Vehicle blind area automatic calibration method, system and device based on AI visual recognition

Abstract

This application relates to a method, system, and device for automatic vehicle blind spot calibration based on AI visual recognition, belonging to the fields of image processing and intelligent visual calibration technology. The method includes: S1, target placement and image acquisition; S2, AI feature point recognition and extraction; S3, establishing a mapping relationship between image coordinates and the actual vehicle space; S4, single-reference full-domain deduction of the blind spot alarm area boundary; S5, automatic configuration of blind spot alarm area parameters. The system includes an image acquisition module, an AI feature point recognition module, a coordinate mapping construction module, a blind spot boundary deduction module, and an automatic parameter configuration module. The device includes a main control unit, memory, camera interface, and communication interface, executing the calibration method and distributing parameters. This application requires only a single calibration target to achieve automatic blind spot alarm area calibration under static conditions, with high calibration efficiency and accuracy, suitable for scenarios such as vehicle assembly line production and after-sales maintenance.

Description

Technical Field

[0001] This application relates to the fields of image processing and intelligent visual calibration technology, and in particular to a method, system and device for automatic vehicle blind spot calibration based on AI visual recognition. Background Technology

[0002] Vehicle blind spot monitoring systems are a core active safety feature for improving vehicle driving safety. They use side and rear AI cameras to monitor pedestrians, non-motorized vehicles, and motorized vehicles in real time within the vehicle's blind spot area, providing collision risk warnings. Currently, the mainstream calibration method in the industry is a multi-target manual calibration mode: after the equipment is installed, staff need to manually place multiple calibration targets / plates within the standard blind spot area on the side and rear of the vehicle, adjusting the blind spot alarm boundary coordinates point by point, repeatedly debugging until the alarm area matches the actual blind spot range of the vehicle.

[0003] The traditional calibration scheme has significant technical drawbacks: First, the calibration process is cumbersome and involves many steps, relying on specialized sites and equipment. Calibration of a single vehicle is time-consuming, and batch installation and debugging are extremely inefficient. Second, the calibration process is highly dependent on manual operation experience. Deviations in manually placing the target and manual adjustment errors will directly affect the calibration accuracy of the blind zone area, easily leading to missed or false alarms. Third, after vehicle bumps, sensor disassembly and assembly, and slight chassis deformation, the blind zone parameters are prone to shift, making recalibration difficult and costly, and preventing rapid re-calibration.

[0004] To address the low efficiency of manual calibration, existing technologies propose online self-calibration schemes for BSD (Blind Spot Detection) cameras based on road environment features. For example, patent document CN113256739B discloses a self-calibration method, device, and storage medium for an in-vehicle BSD camera. This scheme uses a camera to capture road images during vehicle movement, and AI identifies and extracts linear road markers such as lane lines and curbs. Based on the identified road features, iteratively corrects the external parameters of the BSD camera, such as the yaw angle, roll angle, and pitch angle, indirectly updating the blind spot alarm area to achieve online self-calibration during driving. However, this scheme relies entirely on naturally existing lane lines and curbs as its calibration benchmark, which has significant limitations: calibration cannot be carried out in open factory areas, repair sites, or other unmarked locations; when road markings are worn or covered by rain, snow, or dirt, feature recognition fails, and the calibration process cannot be completed normally; this scheme is an online calibration during driving and is not suitable for static calibration scenarios after vehicle production or repair.

[0005] Patent document CN121559503A discloses a BSD vehicle blind spot monitoring system and its calibration method based on video-radar fusion. It employs a multi-sensor fusion approach combining cameras and millimeter-wave radar to achieve automatic BSD system calibration. It relies on natural targets in the environment to complete the spatiotemporal registration of vision and radar, generating blind spot alarm areas and supporting both local and remote calibration. While this solution achieves self-calibration through radar-vision fusion without requiring manual deployment of calibration targets, it necessitates the simultaneous use of two sets of sensing hardware—camera and millimeter-wave radar—resulting in high hardware costs and complex algorithm logic. The calibration effectiveness depends on sufficient static targets detectable by radar in the environment; calibration accuracy decreases in open environments lacking effective targets. Furthermore, this solution does not use manually deployed physical calibration targets as calibration benchmarks, making it difficult to adapt to the rapid calibration needs in featureless, static environments. Summary of the Invention

[0006] Therefore, it is necessary to provide an AI-based visual recognition-based method, system, and device for automatic vehicle blind spot calibration to address the aforementioned technical issues. This method can automatically calculate and configure blind spot alarm areas under static conditions using only a single manually set calibration target, thus balancing calibration efficiency and accuracy while being adaptable to various application scenarios such as vehicle manufacturing and after-sales maintenance.

[0007] The automatic vehicle blind spot calibration method based on AI vision recognition provided in this application adopts the following technical solution: An automatic vehicle blind spot calibration method based on AI visual recognition, applied to the configuration of blind spot monitoring areas for passenger cars or trucks, includes the following steps: S1. Calibration target deployment and image acquisition: A single standardized calibration target is statically deployed within the field of view of the vehicle's side and rear camera. The vehicle's BSD camera is activated to acquire image information containing the complete calibration target. Image preprocessing is performed on the acquired image to obtain a preprocessed image. The standardized calibration target has a known fixed physical size and is equipped with several feature points that can be recognized by AI. S2, AI Feature Point Recognition and Extraction: The preprocessed image is processed using an AI visual recognition model to identify and locate the overall area of ​​the target, extract the pixel position data of each feature point on the target in the image coordinate system, and output the set of pixel coordinates of the feature points. S3. Establish the mapping relationship between image coordinates and actual vehicle space: Obtain the camera intrinsic parameters of the BSD camera and its installation pose parameters on the vehicle body. Combine the known real physical dimensions of the calibration target and the feature point pixel position data obtained in S2, solve the extrinsic parameters of the camera device relative to the calibration target plane through the PnP algorithm, and transform the spatial points in the calibration target coordinate system to the vehicle body world coordinate system according to the installation pose parameters, and construct a two-way coordinate mapping relationship between the image pixel coordinate system and the vehicle body world coordinate system. S4. Single-reference global deduction of blind spot alarm area boundary: Using the spatial coordinates corresponding to the feature points of a single calibrated target as the spatial reference, the vehicle's own body parameters are read in, the preset vehicle blind spot geometric model is retrieved, and the theoretical physical coordinates of each corner point of the blind spot alarm area in the vehicle world coordinate system are calculated; using the bidirectional coordinate mapping relationship, the theoretical physical coordinates of each corner point of the blind spot alarm area are projected in reverse to the image coordinate system to generate a closed polygonal boundary of the blind spot alarm area. S5. Automatic configuration of blind zone alarm area parameters: The generated polygon boundary parameters of the blind zone alarm area are automatically written into the blind zone monitoring system, completing the solidification and saving of the alarm area parameters and completing the calibration process.

[0008] By adopting the above technical solution, in S1, starting the vehicle's BSD camera only requires setting up a single calibration target to initiate the calibration process, eliminating the cumbersome traditional method of manually setting up multiple calibration targets. In this application, the BSD camera is a camera device with AI visual recognition capabilities; the two terms have the same meaning. Setting AI-recognizable feature points on the calibration target with known physical dimensions provides a reliable benchmark for subsequent vision-based automatic recognition and spatial coordinate calculation. S2 utilizes an AI visual recognition model to automatically complete the detection and feature point localization of the calibration target, requiring no manual intervention throughout the process. This avoids subjective errors caused by manual point selection and ensures the consistency and repeatability of feature point extraction. In S3, the PnP algorithm is based on known... The 3D-2D correspondence is established to accurately solve the extrinsic parameters of the camera, thereby establishing a two-way mapping between the image coordinate system and the vehicle world coordinate system, laying a mathematical foundation for the accurate projection of the blind spot boundary. Based on the spatial reference established by a single calibration target, combined with the vehicle's own size parameters and the preset blind spot geometric model, the complete blind spot alarm area boundary can be deduced across the entire domain without having to set up calibration targets at multiple locations one by one. The theoretical physical coordinates are projected back to the image coordinate system, and the generated polygonal boundary of the blind spot alarm area can be directly used for the alarm judgment of the blind spot monitoring system. After calibration, the parameters are automatically written and fixed and saved in S5, without the need for manual input or point-by-point configuration, realizing the automated closed loop of the entire calibration process.

[0009] Preferably, the standardized calibration target in S1 is a planar calibration board, and the feature points of the calibration target include the four vertices of the calibration board and the central reference point; the calibration target adopts a high-contrast pattern to adapt to the imaging characteristics of the vehicle-mounted fisheye camera.

[0010] By adopting the above technical solutions, the planar calibration plate is easy to process and has low cost. The feature point distribution formed by the four vertices and the central reference point can provide sufficient geometric constraints, which is conducive to improving the accuracy and robustness of PnP pose calculation. The high-contrast pattern helps to improve the feature extraction success rate of AI visual recognition models under complex lighting conditions.

[0011] Preferably, the image preprocessing in S1 includes fisheye distortion correction and image noise reduction.

[0012] By adopting the above technical solutions, fisheye distortion correction eliminates the inherent imaging distortion of vehicle-mounted wide-angle cameras, establishing an accurate correspondence between pixel coordinates in the image coordinate system and real spatial orientation; image noise reduction filters out sensor noise and environmental interference, providing high-quality input images for subsequent AI feature point recognition.

[0013] Preferably, the AI ​​visual recognition model described in S2 is a lightweight AI visual recognition model. After extracting the pixel position data of each feature point on the calibrated target, it filters out background interference pixels and outputs a set of pixel coordinates of the feature points.

[0014] By adopting the above technical solutions, the lightweight model reduces the consumption of computing resources and can be deployed on vehicle-mounted edge computing platforms to meet real-time requirements; the background interference pixel filtering mechanism further eliminates false detection points in the environment and improves the accuracy of feature point coordinate output.

[0015] Preferably, the vehicle body parameters mentioned in S4 include at least the vehicle body width, the mounting height of the camera on the vehicle body, and the mounting offset.

[0016] By adopting the above technical solution, the vehicle width determines the lateral occupancy of the vehicle body, while the camera installation height and installation offset directly affect the relative pose relationship between the camera's field of view and the vehicle body coordinate system. These parameters together ensure the accuracy of the blind spot alarm area's positioning in the vehicle body's world coordinate system.

[0017] Preferably, the preset vehicle blind spot geometry model in S4 is pre-stored according to the vehicle type, and different vehicle types correspond to different blind spot distance boundary thresholds.

[0018] By adopting the above technical solution, the blind spot range of different vehicle models varies significantly. For example, the side and rear blind spot of heavy trucks is much larger than that of passenger cars. By pre-storing the blind spot geometric models corresponding to multiple vehicle models, this method has good universality and adaptability, and there is no need to redesign the calibration process for different vehicle models.

[0019] Preferably, in S5, when the polygon boundary parameters of the blind zone alarm area are sent to the BSD camera, the polygon coordinate data is sent and stored in the local storage unit of the BSD camera through the communication interface.

[0020] By adopting the above technical solution, parameters are sent using the communication interface and standard BSD protocol commands, which is compatible with the communication specifications of existing BSD camera products, such as the UART communication interface and the calibration copy / paste command of the BSD Camera communication protocol. There is no need to develop additional dedicated communication protocols or modify hardware interfaces, which reduces the difficulty of technology implementation and the cost of modification.

[0021] Preferably, after step S5, the system further includes the following steps: after calibration, the system performs a calibration self-test, reads the saved alarm area parameters, and outputs a calibration failure message if the parameters are abnormal; otherwise, it outputs a calibration completion message.

[0022] By adopting the above technical solution, the calibration self-checking mechanism verifies the validity of the calibration results, promptly detects abnormal parameter writing or calculation errors, avoids missed alarms or false alarms in the blind zone monitoring system during subsequent operation due to calibration failure, and improves the reliability of the calibration process.

[0023] Furthermore, this application also provides an automatic vehicle blind spot calibration system based on AI vision recognition, which, based on any of the above-mentioned automatic vehicle blind spot calibration methods based on AI vision recognition, includes: The image acquisition module is used to statically deploy a single standardized calibration target within the field of view of the vehicle's side and rear camera, then activate the vehicle's BSD camera to acquire image information containing the complete calibration target, and perform image preprocessing on the acquired image to obtain a preprocessed image; the standardized calibration target has a known fixed physical size and is provided with several feature points that can be recognized by AI. The AI ​​feature point recognition module is used to process the preprocessed image using an AI visual recognition model, identify and locate the overall area of ​​the target, extract the pixel position data of each feature point on the target in the image coordinate system, and output the pixel coordinate set of the feature points. The coordinate mapping construction module is used to obtain the camera intrinsic parameters of the BSD camera and its mounting pose parameters on the vehicle body. Combined with the known real physical size of the calibration target and the pixel position data of the feature points, the extrinsic parameters of the camera device relative to the calibration target plane are solved by the PnP algorithm. Based on the mounting pose parameters, the spatial points in the calibration target coordinate system are transformed to the vehicle world coordinate system, and a two-way coordinate mapping relationship between the image pixel coordinate system and the vehicle world coordinate system is constructed. The blind spot boundary deduction module is used to read the vehicle's own body parameters, retrieve the preset vehicle blind spot geometric model, calculate the theoretical physical coordinates of each corner point of the blind spot alarm area in the vehicle world coordinate system, and use the bidirectional coordinate mapping relationship to project the theoretical physical coordinates back to the image coordinate system to generate a closed polygonal boundary of the blind spot alarm area. The automatic parameter configuration module is used to automatically write the generated polygon boundary parameters of the blind zone alarm area into the blind zone monitoring system, thus completing the solidification and saving of the alarm area parameters.

[0024] By adopting the above technical solution, the image acquisition module, AI feature point recognition module, coordinate mapping construction module, blind zone boundary inference module, and parameter automatic configuration module work together in sequence, respectively corresponding to the image acquisition and preprocessing, AI model recognition and feature point extraction, PnP pose calculation and coordinate mapping construction, blind zone alarm area boundary inference, and parameter distribution and fixed storage functions in the method flow. The entire process of the above calibration method is fully automated. The modules transmit information through data interfaces. The modules are clearly divided and have low coupling, which makes it easy to deploy and implement on the vehicle edge computing platform or calibration control terminal.

[0025] Furthermore, this application also provides an AI-based visual recognition-based automatic vehicle blind spot calibration device, applicable to passenger cars or trucks, including: The main control unit is used to perform calibration calculations for blind zone alarm areas and system logic control; A memory, connected to the main control unit, is used to store calibration parameters and system programs; At least one camera interface is provided for connecting a BSD camera to capture image information from the side and rear of the vehicle. A communication interface is used to communicate with a host computer or calibration control terminal, receive calibration instructions, and send calibration parameters to the BSD camera. The main control unit acquires images captured by the BSD camera through the camera interface, executes the above calibration method, and sends the generated blind zone alarm area polygon boundary parameters to the BSD camera through the communication interface to complete parameter solidification and storage.

[0026] By adopting the above technical solution, the main control unit, as the core computing component, undertakes computationally intensive tasks such as PnP pose calculation, coordinate mapping construction, and blind zone boundary deduction in the calibration method. The memory provides reliable local storage for calibration parameters and system programs. The camera interface supports the access of multiple BSD cameras to meet the blind zone calibration requirements of passenger cars or trucks. The communication interface realizes instruction and data transmission with the host computer and BSD cameras. The above hardware composition constructs a complete calibration calculation and parameter configuration platform with low hardware cost. The physical connection relationship between each hardware unit is clear, the data flow is clear, and the above calibration method can be run stably. It has good reliability and maintainability, and is easy to deploy in batches on various vehicle models.

[0027] In summary, this application includes the following beneficial technical effects: 1. This application only requires setting up a single standardized calibration target to complete the calibration, abandoning the traditional operation mode of manually adjusting multiple calibration targets point by point, reducing the number of calibration tools, simplifying the calibration operation process, greatly shortening the calibration time, and improving the efficiency of batch debugging of the whole vehicle at the factory and recalibration after after-sales maintenance.

[0028] 2. This application uses AI vision to automatically identify and calibrate target feature points, and automatically completes coordinate mapping and blind zone boundary deduction throughout the process. This avoids human error caused by manual target placement and parameter adjustment, improves the consistency and accuracy of blind zone alarm area calibration, and reduces the probability of missed alarms and false alarms in the blind zone monitoring system.

[0029] 3. This application is a static calibration scheme that does not rely on road environment features such as lane lines and curbs. Calibration can be carried out directly in open factory areas, repair workshops and other unmarked areas. It is suitable for both vehicle assembly line calibration scenarios and after-sales calibration scenarios at repair stations.

[0030] 4. This application can complete the entire calibration process solely using the vehicle's BSD camera, without requiring additional millimeter-wave radar hardware, thus reducing hardware costs; at the same time, it can call up the corresponding blind spot geometric model according to different passenger car and truck models, making it highly versatile.

[0031] 5. This application can reuse the existing BSD camera and on-board edge computing power of the vehicle, which can be achieved by simply upgrading the algorithm, without the need for large-scale hardware modification, making it easy to upgrade and implement existing BSD blind spot monitoring equipment.

[0032] 6. This application also provides system claims and devices corresponding to the above methods, with a clear system architecture, low hardware cost, and easy deployment on existing vehicle platforms. Attached Figure Description

[0033] Figure 1 This is a flowchart of the calibration method in this application; Figure 2 This is a schematic diagram of the calibration board in an embodiment of this application; Figure 3 This is a functional block diagram of the AI-BSD blind spot monitoring device in the embodiments of this application; Figure 4 This is a schematic diagram of the marker line type options in the calibration interface in an embodiment of this application; Figure 5 This is a hardware connection diagram of the AI-BSD blind spot monitoring device in the embodiments of this application. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0035] In this application, BSD camera refers to a camera device with AI visual recognition capabilities; the two terms have the same meaning.

[0036] An automatic vehicle blind spot calibration method based on AI vision recognition, referring to Figure 1 It includes the following steps: S1. Calibration target deployment and image acquisition: A single standardized calibration target is statically deployed within the field of view of the vehicle's side and rear camera. The vehicle's BSD camera is activated to acquire image information containing the complete calibration target. Image preprocessing is performed on the acquired image to obtain a preprocessed image. The standardized calibration target has a known fixed physical size and is equipped with several feature points that can be recognized by AI. S2, AI Feature Point Recognition and Extraction: The preprocessed image is processed using an AI visual recognition model to identify and locate the overall area of ​​the target, extract the pixel position data of each feature point on the target in the image coordinate system, and output the set of pixel coordinates of the feature points. S3. Establish the mapping relationship between image coordinates and actual vehicle space: Obtain the camera intrinsic parameters of the BSD camera and its installation pose parameters on the vehicle body. Combine the known real physical dimensions of the calibration target and the feature point pixel position data obtained in S2, solve the extrinsic parameters of the camera device relative to the calibration target plane through the PnP algorithm, and transform the spatial points in the calibration target coordinate system to the vehicle body world coordinate system according to the installation pose parameters, and construct a two-way coordinate mapping relationship between the image pixel coordinate system and the vehicle body world coordinate system. S4. Single-reference global deduction of blind spot alarm area boundary: Using the spatial coordinates corresponding to the feature points of a single calibrated target as the spatial reference, the vehicle's own body parameters are read in, the preset vehicle blind spot geometric model is retrieved, and the theoretical physical coordinates of each corner point of the blind spot alarm area in the vehicle world coordinate system are calculated; using the bidirectional coordinate mapping relationship, the theoretical physical coordinates of each corner point of the blind spot alarm area are projected in reverse to the image coordinate system to generate a closed polygonal boundary of the blind spot alarm area. S5. Automatic configuration of blind zone alarm area parameters: The generated polygon boundary parameters of the blind zone alarm area are automatically written into the blind zone monitoring system, completing the solidification and saving of the alarm area parameters and completing the calibration process.

[0037] Example 1 This embodiment provides an automatic vehicle blind spot calibration method based on AI visual recognition, which is applied to the factory calibration scenario of the BSD blind spot monitoring system for passenger vehicles.

[0038] I. Preparation of the calibration environment and deployment of calibration targets Calibration is performed at the vehicle assembly line testing station, with the vehicle stationary, engine off or idling, and braking system locked to ensure the vehicle does not move during calibration.

[0039] The operator retrieves a pre-prepared standardized planar calibration plate. This calibration plate is a rectangular planar plate with a high-contrast black and white checkerboard pattern or circular array pattern. The physical dimensions (length × width) of the calibration plate are precisely measured beforehand and stored in the calibration system's configuration file. (Reference) Figure 2 In this embodiment, the calibration plate is preferably 600mm×600mm in size. A circular pattern is provided at the center of the calibration plate. The circular pattern is divided into four equal-sized fan-shaped sub-regions by orthogonal cross-shaped dividing lines. The four fan-shaped sub-regions are arranged in alternating black and white. The geometric center of the calibration target is the intersection of the cross-shaped dividing lines. This intersection, the circular outline, and the vertices of the four corner triangles together constitute a set of visual recognition feature points, which are used by the AI ​​visual recognition model of the BSD camera to extract coordinate information in the image coordinate system.

[0040] The operator places the calibration plate upright or slightly tilted on the ground reference position at the left or right rear side of the passenger vehicle. It should be noted that this embodiment does not require strictly precise alignment of the calibration plate—no laser rangefinder or specialized positioning fixtures are needed to ensure the precise distance and angle between the calibration plate and the vehicle. The operator only needs to visually confirm that the BSD camera at the side and rear of the vehicle can completely capture the entire calibration plate. This relaxed placement requirement significantly reduces the barrier to entry and time consumption for calibration operations.

[0041] II. Image Acquisition and Preprocessing After the calibration board is deployed, the operator sends an image acquisition command to the BSD camera via a calibration control terminal, such as a laptop, tablet, or dedicated handheld calibration device. The BSD camera responds to the command and acquires a single frame of raw fisheye image containing the entire calibration board at the set resolution and frame rate.

[0042] Because automotive BSD cameras typically use wide-angle or fisheye lenses to obtain a larger field of view, the raw images exhibit significant radial and tangential distortion. Therefore, this embodiment performs a preprocessing procedure immediately after image acquisition: First, the camera intrinsic parameters, calibrated and stored in the system at the time of camera manufacturing, are used to correct the distortion of the original fisheye image based on a polynomial distortion model. The camera intrinsic parameters include focal lengths fx and fy, principal point coordinates cx and cy, and radial distortion coefficients k1, k2, and k3, and tangential distortion coefficients p1 and p2. The essence of distortion correction is to establish a mapping relationship between the pixel coordinates of the distorted image and the pixel coordinates of the corrected image. Pixel remapping is achieved through bilinear interpolation or bicubic interpolation, and the distortion-corrected image is output.

[0043] Secondly, image noise reduction processing is performed on the corrected image. This embodiment uses a Gaussian filtering algorithm, employing a 5×5 Gaussian kernel to perform convolution operations on the image, effectively suppressing thermal noise generated by the CMOS sensor under low illumination conditions and high-frequency noise introduced during image acquisition, while maintaining the clarity of the calibration board edges and feature points.

[0044] After preprocessing, a high-quality corrected and denoised image is obtained, which can be used for subsequent AI feature point recognition.

[0045] III. AI Feature Point Recognition and Extraction The preprocessed image is input into a lightweight AI visual recognition model deployed on an in-vehicle edge computing platform or calibration control terminal. The AI ​​visual recognition model used in this embodiment is a lightweight convolutional neural network based on the MobileNetV3 backbone network, which, after pre-training, can accurately identify the entire calibration board area and individual feature points on the calibration board.

[0046] The processing flow of the AI ​​model is as follows: The first step involves the model performing global semantic segmentation on the input image, locating the entire region of the calibration board within the image, and outputting the bounding box of the calibration board region. This step effectively eliminates interference from background objects such as ground debris, surrounding equipment, and people in the image.

[0047] The second step involves the model precisely locating the pixel coordinates of the four vertices and the center reference point of the calibration board within the image coordinate system using a keypoint detection branch. Specifically, the model outputs the two-dimensional coordinates (ui, vi) of five feature points, where i = 1, 2, 3, 4, 5 correspond to the top-left vertex, top-right vertex, bottom-right vertex, bottom-left vertex, and center reference point, respectively.

[0048] The third step is for the model to assess the confidence level of the calibration board region. If the confidence level of the calibration board region is lower than a preset threshold, such as 0.85, or if the location confidence level of any feature point is lower than a preset threshold, the current image quality is deemed unqualified, and the system automatically re-triggers image acquisition until an image that meets the quality requirements is obtained.

[0049] The fourth step, after feature point extraction, involves background interference pixel filtering: all falsely detected feature points falling outside the bounding box of the calibration board area are removed, retaining only the five valid feature points within the calibration board area. The final output is a set of pixel coordinates for the five feature points, P_img = {(u1,v1), (u2,v2), (u3,v3), (u4,v4), (u5,v5)}.

[0050] IV. Establishing the mapping relationship between image coordinates and the actual space of the vehicle After extracting the pixel coordinates of the feature points, the system reads the camera intrinsic parameter matrix K (containing fx, fy, cx, cy) and distortion coefficients from the BSD camera. Simultaneously, based on the known physical dimensions of the calibration board, the system constructs the three-dimensional coordinates of five feature points on the calibration board plane in the calibration board coordinate system. Using the center reference point of the calibration board as the origin and the calibration board plane as the XOY plane, the calibration board coordinate system is established. The three-dimensional coordinates of the five feature points are as follows: Top left vertex: P1_3D = (-300, 300, 0) mm Top right vertex: P2_3D = (300, 300, 0) mm Bottom right vertex: P3_3D = (300, -300, 0) mm Bottom left vertex: P4_3D = (-300, -300, 0) mm Center reference point: P5_3D = (0, 0, 0) mm At this point, the system has obtained five pairs of 3D-2D corresponding points: { (P1_3D, (u1,v1)), (P2_3D, (u2,v2)), (P3_3D, (u3,v3)), (P4_3D, (u4,v4)), (P5_3D, (u5,v5))}.

[0051] The system inputs the five sets of corresponding 3D-2D point pairs into the PnP (Perspective-n-Point) solver. The PnP algorithm, based on the camera perspective projection model, uses the known coordinates of 3D points and their corresponding 2D projection points in the image to solve for the rotation matrix R and translation vector t of the camera coordinate system relative to the calibration board coordinate system (i.e., the world coordinate system). This embodiment uses the EPnP (Efficient PnP) algorithm, which has the advantages of high computational efficiency and low requirement for the number of point pairs; the five pairs of feature points provide sufficient constraints to ensure solution accuracy.

[0052] The obtained extrinsic parameters R and t, together with the known camera intrinsic parameters K, constitute the complete camera projection matrix P = K[R|t]. Based on this projection matrix, the system establishes a bidirectional coordinate mapping relationship between the image pixel coordinate system and the calibration board coordinate system, and further forms a mapping relationship with the vehicle body world coordinate system: Forward mapping (transformation from world coordinates to image coordinates): Given any 3D point X_w in the vehicle's world coordinate system, transform it to the camera coordinate system through extrinsic parameter transformation, and then obtain the corresponding image pixel coordinates (u,v) through intrinsic parameter projection; Inverse mapping (conversion from image coordinates to world coordinates): Given the image pixel coordinates (u,v) and the height constraint of the spatial plane where the point is located (such as the ground plane Z=0), the corresponding vehicle world coordinates can be obtained by inverse calculation through the projection matrix.

[0053] V. Single-benchmark global simulation of blind zone alarm area boundaries Using the spatial coordinates corresponding to the center reference point of the calibration board as the sole spatial reference, the system begins to deduce the boundary of the blind zone alarm area.

[0054] First, the system reads the vehicle's body parameter configuration file to obtain the following key parameters: Vehicle width W_vehicle, in this embodiment the passenger vehicle width is 1800mm; BSD camera mounting height on the vehicle body H_camera, relative to the ground, in this embodiment the mounting height is 650mm; lateral mounting offset Δx and longitudinal mounting offset Δy of the BSD camera relative to the rear axle center or side of the vehicle body; nominal pitch angle, yaw angle and other mounting angle parameters of the BSD camera.

[0055] Secondly, the system retrieves the corresponding blind spot geometry model from a pre-set blind spot geometry model library based on the vehicle type. This model defines the geometric shape and size parameters of the passenger vehicle's side and rear blind spot warning area in the vehicle's world coordinate system, specifically including: The starting longitudinal distance of the blind spot is calculated from the rear of the vehicle, and is typically 0.5m for passenger cars; The longitudinal distance at the end of the blind spot is typically 10m for passenger vehicles; The lateral width of the blind spot extends outward from the side of the vehicle body, typically 3.5m for passenger cars; The shape of the blind spot area is generally trapezoidal or rectangular polygon.

[0056] Based on the aforementioned vehicle body parameters and blind spot geometry model, the system calculates the theoretical physical coordinates of each corner point of the blind spot warning area in the vehicle's world coordinate system. Taking the right-side BSD blind spot as an example, the blind spot warning area is a quadrilateral, and the coordinates of its four corner points in the vehicle's world coordinate system are as follows: A_w = Starting longitudinal distance of blind spot, right edge of vehicle body + starting lateral width of blind spot, ground height 0; B_w = starting longitudinal distance of blind spot, right edge of vehicle body, ground height 0; C_w = Blind spot termination longitudinal distance, right edge of vehicle body, ground height 0; D_w = Longitudinal distance of blind spot termination, right edge of vehicle body + lateral width of blind spot termination, ground height 0.

[0057] Subsequently, the system uses the bidirectional coordinate mapping relationship established in step four to back-project the theoretical physical coordinates of the four corner points, obtaining their corresponding four pixel coordinate points in the image coordinate system. Connecting these four pixel coordinate points sequentially generates a closed polygonal boundary for the blind spot alarm area. This polygon precisely corresponds to the physical blind spot range behind and to the side of the vehicle on the image plane.

[0058] VI. Automatic configuration and calibration self-test of blind zone alarm area parameters After the polygonal boundary of the blind zone alarm area is generated, the system automatically executes the parameter configuration process.

[0059] The system encapsulates the pixel coordinates of the four corner points of the polygon boundary according to the data frame format of the BSD Camera UART communication protocol. In this embodiment, communication uses a UART serial communication interface with a baud rate set to 250kbps.

[0060] The system constructs configuration instruction frames according to the protocol specifications. The frame structure includes: frame header (0xAA), target address (0x01, indicating the left BSD camera; 0x02, indicating the right BSD camera), instruction code (0x09, indicating the calibration data paste / write instruction), data length, polygon coordinate data body (containing the x and y pixel coordinates of the four corner points, each coordinate is represented by a 16-bit integer), and checksum (CRC8 or cumulative checksum).

[0061] The system sends configuration command frames to the BSD camera via the UART interface. After receiving the command frame, the BSD camera parses the polygon coordinate data, stores it in the camera's local non-volatile memory, and replies with an acknowledgment frame to confirm successful writing.

[0062] After the parameters are written, the system automatically executes the calibration self-test process: The first step is for the system to send a parameter read command (0x0A) to the BSD camera, requesting the camera to return the saved polygon coordinate data of the alarm area; The second step involves the system comparing the read coordinate data with the locally calculated coordinate data point by point. If the deviations are both within the preset tolerance range, and the polygon is a valid closed shape, then the parameters are considered normal. Third, if the parameter verification is normal, the system will output a "calibration complete" message on the calibration control terminal interface and upload the calibration results to the factory production management system for recording and archiving. If the parameter reading fails, the data verification fails, or the polygon shape is abnormal, the system will output a "calibration failed" message and display the specific error code, such as "E01: parameter writing failed", "E02: data verification error", "E03: polygon shape abnormal", guiding the operator to check the communication connection and re-execute the calibration process.

[0063] At this point, the factory calibration of the BSD blind spot monitoring system for passenger vehicles has been completed.

[0064] Example 2 This embodiment provides an automatic vehicle blind spot calibration method based on AI visual recognition, which is applied to the recalibration of blind spot monitoring systems in the after-sales maintenance scenario of light trucks.

[0065] I. Application Scenarios A light truck experienced a side collision during operation, resulting in deformation of the right-side BSD camera bracket. Repair personnel replaced the bracket and reinstalled the original camera. Because the camera's installation position and orientation had changed from the factory settings, the original blind spot alarm parameters were no longer accurate, necessitating recalibration of the blind spot monitoring system.

[0066] The vehicles were parked in the repair workshop's work bays. The workshop floor was ordinary cement, lacking any road markings, lane lines, or curbs for reference. Due to limited space in the repair workshop, it was impossible to deploy multiple calibration targets using a traditional multi-target calibration method.

[0067] II. Calibration Implementation Process S1. Target deployment and image acquisition: The maintenance personnel took out the standardized calibration plate from the tool cabinet and, as in Example 1, placed it on the ground behind the right side of the truck. No measuring tools were needed for precise positioning; visual confirmation was sufficient to ensure the entire calibration plate was captured by the right-side BSD camera.

[0068] The technician launches the calibration application, which connects to the vehicle's calibration control unit via Wi-Fi or Bluetooth. After clicking the start button, the system automatically executes the following procedures.

[0069] The BSD camera on the right captures a frame of raw fisheye image containing the calibration board. The system calls the camera's intrinsic parameters to correct fisheye distortion in the image, and then performs median filtering noise reduction using a 3×3 median filter. This removes impulse noise while better preserving the edge information of the calibration board, resulting in the preprocessed image.

[0070] S2, AI Feature Point Recognition and Extraction: This embodiment uses the YOLOv5s-Pose keypoint detection model, and the system calls a lightweight AI visual recognition model to process the pre-processed image. The model detects and locates the entire calibration board area, extracting the pixel coordinates of the four vertices and the center reference point of the calibration board. After the model outputs the pixel coordinate set of five feature points, the system verifies the confidence level of the calibration board detection. Once the confidence level is confirmed to be higher than 0.9, the system proceeds to the next step.

[0071] S3. Establish the mapping relationship between image coordinates and the actual space of the vehicle: The system reads the camera intrinsic parameters of the BSD camera of the light truck, combines the known physical dimensions of the calibration board and the feature point pixel coordinates extracted in step two, and uses the EPnP algorithm to solve the rotation matrix R and translation vector t of the camera relative to the calibration board plane, and establishes a two-way mapping relationship between the image pixel coordinate system and the vehicle world coordinate system.

[0072] S4. Single-benchmark global simulation of blind zone alarm area boundary: The system reads the vehicle body parameter configuration file of the light truck: Vehicle width: 2200mm; BSD camera mounting height: 850mm; Camera installation offset: Pre-entered based on the camera installation position for this vehicle model.

[0073] The system retrieves the corresponding blind spot geometric model from the blind spot geometric model library based on the light truck model. Compared with passenger cars, light trucks have a larger blind spot range: the starting longitudinal distance of the blind spot is 1.0m, the ending longitudinal distance of the blind spot is 15m, and the lateral width of the blind spot is 4.5m. Based on this model and vehicle body parameters, the system calculates the theoretical physical coordinates of the four corner points of the blind spot warning area in the vehicle's world coordinate system.

[0074] Using the coordinate mapping relationship established in step S3, the theoretical physical coordinates of the four corner points are projected in reverse to the image coordinate system to generate a closed polygonal boundary for the blind zone alarm area.

[0075] S5. Automatic configuration and calibration self-test of blind zone alarm area parameters: The system encapsulates the polygon boundary coordinate parameters into a UART configuration command frame according to the BSD Camera communication protocol, and sends it to the right-side BSD camera via the vehicle's CAN bus to the UART gateway. The camera receives and stores the parameters and then sends a confirmation reply.

[0076] The system performs a calibration self-test: it reads the parameters transmitted back from the camera and compares them with local data for verification. After successful verification, the calibration application interface displays "Calibration Complete," and the entire calibration process takes no more than 2 minutes.

[0077] III. Technical Effects of this Embodiment This embodiment fully illustrates the outstanding advantages of the method of this application in after-sales maintenance scenarios: First, no dedicated site is required. The repair workshop floor does not need any road markings, lane lines, or curbs; ordinary cement flooring is sufficient for calibration, significantly lowering the site requirements for after-sales calibration.

[0078] Secondly, no high-precision tooling is required. The calibration board only needs to be placed within the camera's field of view, eliminating the need for specialized tooling equipment such as laser rangefinders and positioning brackets. Maintenance personnel do not need to receive specialized calibration operation training.

[0079] Third, the calibration efficiency is high. The entire calibration process, from setting up the calibration board to completing parameter configuration and self-test, takes no more than 2 minutes, which is far superior to the traditional manual calibration scheme with multiple calibration targets.

[0080] Fourth, static calibration ensures safety and reliability. The entire calibration process is completed while the vehicle is stationary, eliminating the need for road tests and thus eliminating safety risks during driving.

[0081] Example 3 This embodiment provides an automatic vehicle blind spot calibration method based on AI visual recognition, which is applied to the blind spot monitoring system calibration scenario of heavy trucks.

[0082] I. The Special Characteristics of the Calibration Scenarios Heavy-duty trucks, such as semi-trailer tractors, face the following unique challenges in calibrating their blind spot monitoring systems: First, heavy trucks have a large body size and a wide blind spot range. According to regulations, the side and rear blind spot warning area of ​​heavy trucks is much larger than that of passenger cars and light trucks, with the longitudinal distance of the blind spot termination exceeding 20m and the lateral width exceeding 6m.

[0083] Secondly, the relative angle between the trailer and the tractor changes during driving, especially when turning. Therefore, the blind spot monitoring system of heavy trucks usually requires BSD cameras to be installed on both sides of the tractor and the trailer, which makes the calibration work more extensive.

[0084] Third, heavy truck repair stations typically have more limited space and do not have the conditions to set up multiple calibration targets.

[0085] II. Calibration Implementation Process The calibration process in this embodiment is basically the same as that in Embodiments 1 and 2, except that: 1. Target setting Due to the height of heavy trucks, BSD cameras are typically installed at a height of 1200mm-1500mm. To ensure the calibration board is fully captured by the camera, operators can place the calibration board on a simple stand of a certain height, making the plane of the calibration board roughly perpendicular to the optical axis of the camera. The stand does not require precise adjustment of height and angle; it is only necessary to ensure that the calibration board occupies an appropriate proportion within the camera's field of view, with the calibration board area occupying 15%-30% of the total image area.

[0086] 2. Differentiated invocation of blind zone geometric models After reading the heavy-duty truck type identifier, the system retrieves the corresponding heavy-duty truck blind-spot geometric model from the blind-spot geometric model library. The main parameters of this model are: Starting longitudinal distance of blind spot: 1.5m; Blind spot termination longitudinal distance: 25m. Heavy trucks have long braking distances and require a longer warning distance. Blind spot lateral width: 6.0m extending outward from the side of the vehicle; Blind spot shape: trapezoidal, narrower near the vehicle and wider far from the vehicle, to accommodate the sweeping trajectory of the trailer when turning.

[0087] 3. Calibrate both cameras sequentially. Heavy-duty trucks typically require a BSD camera on each of the left and right sides to cover blind spots. The calibration process in this embodiment is performed on the left and right cameras respectively: First, calibrate the right-side BSD camera: Place the calibration board on the right rear side of the vehicle and execute the complete process in Example 1 to complete the configuration and fixation of the right-side blind spot alarm area parameters.

[0088] Then calibrate the left BSD camera: move the calibration board to the left rear side of the vehicle and repeat the above process to complete the configuration and fixation of the parameters of the left blind spot alarm area.

[0089] After the cameras on both sides are calibrated, the system stores the polygonal parameters of the blind spot alarm areas on the left and right sides respectively, without interfering with each other.

[0090] 4. Verification of calibration results After the blind spot monitoring system for heavy trucks is calibrated, the system performs a more rigorous calibration self-test process: Check whether the polygons in the blind zone alarm areas on both the left and right sides are valid closed shapes; Check whether the blind spot warning areas on the left and right sides are basically symmetrical about the longitudinal centerline of the vehicle, with a tolerance of ±5%. Check whether the area of ​​the blind spot alarm zone is within the preset reasonable range to avoid the blind spot area being abnormally large or small due to calibration board recognition errors.

[0091] Once all checks are passed, the system will output a "calibration complete" message and upload the calibration results to the fleet management platform for recording.

[0092] III. Technical Effects of this Embodiment This embodiment demonstrates the broad applicability of the method of this application to different vehicle models: First, by pre-storing blind spot geometric models of different vehicle types, the method of this application can be seamlessly adapted to various vehicle types such as passenger cars, light trucks, and heavy trucks, without the need to develop a separate calibration algorithm for each vehicle type.

[0093] Second, for heavy trucks that require calibration by dual cameras, the method in this application only requires placing the same calibration plate on both sides of the vehicle to complete all calibrations, without the need to prepare two sets of calibration fixtures.

[0094] Third, the method of this application is fully applicable to the limited site conditions of heavy truck repair stations, requiring no road markings or large open spaces, which greatly reduces the difficulty of after-sales maintenance of the heavy truck blind spot monitoring system.

[0095] Furthermore, this application also provides an automatic vehicle blind spot calibration system based on AI vision recognition, which, based on any of the above-mentioned automatic vehicle blind spot calibration methods based on AI vision recognition, includes: The image acquisition module is used to statically deploy a single standardized calibration target within the field of view of the vehicle's side and rear camera, then activate the vehicle's BSD camera to acquire image information containing the complete calibration target, and perform image preprocessing on the acquired image to obtain a preprocessed image; the standardized calibration target has a known fixed physical size and is provided with several feature points that can be recognized by AI. The AI ​​feature point recognition module is used to process the preprocessed image using an AI visual recognition model, identify and locate the overall area of ​​the target, extract the pixel position data of each feature point on the target in the image coordinate system, and output the pixel coordinate set of the feature points. The coordinate mapping construction module is used to obtain the camera intrinsic parameters of the BSD camera and its mounting pose parameters on the vehicle body. Combined with the known real physical size of the calibration target and the pixel position data of the feature points, the extrinsic parameters of the camera device relative to the calibration target plane are solved by the PnP algorithm. Based on the mounting pose parameters, the spatial points in the calibration target coordinate system are transformed to the vehicle world coordinate system, and a two-way coordinate mapping relationship between the image pixel coordinate system and the vehicle world coordinate system is constructed. The blind spot boundary deduction module is used to read the vehicle's own body parameters, retrieve the preset vehicle blind spot geometric model, calculate the theoretical physical coordinates of each corner point of the blind spot alarm area in the vehicle world coordinate system, and use the bidirectional coordinate mapping relationship to project the theoretical physical coordinates back to the image coordinate system to generate a closed polygonal boundary of the blind spot alarm area. The automatic parameter configuration module is used to automatically write the generated polygon boundary parameters of the blind zone alarm area into the blind zone monitoring system, thus completing the solidification and saving of the alarm area parameters.

[0096] Each step described in the above method embodiments can be implemented by a corresponding functional module. The image acquisition module corresponds to the image acquisition and preprocessing function in S1, the AI ​​feature point recognition module corresponds to the AI ​​model recognition and feature point extraction function in S2, the coordinate mapping construction module corresponds to the PnP pose calculation and coordinate mapping construction function in S3, the blind zone boundary extrapolation module corresponds to the blind zone alarm area boundary extrapolation function in S4, and the automatic parameter configuration module corresponds to the parameter distribution and saving function in S5.

[0097] Furthermore, this application also provides an automatic calibration system for vehicle blind spot warning areas based on AI vision recognition, which is based on any of the aforementioned automatic calibration methods for vehicle blind spot warning areas based on AI vision recognition. This system is software-based. The hardware co-working architecture, whose various functional modules can rely on this application Figure 3 The board-level hardware circuit shown is implemented as follows. The image acquisition module relies on the SC2337P camera sensor and the internal ISP image processing unit of the XM350AIV3 main control chip to complete image acquisition and preprocessing. The AI ​​feature point recognition module, coordinate mapping construction module, and blind zone boundary deduction module run on the XM350AIV3 chip Dual... The software algorithm module on the CPU core sequentially completes feature point recognition and extraction, PnP extrinsic parameter calculation, construction of bidirectional coordinate mapping relationship, and deduction of the polygonal boundary of the blind spot alarm area in the vehicle world coordinate system; the visualization effect of the polygonal boundary generated by the blind spot boundary deduction module can be seen in [reference]. Figure 4 , Figure 4 The marking line patterns corresponding to different types of blind zone alarm areas are shown; the automatic parameter configuration module relies on the SPI FLASH memory and Uart0 communication interface to complete the storage of calibration parameters and the sending and fixing of parameters to the BSD camera.

[0098] This application also provides an AI-based visual recognition-based automatic calibration device for vehicle blind spot warning areas, applicable to passenger cars or trucks. The device's hardware components include a main control unit, memory, at least one camera interface, and a communication interface; the internal hardware board-level architecture is as follows: Figure 3 As shown, the physical connections of the equipment to external peripherals, cables, and connectors are as follows: Figure 5 As shown. The main control unit consists of an XM350AIV3 main control chip and its peripheral circuitry; the memory is an SPI FLASH chip; the camera interface corresponds to... Figure 3 The MIPI and IICO interfaces are manifested at the external level of the entire machine as follows: Figure 5 The signal path consists of Connector A, extension cord, and Connector B. The communication interface includes an onboard Uart0 (Host COM) communication interface and can also be configured with a Bluetooth module, enabling human-machine interaction via a remote control interface. The main control unit acquires images from the BSD camera through the camera interface, executes the aforementioned automatic calibration method, and sends the generated blind zone alarm area polygon boundary parameters to the BSD camera via the communication interface for parameter storage.

[0099] refer to Figure 3 AI, which is the subject of this application Board-level functional module block diagram of the BSD blind spot monitoring device. (See diagram below.) Figure 3 As shown, the main control unit of the device uses the XM350AIV3 main control chip, which integrates an ISP image processing module and a Dual-channel ISP. The CPU core comprises an H.265 codec module and peripheral blocks. An external SC2337P camera sensor is connected, transmitting image data to the XM350AIV3 main control chip via a MIPI interface and handling register configuration via an IICO interface. The XM350AIV3 main control chip connects to an external SPI FLASH memory via an SPI interface to store system programs and calibration parameters. An external Crystal oscillator is connected via an OSC interface to provide the system clock. A Debug interface is provided via Uart0 for program debugging and log output. The main control chip has a USB interface for connecting external USB devices and an external IR sensor via a GPIO interface. The system includes an OUT output module; an external Lamp panel connected via an LED interface; and an external Audio Codec connected to the main control chip. The Audio Codec connects to the Audio in / Audio out peripherals to enable audio signal encoding, decoding, and transmission / reception. In this embodiment, the AI ​​feature point recognition module, coordinate mapping construction module, blind zone boundary deduction module, and automatic parameter configuration module all rely on the internal dual-channel memory of the XM350AIV3 chip. The CPU, ISP, and built-in algorithm program are implemented; the SPI FLASH memory completes the calibration parameter fixation; Uart0 can be used as the Host COM upper computer communication interface to receive upper computer calibration instructions and send calibration parameters to the BSD camera.

[0100] refer to Figure 4 This is a schematic diagram of the marking line types on the calibration interface. After completing the blind spot calibration calculation, the system's human-machine calibration interface offers several marking line types to choose from, including "Horizontal trapezoid," "Left horizontal line," and "Right horizontal line." The "Horizontal trapezoid" marking line consists of a green, yellow, and red trapezoidal frame, matching the trapezoidal / rectangular outline of the polygonal blind spot alarm area. The "Left horizontal line" and "Right horizontal line" correspond to the marking display of the left and right blind spots of the vehicle, respectively. Each marking line can be independently turned on or off. When the marking lines are on, the video image is overlaid with green, yellow, and red marking lines, allowing operators to visually verify the calibration effect of the blind spot alarm area. When the marking lines are off, the video image returns to its original state, meeting different debugging and usage needs.

[0101] refer to Figure 5 Go to apply for AI A schematic diagram illustrating exemplary external hardware connections for a BSD blind spot monitoring device. Figure 5 Only the external peripherals, connectors, and cable connections of the entire device are shown; the internal PCB motherboard, XM350AIV3 main control chip, SPI FLASH memory, onboard Host COM communication interface, and Bluetooth module are internal components and are not shown in the diagram. Figure 5 The diagram illustrates that the internal hardware of this part corresponds one-to-one with the components described in the device claims. The device internally uses a PCB motherboard as the main control unit carrier. The PCB motherboard integrates an XM350AIV3 main control chip and supporting circuitry, performing blind zone alarm area calibration calculations and overall system logic control. An SPI FLASH memory is mounted on the PCB motherboard, connected to the main control unit, and stores calibration parameters and system programs. (See diagram for details.) Figure 5 As shown, the device supports access to 1 The device includes 4 AI cameras. These cameras connect to Connector B via Connector A and an extension cord, forming a signal path for simultaneous acquisition of multiple images from the side and rear of the vehicle. Each camera features a speaker for audio alarm output, a remote control interface for calibration, and a Connector C for the monitor to display video feeds showing the calibration interface and alarm areas in real-time. A Connector D connects to external LED alarm indicators, providing alarm indications based on the pedestrian's front, left, right, or rear position. Communication interfaces include a Host COM interface, Bluetooth module, and Remote interface. The Host COM and Bluetooth modules are onboard interfaces for wired communication with the host computer and wireless communication with the calibration terminal, respectively receiving calibration commands and sending calibration parameters to the BSD camera. The Remote interface is for external remote control. The entire unit is powered by a power connector, and the video signal is output to the monitor via Connector C. During operation, the main control unit acquires images captured by the BSD camera via the camera interface, executes the aforementioned blind zone calibration method, and sends the generated polygon boundary parameters of the blind zone alarm area to the BSD camera through the communication interface to complete the parameter solidification and storage. Figure 5 The physical connection relationships between the external hardware components of the device and the direction of external signal flow are clearly shown, providing an exemplary external hardware implementation scheme for the device claims of this application.

[0102] It should be noted that the above embodiments are merely illustrative examples of the technical solutions of this application and are not intended to limit the scope of protection of this application. Conventional modifications and equivalent substitutions made by those skilled in the art to the technical solutions without departing from the concept of this application all fall within the scope of protection of this application.

Claims

1. An automatic vehicle blind spot calibration method based on AI visual recognition, applied to the configuration of blind spot monitoring areas for passenger cars or trucks, characterized in that, Includes the following steps: S1. Calibration target deployment and image acquisition: A single standardized calibration target is statically deployed within the field of view of the vehicle's side and rear camera. The vehicle's BSD camera is activated to acquire image information containing the complete calibration target. Image preprocessing is performed on the acquired image to obtain a preprocessed image. The standardized calibration target has a known fixed physical size and is equipped with several feature points that can be recognized by AI. S2, AI Feature Point Recognition and Extraction: The preprocessed image is processed using an AI visual recognition model to identify and locate the overall area of ​​the target, extract the pixel position data of each feature point on the target in the image coordinate system, and output the set of pixel coordinates of the feature points. S3. Establish the mapping relationship between image coordinates and actual vehicle space: Obtain the camera intrinsic parameters of the BSD camera and its installation pose parameters on the vehicle body. Combine the known real physical dimensions of the calibration target and the feature point pixel position data obtained in S2, solve the extrinsic parameters of the camera device relative to the calibration target plane through the PnP algorithm, and transform the spatial points in the calibration target coordinate system to the vehicle body world coordinate system according to the installation pose parameters, and construct a two-way coordinate mapping relationship between the image pixel coordinate system and the vehicle body world coordinate system. S4. Single-reference global deduction of blind spot alarm area boundary: Using the spatial coordinates corresponding to the feature points of a single calibrated target as the spatial reference, the vehicle's own body parameters are read in, the preset vehicle blind spot geometric model is retrieved, and the theoretical physical coordinates of each corner point of the blind spot alarm area in the vehicle world coordinate system are calculated; using the bidirectional coordinate mapping relationship, the theoretical physical coordinates of each corner point of the blind spot alarm area are projected in reverse to the image coordinate system to generate a closed polygonal boundary of the blind spot alarm area. S5. Automatic configuration of blind zone alarm area parameters: The generated polygon boundary parameters of the blind zone alarm area are automatically written into the blind zone monitoring system, completing the solidification and saving of alarm area parameters and completing the calibration process.

2. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: The standardized calibration target mentioned in S1 is a planar calibration board. The feature points of the calibration target include the four vertices of the calibration board and the central reference point. The calibration target adopts a high-contrast pattern to adapt to the imaging characteristics of the vehicle-mounted fisheye camera.

3. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: The image preprocessing in S1 includes fisheye distortion correction and image noise reduction.

4. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: The AI ​​visual recognition model in S2 is a lightweight AI visual recognition model. After extracting the pixel position data of each feature point on the calibrated target, it filters out background interference pixels and outputs the set of pixel coordinates of the feature points.

5. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: The vehicle parameters mentioned in S4 include at least the vehicle width, the mounting height of the camera on the vehicle, and the mounting offset.

6. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: The preset vehicle blind spot geometry model in S4 is pre-stored according to the vehicle type, and different vehicle types correspond to different blind spot distance boundary thresholds.

7. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: In step S5, when the polygon boundary parameters of the blind zone alarm area are sent to the BSD camera, the polygon coordinate data is sent and stored in the local storage unit of the BSD camera through the communication interface.

8. The method for automatic vehicle blind spot calibration based on AI visual recognition according to claim 1, characterized in that: Following step S5, the system performs a calibration self-test after calibration is completed, reads the saved alarm area parameters, and outputs a calibration failure message if the parameters are abnormal; otherwise, it outputs a calibration completion message if the parameters are normal.

9. An automatic vehicle blind spot calibration system based on AI visual recognition, based on the automatic vehicle blind spot calibration method based on AI visual recognition as described in any one of claims 1 to 8, characterized in that, include: The image acquisition module is used to statically deploy a single standardized calibration target within the field of view of the vehicle's side and rear camera, then activate the vehicle's BSD camera to acquire image information containing the complete calibration target, and perform image preprocessing on the acquired image to obtain a preprocessed image; the standardized calibration target has a known fixed physical size and is provided with several feature points that can be recognized by AI. The AI ​​feature point recognition module is used to process the preprocessed image using an AI visual recognition model, identify and locate the overall area of ​​the target, extract the pixel position data of each feature point on the target in the image coordinate system, and output the pixel coordinate set of the feature points. The coordinate mapping construction module is used to obtain the camera intrinsic parameters of the BSD camera and its mounting pose parameters on the vehicle body. Combined with the known real physical size of the calibration target and the pixel position data of the feature points, the extrinsic parameters of the camera device relative to the calibration target plane are solved by the PnP algorithm. Based on the mounting pose parameters, the spatial points in the calibration target coordinate system are transformed to the vehicle world coordinate system, and a two-way coordinate mapping relationship between the image pixel coordinate system and the vehicle world coordinate system is constructed. The blind spot boundary deduction module is used to read the vehicle's own body parameters, retrieve the preset vehicle blind spot geometric model, calculate the theoretical physical coordinates of each corner point of the blind spot alarm area in the vehicle world coordinate system, and use the bidirectional coordinate mapping relationship to project the theoretical physical coordinates back to the image coordinate system to generate a closed polygonal boundary of the blind spot alarm area. The automatic parameter configuration module is used to automatically write the generated polygon boundary parameters of the blind zone alarm area into the blind zone monitoring system, thus completing the solidification and saving of the alarm area parameters.

10. An automatic blind spot calibration device for vehicles based on AI visual recognition, applicable to passenger cars or trucks, characterized in that, include: The main control unit is used to perform calibration calculations for blind zone alarm areas and system logic control; A memory, connected to the main control unit, is used to store calibration parameters and system programs; At least one camera interface is provided for connecting a BSD camera to capture image information from the side and rear of the vehicle. A communication interface is used to communicate with a host computer or calibration control terminal, receive calibration instructions, and send calibration parameters to the BSD camera. The main control unit acquires images captured by the BSD camera through the camera interface, executes the vehicle blind spot automatic calibration method based on AI visual recognition as described in any one of claims 1 to 8, and sends the generated blind spot alarm area polygon boundary parameters to the BSD camera through the communication interface to complete parameter solidification and storage.

Images