional testing facilities, it achieves physically interpretable adaptive correction of calibration parameters and continuous self-improvement of the system. This provides a reliable error benchmark for subsequent intelligent calibration, significantly reducing evaluation costs and deployment barriers in aftermarket scenarios, improving practicality in these scenarios, and handling dynamic parameter drift during daily use. Simultaneously, it improves adaptability and parameter coverage, handling complex deviations caused by multi-parameter coupling. Furthermore, the parameter prediction results have clear physical basis, enhancing the physical interpretability and reliability of the model and reducing the likelihood of physical distortion. Thus, it achieves performance evaluation and calibration that balances high accuracy and low cost.

[0082] like Figure 4 As shown in the embodiment of this application, an electronic device 400 includes a processor 401, a memory 402, and a bus. The memory 402 stores machine-readable instructions that can be executed by the processor 401. When the electronic device is running, the processor 401 communicates with the memory 402 via the bus. The processor 401 executes the machine-readable instructions to perform the steps of the physical constraint calibration method for vision ADAS devices described above.

[0083] Specifically, the memory 402 and processor 401 mentioned above can be general-purpose memory and processor, without any specific limitations. When the processor 401 runs the computer program stored in the memory 402, it can execute the physical constraint calibration method for vision ADAS devices.

[0084] Corresponding to the above-described physical constraint calibration method for vision-based ADAS devices, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described physical constraint calibration method for vision-based ADAS devices.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0086] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0087] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0088] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the deployment methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0089] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A physical constraint calibration method for vision-based ADAS devices, characterized in that, The method includes: The ranging output data of multiple vision ADAS devices to be optimized are collected and the ranging output data is preprocessed in multiple dimensions to obtain a unified input vector for each vision ADAS device. The unified input vector is input into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value; wherein, the ranging pseudo-true value represents the calibrated error reference benchmark; Based on the ranging output data of the visual ADAS device and the ranging pseudo-true value, an error feature vector for parameter diagnosis is generated, and the error feature vector is input into a preset physical constraint multi-branch network for diagnosis, and the corresponding calibration parameter deviation is output. The camera calibration parameters of the visual ADAS device are corrected based on the calibration parameter deviation to obtain corrected ranging output data, and a corrected ranging error is constructed based on the corrected ranging output data and the ranging pseudo-true value. The effectiveness of this round of calibration parameter correction is evaluated based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value.

2. The method according to claim 1, characterized in that, The process of collecting ranging output data from multiple visual ADAS devices to be optimized and performing multi-dimensional preprocessing on the ranging output data includes: The message signals output by multiple vision ADAS devices to be optimized are collected, and the message signals are parsed to extract the ranging data output by each vision ADAS device. Transform the ranging data of all vision ADAS devices from the local coordinate system to the unified vehicle coordinate system, and add a unified timestamp to each frame of ranging data; Identify and eliminate the dimensional and numerical range differences among different vision ADAS devices.

3. The method according to claim 1, characterized in that, The multi-source, multi-stage fusion network includes a parallel feature extraction module and a multi-stage interactive fusion module; the step of inputting the unified input vector into the preset multi-source, multi-stage fusion network to generate ranging pseudo-true values ​​includes: In the parallel feature extraction module, the proprietary feature representation of the visual ADAS device is extracted based on the preset feature mapping function and the unified input vector of the visual ADAS device. In the multi-stage interactive fusion module, the weight coefficient of each visual ADAS device in the fusion process is determined, and based on the weight coefficients of all visual ADAS devices and their corresponding proprietary feature representations, a comprehensive feature representation after feature fusion of all visual ADAS devices is obtained, and a ranging pseudo-true value is generated based on the comprehensive feature representation.

4. The method according to claim 1, characterized in that, The generation of an error feature vector for parameter diagnosis based on the ranging output data of the visual ADAS device and the ranging pseudo-true value includes: The ranging output data of the visual ADAS device is compared with the ranging pseudo-true value to construct a ranging error sequence of the visual ADAS device relative to the ranging pseudo-true value; Multiple features are extracted based on the ranging error sequence, and an error feature vector for parameter diagnosis is generated based on the multiple features.

5. The method according to claim 1, characterized in that, The calibration parameter deviations include at least focal length deviation, pitch angle deviation, and installation height deviation; the step of inputting the error feature vector into a preset physical constraint multi-branch network and outputting the corresponding calibration parameter deviations includes: The error feature vector is divided into multiple branches. Based on the error feature vector, the feature representations of different branches and the attention weight coefficients of the branches are obtained. Based on the attention weight coefficients and the feature representations, the feature representations of different branches are adaptively weighted and fused to obtain the comprehensive feature representation after multi-head attention fusion. The geometric prior of the pinhole camera model is embedded into the physical constraint multi-branch network. The geometric prior represented by the pinhole camera model is used to decouple the parameters of the comprehensive feature representation and output the focal length deviation, pitch angle deviation and installation height deviation.

6. The method according to claim 4, characterized in that, The evaluation of the effectiveness of the calibration parameter correction in this round based on the corrected ranging error, the corrected ranging output data, and the true value of the ranging pseudo-value includes: The corrected ranging error, the corrected ranging output data, and the true ranging value are compared to generate a corrected performance index, and the effect of this round of calibration parameter correction is evaluated based on the performance index. When the corrected performance index meets the corresponding preset threshold, the final camera calibration parameters of the visual ADAS device are output. When the corrected performance index does not meet the corresponding preset threshold, the ranging error sequence is reconstructed to correct the camera calibration parameters of the visual ADAS device.

7. The method according to claim 1, characterized in that, The step of comparing the corrected ranging error, the corrected ranging output data, and the true ranging value to generate a corrected performance index, and evaluating the effect of this round of calibration parameter correction based on the performance index, includes: Calculate the absolute error, relative error, mean square error, and root mean square error between the corrected ranging output data and the true ranging value, and evaluate the ranging accuracy after the calibration parameters are corrected in this round based on the absolute error, the relative error, the mean square error, and the root mean square error. Determine the corrected mean square error and the original mean square error, and calculate the performance improvement rate based on the corrected mean square error and the original mean square error. Quantitatively evaluate the improvement effect of this round of calibration parameter correction on the equipment ranging performance based on the performance improvement rate.

8. A physical constraint calibration device for vision ADAS devices, characterized in that, The device includes: The first acquisition module is used to collect ranging output data from multiple visual ADAS devices to be optimized and to perform multi-dimensional preprocessing on the ranging output data to obtain a unified input vector for each visual ADAS device. The generation module is used to input the unified input vector into a preset multi-source multi-stage fusion network to generate a ranging pseudo-true value; wherein, the ranging pseudo-true value represents the calibrated error reference benchmark; The second acquisition module is used to generate an error feature vector for parameter diagnosis based on the ranging output data of the visual ADAS device and the ranging pseudo-true value, and input the error feature vector into a preset physical constraint multi-branch network for diagnosis, and output the corresponding calibration parameter deviation. The module is used to correct the camera calibration parameters of the visual ADAS device based on the calibration parameter deviation, obtain the corrected ranging output data, and construct the corrected ranging error based on the corrected ranging output data and the ranging pseudo-true value. The evaluation module is used to evaluate the effect of the calibration parameter correction in this round based on the corrected ranging error, the corrected ranging output data, and the ranging pseudo-true value.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the physical constraint calibration method for vision-oriented ADAS devices as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the physical constraint calibration method for vision-based ADAS devices as described in any one of claims 1 to 7.

Images