Panoramic image generation method for assisting vehicle parking navigation

Abstract

The invention relates to the technical field of image processing, in particular to a panoramic image generation method for assisting vehicle parking navigation, which comprises the following steps: acquiring images acquired by a plurality of image acquisition devices deployed on a vehicle at the current moment; determining a content validity parameter of the image according to the pixel value distribution information of the image; effective view angles are screened out from all the collection view angles, and the image of each effective view angle is converted into a corresponding feature map; determining a plurality of adjacent view angle pairs formed by the effective view angles; extracting a plurality of feature points from the feature maps of the two view angles included in the adjacent view angle pairs, generating a plurality of candidate splicing arrays according to pixel coordinates and feature intensity values of the feature points, and determining a target splicing array corresponding to the adjacent view angle pairs; and on the basis of the target splicing array, splicing the feature maps of all the effective visual angles, and carrying out up-sampling processing on the spliced feature maps to generate a panoramic image of the vehicle. According to the invention, the panoramic image splicing precision is improved, and an effective basis is provided for parking navigation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for generating panoramic images to assist vehicle parking navigation. Background Technology

[0002] With the continuous development of vehicle intelligence and mobile computing technology, the number of visual sensors that can be deployed around vehicles is constantly increasing. These devices have the advantages of convenient deployment and strong real-time data acquisition, and can quickly capture the actual environmental information around the vehicle, providing basic support for vehicle-centric navigation. As the core carrier for intuitively presenting the 360° environment around the vehicle in parking navigation, the quality of panoramic imaging directly affects the safety and convenience of parking operations, and has become an important part of vehicle intelligent assisted driving systems.

[0003] Currently, most panoramic image generation solutions involve deploying multiple image acquisition devices around the vehicle to capture environmental images from different perspectives, which are then stitched together to achieve all-around environmental perception. However, these solutions typically stitch together images directly from the original data, failing to adequately consider factors such as differences in perspective, inconsistent rates of image change, partial occlusion, or inconsistent image quality due to proximity to obstacles. These factors result in panoramic images generated by existing solutions often exhibiting large stitching errors, discontinuous environmental feature presentation, and blurred details, leading to insufficient overall perception accuracy and making it difficult to meet the precise identification requirements of key information such as the distribution of surrounding obstacles, parking space boundaries, and curbs during parking navigation. Summary of the Invention

[0004] To address the technical problems of insufficient stitching accuracy and coherence in existing panoramic image generation schemes for assisting vehicle parking navigation, which prevent accurate support for parking navigation, the present invention aims to provide a panoramic image generation method for assisting vehicle parking navigation. The specific technical solution adopted is as follows: Firstly, a panoramic image generation method for assisting vehicle parking navigation is provided. The method includes: acquiring images captured at the current moment by multiple image acquisition devices deployed on the vehicle, each image acquisition device corresponding to an acquisition viewpoint; for each image, determining a content validity parameter based on the pixel value distribution information of the image, the content validity parameter being used to characterize the richness and saliency of information in the image used for panoramic stitching; based on the content validity parameter, selecting valid viewpoints from all acquisition viewpoints, and converting the image of each valid viewpoint into a corresponding feature map, where each pixel in the feature map has a feature intensity value; determining multiple adjacent viewpoint pairs composed of valid viewpoints based on the deployment position relationship of the image acquisition devices; for each adjacent viewpoint pair, extracting multiple feature points from the feature maps of the two viewpoints included in the adjacent viewpoint pair, generating multiple candidate stitching arrays based on the pixel coordinates and feature intensity values ​​of the feature points, and determining the target stitching array corresponding to the adjacent viewpoint pair from the multiple candidate stitching arrays; stitching the feature maps of all valid viewpoints based on the target stitching array corresponding to each adjacent viewpoint pair, and upsampling the stitched feature maps to generate a panoramic image of the vehicle.

[0005] In one possible design, the content validity parameters of the image are determined based on the pixel value distribution information, including: generating a pixel value distribution histogram of the image; determining a pixel value concentration parameter and a pixel value difference parameter based on the pixel value distribution histogram, wherein the pixel value concentration parameter is used to characterize the degree of concentration of pixel value distribution in the image, and the pixel value difference parameter is used to characterize the overall degree of difference between different pixel value regions in the image; and determining the content validity parameters based on the pixel value concentration parameter and the pixel value difference parameter.

[0006] In one possible design, determining the pixel value concentration parameter and pixel value difference parameter of the image includes: determining the target pixel value with the largest number of pixels from the pixel value distribution histogram, and determining the ratio of the number of target pixels with the target pixel value to the total number of pixels in the image as the pixel value concentration parameter; for each target pixel, determining the pixel value difference between the target pixel and its neighboring pixels; and determining the pixel value difference parameter based on the statistical characteristics of the pixel value differences corresponding to all target pixels.

[0007] In one possible design, valid viewpoints are selected from all captured viewpoints based on content validity parameters. This includes: sorting all captured viewpoints in descending order according to their corresponding content validity parameters to form a validity sequence, and determining the interval between adjacent content validity parameters in the validity sequence; determining the target interval with the largest value from all calculated intervals; dividing the validity sequence into a first subsequence and a second subsequence based on the position of the target interval in the validity sequence, where any content validity parameter in the first subsequence is higher than any content validity parameter in the second subsequence; and determining the captured viewpoints corresponding to the content validity parameters in the first subsequence as valid viewpoints.

[0008] In one possible design, multiple candidate stitching arrays are generated based on the pixel coordinates and feature intensity values ​​of feature points. This includes: for any two feature points in the feature map of each viewpoint in an adjacent viewpoint pair, determining the binding correlation degree between the two feature points based on the continuity of the pixel gradient direction on the path connecting the two feature points and the difference between the feature intensity values ​​of the two feature points. The binding correlation degree is used to characterize the degree of fit between the two feature points as stitching reference points. For each feature point in the feature map, a candidate stitching array is formed by combining the feature point and feature points with a binding correlation degree higher than a preset correlation degree threshold. After traversing all feature points in the feature map, multiple candidate stitching arrays are obtained.

[0009] In one possible design, determining the target stitching array corresponding to an adjacent viewpoint pair from multiple candidate stitching arrays includes: determining the stitching accuracy between the first candidate stitching array and the second candidate stitching array for the first candidate stitching array of the first viewpoint feature map and the second candidate stitching array of the second viewpoint feature map in the adjacent viewpoint pair; and determining the target stitching array from multiple candidate stitching arrays based on the stitching accuracy.

[0010] In one possible design, determining the stitching accuracy between the first candidate stitching array and the second candidate stitching array includes: determining matching feature points in the first and second candidate stitching arrays based on pixel coordinate matching; obtaining the number of matching feature points, the number of first feature points contained in the first candidate stitching array, and the number of second feature points contained in the second candidate stitching array; determining a first evaluation factor based on the number of matching feature points, the number of first feature points, and the number of second feature points; obtaining the mean of the binding correlation degree corresponding to the matching feature points in the first and second candidate stitching arrays; obtaining the sum of the binding correlation degrees corresponding to other feature points besides the matching feature points in the first and second candidate stitching arrays; determining a second evaluation factor based on the mean and the sum; and determining the stitching accuracy based on the first and second evaluation factors.

[0011] In one possible design, based on the target stitching array corresponding to each adjacent viewpoint pair, the feature maps of all effective viewpoints are stitched together, including: for each adjacent viewpoint pair, aligning the matching feature points in the target stitching array corresponding to the adjacent viewpoint pair according to pixel coordinates; determining the sum of the combination correlation of each candidate stitching array in the target stitching array, determining the acquisition viewpoint corresponding to the candidate stitching array with the larger sum as the primary viewpoint, and determining the other acquisition viewpoint as the auxiliary viewpoint; determining the feature point position adjustment amount for fusing the primary viewpoint and the auxiliary viewpoint based on the stitching accuracy of the target stitching array; adjusting the feature point positions in the primary viewpoint and the auxiliary viewpoint respectively according to the position adjustment amount; and fusing the feature maps of the primary viewpoint and the auxiliary viewpoint based on the adjusted feature point positions.

[0012] In one possible design, determining matching feature points based on pixel coordinate matching in the first candidate stitching array and the second candidate stitching array includes: obtaining the geometric center coordinates of the first candidate stitching array and the second candidate stitching array; matching between the candidate stitching array of the first view feature map and the candidate stitching array of the second view feature map according to the prior orientation relationship of adjacent view pairs and the proximity of the geometric center coordinates; and determining the feature points with position coordinate deviations less than a preset deviation threshold in the successfully matched candidate stitching array pairs as matching feature points.

[0013] In one possible design, multiple feature points are extracted from the feature maps of the two viewpoints included in the adjacent viewpoint pair, including: for each feature map, determining the pixels in the feature map whose feature intensity value is higher than a preset intensity threshold as candidate feature points; for each candidate feature point, if there are multiple candidate feature points within a preset range centered on the candidate feature point, retaining the candidate feature point with the largest feature intensity value as the feature point.

[0014] The present invention has the following beneficial effects: In the panoramic image generation method for assisting vehicle parking navigation provided by this invention, high-value images and inefficient images are accurately identified through content validity parameters, and effective viewpoints with sufficient stitching information are selected. This reduces the interference of invalid data on subsequent processes and lowers computational power consumption. Then, feature map transformation is used to strengthen the core features of the image, and the construction of candidate stitching arrays and target stitching arrays achieves accurate cross-view feature matching, providing a reliable reference for stitching. Finally, the target stitching array is used as the core to complete the accurate stitching of the feature maps, and upsampling is used to supplement details and improve resolution, ensuring that the generated panoramic image can clearly and coherently present key navigation information such as obstacle distribution, parking space boundaries, and curbs in the 360° environment around the vehicle. The entire process follows an "end-to-end" processing logic, with all calculations revolving around high-quality data from effective viewpoints. This ensures the real-time generation of panoramic images and significantly improves the stitching accuracy and visual clarity of the images, providing intuitive and accurate environmental perception support for vehicle parking navigation. This effectively reduces the operational difficulty caused by blind spots and inaccurate environmental judgment during parking, improving parking safety and convenience. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a panoramic image generation method for assisting vehicle parking navigation, provided in one embodiment of the present invention. Figure 2 This is a schematic diagram of a panoramic image generation system for assisting vehicle parking navigation, provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a panoramic image generation method for assisting vehicle parking navigation proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0019] In the description of this invention, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" and "more than one" refer to two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] The following describes in detail, with reference to the accompanying drawings, a specific scheme of the panoramic image generation method for assisting vehicle parking navigation provided by the present invention.

[0022] Please see Figure 1 The diagram illustrates a method flowchart for generating panoramic images to assist vehicle parking navigation according to an embodiment of the present invention, including the following steps S101-S107.

[0023] S101. Acquire images captured at the current moment by multiple image acquisition devices deployed on the vehicle.

[0024] Each image acquisition device corresponds to a single acquisition viewpoint.

[0025] In some embodiments, the image acquisition device is a camera used to collect visual information about the environment around the vehicle. It is evenly deployed around the vehicle, for example, at the center of the front bumper, the center of the rear bumper, the left door handle, the right door handle, the front wheel arch, and the rear wheel arch, with a total of 6-8 cameras deployed to ensure that the acquisition range covers 360° around the vehicle without blind spots. Each image acquisition device corresponds to a unique acquisition angle and is used to independently acquire environmental feature information around the vehicle from the corresponding angle.

[0026] One possible implementation involves sending a synchronous acquisition command to all image acquisition devices via the onboard central processing unit. This command ensures that all devices expose at the same physical moment, resulting in the acquisition of multiple images with strictly synchronized timestamps at the current time. During acquisition, the image data is timestamped and aligned with the vehicle's real-time position data (provided by the onboard Global Positioning System (GPS) module) and attitude data (provided by the onboard gyroscope) via the Controller Area Network (CAN) bus, ensuring that each frame is precisely correlated with the vehicle's current motion state.

[0027] It should be noted that all acquired images are RGB color images with a resolution of 1920×1080 pixels. The images contain core environmental feature information required for panoramic stitching, such as obstacles around the vehicle (e.g., other parked vehicles, curbs, pillars, pedestrians, etc.), ground parking markings, parking space boundary lines, and spatial distance reference marks (e.g., ground scale lines, wall markings, etc.). All image data is stored in real time on the vehicle's local storage unit, using an "end-to-end" processing method to avoid delays caused by data transmission to the cloud and ensure the real-time performance of subsequent image processing.

[0028] To achieve continuous, real-time vehicle-centric navigation while the vehicle is in motion, the image acquisition device continuously acquires images at a preset frame rate (e.g., 30 frames per second) to form a video stream. "Current moment" refers to the specific synchronized frame moment extracted from the continuous video stream used to generate a single-frame panoramic image. In subsequent processing, a set of synchronized images acquired at the synchronized frame moment is processed to generate the corresponding panoramic image.

[0029] In some embodiments, after the image is acquired, it is preprocessed, including depigmentation, automatic white balance, and gamma correction, to correct the inherent optical distortion of the image acquisition device and improve image quality, providing standardized input for subsequent feature extraction and stitching.

[0030] S102. For each image, determine the content validity parameter of the image based on the pixel value distribution information of the image.

[0031] Among them, the content validity parameter is used to characterize the richness and salience of the information in the image used for panoramic stitching.

[0032] One possible implementation is to determine the pixel value of each pixel in each image. This pixel value can be the average of the RGB three color channels, or it can be directly based on the pixel value after grayscale processing. Then, based on the pixel value of each pixel in the image, a pixel value distribution histogram is generated for the corresponding image.

[0033] In some embodiments, in the generated pixel value distribution histogram, the horizontal axis represents the pixel value, and the vertical axis represents the number of pixels corresponding to the pixel value in the image.

[0034] Furthermore, based on the pixel value distribution histogram, the pixel value concentration parameter of the image is determined. The pixel value concentration parameter is used to characterize the degree of concentration of pixel value distribution in the image.

[0035] In some embodiments, the target pixel value with the largest number of pixels is first determined from the pixel value distribution histogram. The ratio of the number of target pixels to the total number of pixels in the image is then determined as the pixel value concentration parameter of the image, expressed by the formula: In the formula, For the first The number of target pixels in an image For the first The total number of pixels in the image is not zero. Pixel value concentration parameter. The larger the value, the closer the gray values ​​of most pixels in the image are, which may correspond to areas with simple content, sparse texture, or large areas of occlusion, and thus have low information richness.

[0036] Furthermore, based on the pixel value distribution histogram, a pixel value difference parameter is determined, which is used to characterize the overall degree of difference between different pixel value regions in the image.

[0037] In some embodiments, all target pixels are located in the image. For each target pixel, its neighboring pixels are determined. Neighboring pixels can be pixels within the four-neighbor or eight-neighbor range of the target pixel, excluding other target pixels. The pixel value difference between the target pixel and its neighboring pixels is then determined, as expressed by the formula: , The pixel value of the target pixel. For the target pixel point The pixel value of each adjacent pixel. This represents the total number of neighboring pixels of the target pixel. After obtaining the pixel value differences between all target pixels and their neighboring pixels, the mean of all pixel value differences is determined and denoted as . and the standard deviation of all pixel value differences. Based on the statistical characteristics of the pixel value differences corresponding to all target pixels, the pixel value difference parameter is determined, and its formula is expressed as follows: , For a very small positive number, an empirical value of 0.001 can be taken to prevent the denominator from being zero. Numerator Reflects the average difference intensity between the main area and the surrounding areas, denominator The pixel value difference parameter reflects the spatial variability of this difference. The larger the value, the stronger the average contrast between the main image area and the surrounding environment, and the clearer and more consistent the boundary, and the higher its information saliency.

[0038] Finally, based on the calculated pixel value concentration parameter and pixel value difference parameter Determine the content validity parameters.

[0039] In some embodiments, the formula for calculating the content validity parameter is as follows: In the formula, For content validity parameters, This is a parameter representing the concentration of pixel values. This is a parameter representing the pixel value difference. , These are the pixel value concentration parameters. and pixel value difference parameter The corresponding weight coefficients all have values ​​greater than zero and The weighting coefficients can be determined based on prior knowledge or experimental data, and are used to adjust the pixel value concentration parameter. and pixel value difference parameter The relative importance in the final evaluation. For example, it can be increased in scenarios requiring rich features. ,like , In scenarios requiring strong feature contrast and stability, it can increase ,like , .

[0040] Based on the above implementation process, the content validity parameters corresponding to each image can be obtained.

[0041] In some embodiments, if the vehicle speed sensor determines that the vehicle is stationary (vehicle speed is 0 km / h), then there is no need to recalculate the content validity parameters of the current image. Instead, the content validity parameters of the corresponding image at the previous moment, as well as the pixel value concentration parameters and pixel value difference parameters obtained during the calculation process, can be directly reused. This ensures the consistency of the parameters and avoids the waste of computing power caused by repeated calculations in a stationary state.

[0042] Understandably, this invention uses histograms to fully capture the pixel distribution patterns of images. Two core parameters characterize the concentration of pixel distribution and the degree of difference between different regions, avoiding the one-sidedness caused by single-dimensional evaluation. It can effectively distinguish between low-quality images such as those with camera obstruction, those too close to obstacles, or those with monotonous scenes, and high-quality images with rich environmental features. This makes the calculation results of the content effectiveness parameter more in line with the actual needs of panoramic stitching materials, providing a scientific and reliable quantitative basis for the subsequent selection of effective viewpoints, and ensuring the quality of basic materials for panoramic image stitching from the source.

[0043] S103. Based on the content validity parameter, select valid viewpoints from all collected viewpoints and convert the images of each valid viewpoint into corresponding feature maps.

[0044] In the feature map, each pixel has a feature intensity value.

[0045] One possible implementation is to sort the acquisition viewpoints of all image acquisition devices around the vehicle in descending order according to their respective content validity parameters, forming an ordered validity sequence. In the validity sequence, each element is the content validity parameter corresponding to a single acquisition viewpoint. The interval between adjacent content validity parameters in the validity sequence is then determined, i.e., the difference between adjacent content validity parameters (previous term minus subsequent term). This interval reflects the drastic decrease in content validity across sorted viewpoints.

[0046] Furthermore, the target interval with the largest value is determined from all calculated intervals. Since the validity sequence is arranged in descending order, the largest interval corresponds to a significant transition from "high at the beginning to low at the end," which is the boundary between valid and inefficient perspectives. Then, based on the position of the target interval in the validity sequence, the validity sequence is divided into a first subsequence and a second subsequence. The first subsequence consists of the perspective preceding the target interval and all perspectives before it, while the second subsequence consists of the perspective following the target interval and all perspectives after it. Any content validity parameter in the first subsequence is higher than any content validity parameter in the second subsequence.

[0047] Furthermore, the collection perspective corresponding to the validity parameter of the content contained in the first subsequence is determined as the valid perspective.

[0048] In some embodiments, a minimum interval threshold is preset, which can be 10% of the average content validity parameter. When the target interval is less than the minimum interval threshold, it is considered that the difference in validity of all viewpoints is not significant, and all viewpoints, or the top few viewpoints (a fixed number, such as 90% of the total number of image acquisition devices) are determined as valid viewpoints.

[0049] Finally, the image from each effective viewpoint is converted into a corresponding feature map, where each pixel has a feature intensity value.

[0050] In some embodiments, for each effective viewpoint image, a preset feature extraction module is used to perform image transformation. This preset feature extraction module is a deep learning module containing five convolutional layers, deployed on the vehicle's local computing unit, following the "end-to-end" processing principle to ensure the real-time nature of the transformation process (matching the real-time requirements of parking navigation).

[0051] During the conversion, the preprocessed image of the effective viewpoint is first input into the first convolutional layer of the feature extraction module. Convolution is performed using a 3×3 kernel to perform secondary noise filtering (further removing environmental interference) and initially extract low-order features (such as basic edges and areas of abrupt brightness changes). Subsequently, the feature abstraction level is gradually improved through four subsequent convolutional layers. Each convolutional layer enhances the non-linear expressive power of the features through activation functions, while pooling operations preserve key features and reduce data dimensionality. Finally, a feature map with the same size as the original image is output.

[0052] The feature intensity value of each pixel in the feature map is obtained by normalizing the output signal value of the last convolutional layer of the feature extraction module, and the value range is [0,1]. The magnitude of the feature intensity value is positively correlated with the saliency of the environmental features at the corresponding location: if a pixel in the original image is a strong feature region such as the edge of an obstacle or the boundary of a parking space line, its feature intensity value in the feature map will approach 1; if a pixel in the original image is a background region without obvious features (such as a flat ground without texture), its feature intensity value in the feature map will approach 0. By quantifying the feature intensity value, the accurate location and representation of effective stitching reference features in the image can be achieved.

[0053] S104. Based on the deployment location relationship of the image acquisition equipment, determine multiple adjacent viewpoint pairs consisting of effective viewpoints.

[0054] In one example, image acquisition devices are evenly deployed at key locations around the vehicle, including the center of the front bumper, the right front wheel arch, the center of the right door, the right rear wheel arch, the center of the rear bumper, the left rear wheel arch, the center of the left door, and the left front wheel arch, totaling eight deployment points, forming 360° coverage without blind spots. Each deployment point corresponds to a unique device identifier, numbered sequentially from 1 to 8 in a clockwise direction. This numbering order is the preset installation order, used to visually reflect the physical adjacency between the devices. Simultaneously, through the vehicle positioning module and device calibration data, the absolute installation coordinates (establishing a coordinate system with the vehicle's center of gravity as the origin) corresponding to each device identifier and the relative angles between adjacent devices (the angle between adjacent deployment points is 45°) are pre-stored, forming a complete deployment position relationship database to ensure the accuracy of adjacency determination.

[0055] As one possible implementation, based on the aforementioned pre-defined deployment location relationship, two perspectives that are physically adjacent and both are valid viewpoints are selected to form adjacent perspective pairs that are feasible for stitching, ensuring that the images of the paired perspectives have effective overlapping areas, thus providing a foundation for subsequent feature matching and image stitching.

[0056] In some embodiments, each valid viewing angle obtained from the previous screening is first associated with the installation coordinates, device number, and relative angle information in the deployment location relationship database through its corresponding image acquisition device identifier. This clarifies the specific physical location of each valid viewing angle around the vehicle and its corresponding preset installation sequence number, making the positional attributes of the valid viewing angle quantifiable and comparable. For example, if the viewing angle corresponding to the image acquisition device with identifier number 2 is determined to be a valid viewing angle, then its associated installation location is the right front wheel arch, numbered 2, with adjacent devices numbered 1 (front) and 3 (right), and the relative angle with the adjacent devices is 45°.

[0057] Secondly, based on the preset installation sequence number (numbers 1 to 8 arranged clockwise), the device numbers corresponding to all valid viewpoints are traversed. Their adjacency relationship can be simplified to sequential adjacency. Each viewpoint in the sequence is then paired with the next viewpoint (the last viewpoint and the first viewpoint) to form a candidate adjacent viewpoint pair.

[0058] Furthermore, the candidate adjacent viewpoint pairs obtained from the initial screening are further verified to ensure that their image acquisition ranges overlap, ensuring that the images from adjacent viewpoints have sufficient common areas for feature matching. Specifically, based on the pre-configured position and shooting angle of each image acquisition device, it can be determined whether the images acquired by each viewpoint in each candidate adjacent viewpoint pair overlap. If there is no overlap, or the overlap range is too small (e.g., the ratio of the number of pixels in the overlap range to the total number of pixels in the image is less than 5%), the candidate adjacent viewpoint pair is eliminated; otherwise, the candidate adjacent viewpoint pair is determined as the adjacent viewpoint pair for subsequent stitching and fusion.

[0059] In some embodiments, adjacent view pairs are assigned a unique view pair identifier, which is composed of the device numbers corresponding to the two valid view pairs (e.g., the valid view pairs composed of numbers 1 and 2 are identified as "1-2") and stored in the local database on the vehicle.

[0060] S105. For each adjacent viewpoint pair, extract multiple feature points from the feature maps of the two views included in the adjacent viewpoint pair, and generate multiple candidate stitching arrays based on the pixel coordinates and feature intensity values ​​of the feature points.

[0061] As one possible implementation, for each adjacent viewpoint pair, multiple feature points are extracted from the feature maps of the two views included in the adjacent viewpoint pair.

[0062] In some embodiments, for each feature map, pixels with feature intensity values ​​higher than a preset intensity threshold are identified as candidate feature points. Since the feature intensity value is a normalized value within the range of 0 to 1, the preset intensity threshold can be set to a fixed value, such as 0.7, to filter out feature points with high recognizability and strong stability. Furthermore, for each candidate feature point, if multiple candidate feature points exist within a preset range centered on the candidate feature point (e.g., a circular range with a radius of 5 pixels), the candidate feature point with the highest feature intensity value is retained as the feature point, and the remaining candidate feature points within this range are removed, ensuring that the extracted feature points are evenly distributed in the feature map.

[0063] After the filtering is completed, record the pixel coordinates and feature intensity value of each feature point. The pixel coordinates are the pixel coordinates of the feature point in the feature map, that is, the pixel coordinates of the feature point in the image.

[0064] Furthermore, for any two feature points in the feature map of each viewpoint in an adjacent viewpoint pair, the binding correlation degree between the two feature points is determined based on the continuity of the pixel gradient direction on the path connecting the two feature points and the difference between the feature intensity values ​​of the two feature points. The binding correlation degree is used to characterize the degree of fit between the two feature points as splicing reference points.

[0065] In some embodiments, the formula for calculating the association degree between two feature points is as follows: In the formula, For feature points With feature points The degree of association between them To connect feature points With feature points The average of the cosine similarity of the gradient directions of all adjacent pixels along the path. For feature points With feature points The ratio of the Euclidean distance (in pixels) between two points to the maximum Euclidean distance between feature points in the current feature map. This is the distance attenuation coefficient, which takes a value greater than zero and is used to control the attenuation rate. An empirical value of 3 is recommended. For feature points The corresponding feature intensity value, For feature points The corresponding feature intensity value, It represents the maximum value of the feature intensity of all feature points in the current feature map. It is the minimum feature intensity value of all feature points in the current feature map. It is a very small positive number, and an empirical value of 0.001 can be taken. It is a natural constant. , They are respectively Item and The weight coefficients corresponding to each item all have values ​​greater than zero, and The weighting coefficients can be determined based on prior knowledge or experimental data, and are used to adjust... Item and The relative importance of items in the final evaluation, for example, a possible empirical value is , ,or , .

[0066] The formula comprehensively evaluates the applicability of feature point pairs through a weighted sum. The first term evaluates the continuity of pixel gradient directions along the path connecting the two feature points, measured by gradient direction consistency. With distance decay function The product of these two terms shows that the value increases with increasing continuity and decreases smoothly with increasing normalized distance. The second term is feature similarity, obtained by subtracting the difference in normalized feature intensity values ​​from 1; its value increases as the difference in feature intensity decreases. Both terms are positive indicators, and their values ​​range from [0,1]. Therefore, combining the correlation The value range is also [0,1], achieving complete normalization. The larger the value, the more likely the feature point pair belongs to the same physical structure and has similar feature responses, making it more suitable as a candidate matching primitive to guide subsequent image stitching.

[0067] Based on this, the structural correlation between any two feature points can be obtained.

[0068] Furthermore, for each feature point in the feature map, a candidate splicing array is formed by combining the feature point and the feature points whose association degree with the feature point is higher than a preset association degree threshold. After traversing all feature points in the feature map, multiple candidate splicing arrays are obtained.

[0069] In some embodiments, since the correlation degree between feature points ranges from [0,1], the preset correlation degree threshold can be set to [0.7,0.9], for example, an empirical value of 0.8. Then, starting with each feature point, all other feature points within the same feature map are traversed. If the correlation degree between a feature point and the starting feature point is higher than the preset correlation degree threshold, then that feature point and the starting feature point are grouped into the same candidate splicing array. This process is repeated to traverse all feature points, thereby obtaining multiple candidate splicing arrays.

[0070] In some embodiments, the candidate splicing arrays obtained by division are verified. If the number of feature points contained in an array is too low (e.g., less than 3), it is determined to be an invalid array and is removed to avoid splicing matching failure due to insufficient number of feature points.

[0071] S106. Determine the target stitching array corresponding to adjacent viewpoint pairs from multiple candidate stitching arrays.

[0072] As one possible implementation, for a first candidate stitching array of the first view feature map and a second candidate stitching array of the second view feature map in adjacent view pairs, the stitching accuracy between the first candidate stitching array and the second candidate stitching array is determined. The first candidate stitching array is any candidate stitching array in the first view feature map, and the second candidate stitching array is any candidate stitching array in the second view feature map.

[0073] In some embodiments, the geometric center coordinates of the first candidate stitching array and the second candidate stitching array are first calculated. Based on the prior orientation relationship between the first and second viewpoints (e.g., if the first viewpoint is a right front angle viewpoint and the second viewpoint is a right side viewpoint, then the prior orientation is "the first viewpoint is in front of the second viewpoint"), preliminary matching is performed according to the proximity of the geometric center coordinates. The Euclidean distance between the geometric center coordinates of the first and second candidate stitching arrays is calculated. If the distance exceeds a preset distance threshold (based on the feature map size, such as 50 pixels), the second candidate stitching array is skipped, and the next second candidate stitching array in the second viewpoint feature map is traversed. If the distance is less than the preset distance threshold, the matching feature points in the first and second candidate stitching arrays are further determined.

[0074] Optionally, the pixel coordinates of feature points in the first candidate stitching array can be mapped to the second-view feature map, or the pixel coordinates of feature points in the second candidate stitching array can be mapped to the first-view feature map. Taking mapping the pixel coordinates of feature points in the second candidate stitching array to the first-view feature map as an example, for each feature point in the first candidate stitching array, the position coordinate deviation (such as Euclidean distance) between its position coordinates and the position coordinates of each feature point in the second candidate stitching array mapped to the first-view feature map is determined. Then, if the position coordinate deviation between the feature points in the first candidate stitching array and the feature points in the second candidate stitching array is less than a preset deviation threshold (such as 5 pixels), the feature point pair is determined to be a matching feature point. If the position coordinate deviation between a feature point and multiple feature points is less than the preset deviation threshold, the feature point pair corresponding to the smallest position coordinate deviation is determined to be a matching feature point. If there are still multiple feature point pairs whose position coordinate deviations are all the smallest position coordinate deviations, the gray value difference of each feature point pair is further determined, and the feature point pair with the smallest gray value difference is determined to be a matching feature point. After traversing all feature points in the first candidate splicing array, the number of feature points in any candidate splicing array that are determined to be matching feature points is recorded as the number of matching feature points.

[0075] Furthermore, the number of matching feature points, the number of first feature points contained in the first candidate splicing array, and the number of second feature points contained in the second candidate splicing array are obtained. Based on the number of matching feature points, the number of first feature points, and the number of second feature points, a first evaluation factor is determined. The mean of the binding correlation of matching feature points in the first and second candidate splicing arrays is obtained. The sum of the binding correlations of all feature points other than matching feature points in the first and second candidate splicing arrays is obtained. Based on the mean and the sum, a second evaluation factor is determined. Finally, the splicing accuracy is determined based on the first and second evaluation factors.

[0076] In some embodiments, the formula for calculating the stitching accuracy between the first candidate stitching array and the second candidate stitching array is as follows: In the formula, The first candidate splicing array in the first-view feature map Second candidate splicing array with the second view feature map The accuracy of splicing between array pairs is in the range of [0, +∞). The larger the value, the stronger the adaptability of the array pair and the higher the feature alignment accuracy after splicing. First candidate splicing array With the second candidate splicing array The number of corresponding matching feature points, First candidate splicing array The number of first feature points included. For the second candidate splicing array The number of second feature points included; First candidate splicing array With the second candidate splicing array The average binding correlation of all matching feature points is calculated by taking the average value of the binding correlation between each matching feature point and other matching feature points in the candidate splicing array. This parameter is used to characterize the internal adaptability of the matching feature points. First candidate splicing array With the second candidate splicing array The combined association degree of all feature points other than all matching feature points is calculated by summing the association degree between the feature points other than matching feature points in the two candidate splicing arrays. This parameter is used to characterize the degree of interference of non-matching regions on splicing. , These are the weight coefficients corresponding to the first and second evaluation factors, respectively, and their values ​​are all greater than zero. The weighting coefficients can be determined based on prior knowledge or experimental data, and are used to adjust the relative importance of the first and second evaluation factors in the final evaluation. For example, an empirical value can be taken as... , ,or , .

[0077] in, The first evaluation factor is used to characterize the spatial fitness of the array pair. The larger the numerator and the smaller the denominator, the more overlapping feature points the two arrays have, the closer their sizes are, and the stronger their spatial fitness. The second evaluation factor is used to characterize the guiding role of matching feature points on the overall splicing. The larger the numerator and the smaller the denominator, the stronger the internal adaptability of the matching feature points, the smaller the interference from non-matching regions, and the higher the splicing reliability.

[0078] Finally, based on the stitching accuracy, the target stitching array is determined from multiple candidate stitching arrays.

[0079] In some embodiments, based on the stitching accuracy corresponding to all matching first and second candidate stitching arrays, the first candidate stitching array with the highest stitching accuracy and the second candidate stitching array are selected as the target stitching array. If multiple stitching accuracies are the same and all are the highest values, the candidate stitching array pair with the most matching feature points is further selected as the target stitching array based on the number of matching feature points of each candidate stitching array pair. If the number of matching feature points is still the same, the arrays are compared. and Mean of the association degree Candidate stitching array pairs with larger mean values ​​are selected to ensure the uniqueness and optimality of the target stitching array.

[0080] S107. Based on the target stitching array corresponding to each adjacent viewpoint, the feature maps of all effective viewpoints are stitched together, and the stitched feature maps are upsampled to generate a panoramic image of the vehicle.

[0081] As one possible implementation, for each adjacent viewpoint pair, the matching feature points in the target stitching array corresponding to the adjacent viewpoint pair are aligned according to pixel coordinates.

[0082] In some embodiments, the array in the target stitched array is extracted. (First-person view feature map) and array All matching feature points in the (second-view feature map) are precisely aligned according to pixel coordinates—the array is then... The pixel coordinates of each matching feature point ( , ) and array The pixel coordinates of the corresponding matching feature points ( , Coordinate overlap processing is performed, and the average coordinates of the matching feature points are used as a reference to fine-tune the spatial position of the two feature maps so that the matching feature points completely overlap, forming a splicing reference anchor point to ensure the spatial consistency of cross-view features.

[0083] The total correlation coefficient of each candidate stitching array in the target stitching array is further determined, and the acquisition viewpoint corresponding to the candidate stitching array with the larger total correlation coefficient is determined as the primary viewpoint, and the other acquisition viewpoint is determined as the auxiliary viewpoint. The feature map of the primary viewpoint provides core stitching information, while the feature map of the auxiliary viewpoint supplements edge details. This determination logic is based on the principle that "the higher the total correlation coefficient, the stronger the adaptability of feature points within the array, and the higher the credibility of feature map information," ensuring that the stitching result is based on highly credible information.

[0084] Furthermore, based on the stitching accuracy of the target stitching array, the adjustment amount of the feature point positions for fusing the main view and the auxiliary view is determined, and the feature point positions in the main view and the auxiliary view are adjusted according to the position adjustment amount; then, based on the adjusted feature point positions, the feature maps of the main view and the auxiliary view are fused.

[0085] In some embodiments, for non-overlapping feature points in the main view feature map (denoted as...) The closest corresponding feature point in the auxiliary view feature map (denoted as ) ), calculate the straight-line distance between the two. (Based on Euclidean distance of pixel coordinates, (and based on the stitching accuracy of the target stitching array) Normalization is performed to obtain the normalized stitching accuracy. (Value range [0,1], normalization logic is) ,in The minimum stitching accuracy for all target stitching arrays. To achieve maximum splicing accuracy.

[0086] Then, based on the normalized stitching accuracy, the straight-line spacing is allocated to achieve feature point position adjustment and pixel fusion. The calculation formula for determining the feature point position adjustment amount used to fuse the main viewpoint and auxiliary viewpoint is as follows: In the formula, Main viewpoint feature points The distance to be moved is used to characterize the weight of the main viewpoint information in the stitching area. auxiliary viewpoint feature points The distance to be moved is used to characterize the weighting of auxiliary viewpoint information in the stitching area. Non-overlapping feature points from the main viewpoint Feature points corresponding to auxiliary view The linear spacing reflects the spatial distance between two points. The normalized stitching accuracy is used to quantify the impact of the fit of the target stitching array on the spacing allocation. The larger the value, the higher the weight of the main viewpoint information.

[0087] The calculated position adjustment amount ( , After that, feature points Along the straight line spacing Towards Directional movement Feature points Move in the opposite direction This allows for a smooth transition between the two points in the splicing area, and also affects the linear spacing. For all pixels within the range, the feature intensity values ​​are processed using a weighted fusion method: the final feature intensity value of a pixel = the feature intensity value of the pixel in the main viewpoint × Auxiliary viewpoint pixel feature intensity value × ( This ensures a natural transition in feature intensity across the spliced ​​areas, without obvious breaks. After fusion, the local spliced ​​feature map of the adjacent viewpoint pair is output.

[0088] After completing the local stitching of all adjacent viewpoint pairs, based on the preset installation order of the image acquisition equipment (clockwise 1-8), all local stitched feature maps are integrated into a global feature map to ensure 360° coverage of the vehicle's surrounding environment without blind spots.

[0089] Starting from the vehicle's frontal view (the effective view corresponding to device 1), the local stitching results of adjacent view pairs are sequentially integrated in a clockwise direction to form a global stitching sequence (such as local stitching image 1-2 → local stitching image 2-3 → … → local stitching image 8-1). This ensures that the stitching order is consistent with the physical correlation of the device deployment location and avoids global view confusion.

[0090] After integrating feature maps from all valid viewpoints, upsampling is required to obtain a high-resolution panoramic image suitable for display. This upsampling employs a preset interpolation algorithm (such as bilinear or bicubic interpolation). The core purpose is to infer and supplement pixel information in the high-resolution image based on the existing feature intensity distribution in the global feature map, enabling the panoramic image to clearly present key navigation information such as the outlines of obstacles around the vehicle, parking line details, and curb boundaries. The upsampling scaling ratio is based on the display resolution of the vehicle control screen (e.g., scaling a 1920×1080 global feature map to 3840×2160 to meet high-definition display requirements). During upsampling, high feature intensity regions in the global feature map (such as obstacle edges and parking lines) undergo edge enhancement processing. Edge contrast is strengthened by calculating pixel gradient values ​​to prevent blurring of details after upsampling. For low feature intensity background regions (such as flat ground), smooth interpolation is used to avoid introducing false features and ensure the image's authenticity.

[0091] After the upsampling process is completed, the high-resolution global feature map is converted into an RGB color panoramic image. Based on the mapping relationship between feature intensity values ​​and original image color information (the original image color data corresponding to each pixel is retained during the upsampling process), each pixel is assigned a corresponding RGB color value to form a visual image.

[0092] The generated panoramic image needs to be linked to the vehicle's real-time position and attitude data, and transmitted to the vehicle control screen for real-time display via the vehicle's CAN bus. Navigation information such as parking space boundary markers and obstacle distance indicators are overlaid on the image to provide the driver with intuitive parking orientation references. For image areas with non-effective viewing angles, which are considered locations with obstructions or unreliable information, when the vehicle moves in the direction of the non-effective viewing angle, it does not rely on visual images as a basis for movement, but instead relies on supplementary means such as ultrasonic radar and lidar to move in the corresponding direction of the non-effective viewing angle. Furthermore, the panoramic image data is stored in a local cache for real-time updates and retrospective analysis during subsequent parking processes, ensuring continuous and accurate vehicle-centric navigation.

[0093] Understandably, in the panoramic image generation method for assisting vehicle parking navigation provided in this embodiment of the invention, high-value images and inefficient images are accurately identified through content validity parameters, and effective viewpoints with sufficient stitching information are selected. This reduces the interference of invalid data on subsequent processes and lowers computational power consumption. Then, the core features of the image are enhanced through feature map transformation, and cross-view feature matching is achieved by constructing candidate stitching arrays and target stitching arrays, providing a reliable reference for stitching. Finally, the feature map is accurately stitched with the target stitching array as the core, and details are supplemented and resolution is improved through upsampling to ensure that the generated panoramic image can clearly and coherently present key navigation information such as obstacle distribution, parking space boundaries, and curbs in the 360° environment around the vehicle. The entire process follows an "end-to-end" processing logic, with all calculations revolving around high-quality data from effective viewpoints. This ensures the real-time generation of panoramic images while significantly improving image stitching accuracy and visual clarity. It provides intuitive and accurate environmental perception support for vehicle parking navigation, effectively reducing the operational difficulty caused by blind spots and inaccurate environmental judgment during parking, and enhancing the safety and convenience of parking.

[0094] Please see Figure 2 The diagram illustrates a panoramic image generation system for assisting vehicle parking navigation according to an embodiment of the present invention. Figure 2 As shown, the panoramic image generation system 20 for assisting vehicle parking navigation includes an image acquisition module 21, a content validity evaluation module 22, a viewpoint filtering module 23, a stitching array construction module 24, and a panoramic stitching module 25.

[0095] The image acquisition module 21 is used to acquire images acquired by multiple image acquisition devices deployed on the vehicle at the current moment. Its specific implementation can be referred to the description in step S101 of the above embodiment, and will not be repeated here.

[0096] The content validity assessment module 22 is used to determine the content validity parameters of each image based on the pixel value distribution information of the image. Its specific implementation can be referred to the description in step S102 in the above embodiment, and will not be repeated here.

[0097] The viewpoint filtering module 23 is used to filter out valid viewpoints from all acquired viewpoints based on the content validity parameter, and convert the image of each valid viewpoint into a corresponding feature map, where each pixel in the feature map has a feature intensity value; then, based on the deployment location relationship of the image acquisition devices, it determines multiple adjacent viewpoint pairs composed of valid viewpoints. Its specific implementation can be referred to in steps S103 and S104 of the above embodiments, and will not be repeated here.

[0098] The stitching array construction module 24 is used to extract multiple feature points from the feature maps of the two views included in each adjacent view pair, generate multiple candidate stitching arrays based on the pixel coordinates and feature intensity values ​​of the feature points, and determine the target stitching array corresponding to the adjacent view pair from the multiple candidate stitching arrays. Its specific implementation can be referred to in steps S105 and S106 of the above embodiments, and will not be repeated here.

[0099] The panoramic stitching module 25 is used to stitch the feature maps of all effective viewpoints based on the corresponding target stitching array for each adjacent viewpoint, and to perform upsampling processing on the stitched feature maps to generate a panoramic image of the vehicle. Its specific implementation can be referred to in step S107 of the above embodiments, and will not be repeated here.

[0100] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0101] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for generating a panoramic image for assisting vehicle parking navigation, characterized in that, The method includes: Acquire images captured at the current moment by multiple image acquisition devices deployed on the vehicle, with each image acquisition device corresponding to a capture viewpoint; For each image, a content validity parameter is determined based on the pixel value distribution information of the image. The content validity parameter is used to characterize the richness and salience of the information used for panoramic stitching in the image. Based on the content validity parameters, valid viewpoints are selected from all captured viewpoints, and the image of each valid viewpoint is converted into a corresponding feature map, where each pixel in the feature map has a feature intensity value. Based on the deployment location relationship of the image acquisition devices, determine multiple adjacent viewpoint pairs composed of the effective viewpoints; For each adjacent viewpoint pair, multiple feature points are extracted from the feature maps of the two viewpoints included in the adjacent viewpoint pair. Multiple candidate stitching arrays are generated based on the pixel coordinates and feature intensity values ​​of the feature points. The target stitching array corresponding to the adjacent viewpoint pair is determined from the multiple candidate stitching arrays. Based on the target stitching array corresponding to each adjacent viewpoint, the feature maps of all effective viewpoints are stitched together, and the stitched feature maps are upsampled to generate a panoramic image of the vehicle.

2. The method for generating a panoramic image for assisting vehicle parking navigation according to claim 1, wherein, The content validity parameters of the image are determined based on the pixel value distribution information of the image, including: Generate a histogram of pixel value distribution for the image; Based on the pixel value distribution histogram, the pixel value concentration parameter and pixel value difference parameter of the image are determined. The pixel value concentration parameter is used to characterize the degree of concentration of pixel value distribution in the image, and the pixel value difference parameter is used to characterize the overall degree of difference between different pixel value regions in the image. The content validity parameter is determined based on the pixel value concentration parameter and the pixel value difference parameter.

3. The method for generating a panoramic image to assist in navigating a vehicle into a parking space according to claim 2, wherein, Determining the pixel value concentration parameter and pixel value difference parameter of the image includes: From the pixel value distribution histogram, the target pixel value with the largest number of pixels is determined, and the ratio of the number of target pixels with the pixel value of the target pixel value to the total number of pixels in the image is determined as the pixel value concentration parameter of the image. For each target pixel, determine the pixel value difference between the target pixel and its neighboring pixels; The pixel value difference parameter is determined based on the statistical characteristics of the pixel value differences corresponding to all target pixels.

4. The method for generating a panoramic image to assist in navigating a vehicle into a parking space according to claim 1, wherein, Based on the content validity parameters, valid viewpoints are selected from all collected viewpoints, including: All collected perspectives are sorted in descending order according to their corresponding content validity parameters to form a validity sequence, and the interval between adjacent content validity parameters in the validity sequence is determined. Determine the target interval with the largest value from all calculated intervals; Based on the position of the target interval in the validity sequence, the validity sequence is divided into a first subsequence and a second subsequence, wherein the validity parameter of any content included in the first subsequence is higher than the validity parameter of any content in the second subsequence; The collection perspective corresponding to the validity parameter of the content contained in the first subsequence is determined as the valid perspective.

5. The method for generating a panoramic image to assist in vehicle parking navigation according to claim 1, wherein, Multiple candidate stitching arrays are generated based on the pixel coordinates and feature intensity values ​​of feature points, including: For any two feature points in the feature map of each viewpoint in the adjacent viewpoint pair, the binding correlation degree between the two feature points is determined based on the continuity of the pixel gradient direction on the path connecting the two feature points and the difference between the feature intensity values ​​of the two feature points. The binding correlation degree is used to characterize the degree of fit between the two feature points as splicing reference points; the binding correlation degree is expressed by the formula... Sure, For feature points With feature points The degree of association between them To connect feature points With feature points The average of the cosine similarity of the gradient directions of all adjacent pixels along the path. For feature points With feature points The ratio of the Euclidean distance between points to the maximum Euclidean distance between feature points in the current feature map. This is the distance attenuation coefficient, which takes a value greater than zero. For feature points The corresponding feature intensity value, For feature points The corresponding feature intensity value, It represents the maximum value of the feature intensity of all feature points in the current feature map. It is the minimum feature intensity value of all feature points in the current feature map. It is a very small positive number, used to prevent the denominator from being zero. , These are the weighting coefficients, and , To ensure the continuity of pixel gradient directions on the path connecting the two feature points; For each feature point in the feature map, the feature point and the feature points whose association degree with the feature point is higher than a preset association degree threshold are formed into a candidate splicing array. After traversing all feature points in the feature map, multiple candidate splicing arrays are obtained.

6. The panoramic image generation method for assisting vehicle parking navigation according to claim 5, characterized in that, Determining the target stitching array corresponding to the adjacent viewpoint pair from the plurality of candidate stitching arrays includes: For the first candidate stitching array of the first view feature map and the second candidate stitching array of the second view feature map in the adjacent view pair, determine the stitching accuracy between the first candidate stitching array and the second candidate stitching array; Based on the splicing accuracy, a target splicing array is determined from the plurality of candidate splicing arrays.

7. The panoramic image generation method for assisting vehicle parking navigation according to claim 6, characterized in that, Determining the stitching accuracy between the first candidate stitching array and the second candidate stitching array includes: Determine the matching feature points in the first candidate stitching array and the second candidate stitching array based on pixel coordinate matching; Obtain the number of matching feature points, the number of first feature points contained in the first candidate splicing array, and the number of second feature points contained in the second candidate splicing array; The first evaluation factor is determined based on the number of matched feature points, the number of the first feature points, and the number of the second feature points; Obtain the mean value of the combination correlation degree of the matching feature points in the first candidate splicing array and the second candidate splicing array; Obtain the sum of the combination correlation degrees corresponding to all feature points other than the matching feature points in the first candidate splicing array and the second candidate splicing array; The second evaluation factor is determined based on the mean and the sum. The splicing accuracy is determined based on the first evaluation factor and the second evaluation factor, and the splicing accuracy is determined by the formula... Sure, The first candidate splicing array in the first-view feature map Second candidate splicing array with the second view feature map The accuracy of splicing between them The first evaluation factor, First candidate splicing array With the second candidate splicing array The number of corresponding matching feature points, First candidate splicing array The number of first feature points included. For the second candidate splicing array The number of second feature points included. This is the second evaluation factor. First candidate splicing array With the second candidate splicing array The mean associativity of all matching feature points in the dataset. First candidate splicing array With the second candidate splicing array The sum of the associativity of all feature points other than all matching feature points. , These are the weighting coefficients, and .

8. The panoramic image generation method for assisting vehicle parking navigation according to claim 7, characterized in that, Based on the target stitching array corresponding to each of the adjacent viewpoints, the feature maps of all effective viewpoints are stitched together, including: For each of the adjacent viewpoint pairs, the matching feature points in the target stitching array corresponding to the adjacent viewpoint pairs are aligned according to pixel coordinates; The sum of the combination correlation of each candidate splicing array in the target splicing array is determined, and the acquisition view corresponding to the candidate splicing array with the larger sum is determined as the main view and the other acquisition view is determined as the auxiliary view. Based on the stitching accuracy of the target stitching array, determine the feature point position adjustment amount used to fuse the main viewpoint and the auxiliary viewpoint; Based on the position adjustment amount, the positions of feature points in the main view and the auxiliary view are adjusted respectively; Based on the adjusted feature point positions, the feature maps of the main viewpoint and the auxiliary viewpoint are fused.

9. The panoramic image generation method for assisting vehicle parking navigation according to claim 7, characterized in that, Determining matching feature points based on pixel coordinate matching in the first candidate stitching array and the second candidate stitching array includes: Obtain the geometric center coordinates of the first candidate splicing array and the second candidate splicing array; Based on the prior orientation relationship of the adjacent viewpoint pairs, and according to the proximity of the geometric center coordinates, a matching is performed between the candidate stitching array of the first viewpoint feature map and the candidate stitching array of the second viewpoint feature map. In the successfully matched candidate splicing array pairs, feature points whose position coordinate deviation is less than a preset deviation threshold are identified as the matched feature points.

10. The panoramic image generation method for assisting vehicle parking navigation according to claim 1, characterized in that, Multiple feature points are extracted from the feature maps of the two viewpoints included in the adjacent viewpoint pair, including: For each feature map, pixels in the feature map whose feature intensity value is higher than a preset intensity threshold are identified as candidate feature points; For each candidate feature point, if there are multiple candidate feature points within a preset range centered on the candidate feature point, the candidate feature point with the largest feature intensity value is retained as the feature point.

Images