A large-parallax full-scene image stitching method and system for an annular camera array

Abstract

This invention discloses a method and system for high parallax full-scene image stitching using a circular camera array. Addressing the issues of ghosting and structural discontinuities caused by parallax due to changes in camera baseline and scene depth when acquiring images in complex 3D and dynamic scenes using multiple cameras arranged in a circular array, this invention proposes an image reconstruction and stitching technology solution for online full-scene applications. Based on a unified geometric organization of the circular camera array, this method, without relying on depth estimation and 3D reconstruction, constructs the local projection domain of each camera image in panoramic space and establishes a continuous confidence weight distribution at the pixel level for different image sources, thereby suppressing geometric inconsistencies caused by parallax. This method supports algorithm-level parallel computation, enabling rapid online updates in dynamic scenes, and can be widely applied to mobile robots, environmental perception, full-scene visual reconstruction, monitoring systems, and other multi-camera vision applications.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to a method and system for stitching large parallax full-scene images using a ring camera array. Background Technology

[0002] With the development of computer vision technology, multi-camera image stitching technology has been widely applied in fields such as environmental perception, panoramic vision construction, and monitoring systems for mobile robots. By acquiring images from different perspectives using multiple cameras and stitching them together, the limitation of the field of view of a single camera can be overcome to some extent, thereby obtaining a wider range of visual information. However, in practical applications, especially in scenarios with significant depth changes and dynamic targets, multi-camera image stitching still faces geometric inconsistencies caused by parallax.

[0003] Existing multi-camera image stitching methods are mainly based on two-dimensional geometric transformation models. They calculate the homography matrix between images through feature extraction and matching, and then use this homography matrix to achieve image alignment and fusion. Common methods include matching methods based on human-generated features such as SIFT and FAST, deep feature matching methods such as SuperGlue, and methods like UDIS++ that directly regress the homography matrix using deep learning. The common goal of these methods is to obtain the most accurate homography matrix possible; some methods even rely on long-term data training or meticulous calibration processes to achieve higher precision in geometric mapping relationships.

[0004] However, the methods described above generally implicitly assume that the scene is approximately coplanar or has small parallax. Under multi-camera circular arrangement, due to the large camera baseline and complex scene depth distribution, even with a relatively accurate homography matrix, image distortion, ghosting, or structural breaks are unavoidable in near-field objects or areas of abrupt depth changes. This parallax-induced error cannot be fundamentally eliminated by further improving the homography matrix accuracy; instead, it becomes more prominent with increasing calibration complexity and system cost.

[0005] Furthermore, existing methods typically require high camera calibration accuracy and a suitable system deployment environment, making them difficult to adapt to the needs of rapid deployment and dynamic adjustments in applications. Their versatility and practicality across different platforms and scenarios are limited. Therefore, effectively suppressing the parallax effect under multi-camera circular arrangement conditions and achieving easily configurable, dynamic, full-scene image reconstruction and stitching without requiring high-precision calibration and complex 3D reconstruction has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of existing multi-camera image stitching methods, which generally rely on accurate homography matrix estimation and high-precision camera calibration, and still inevitably produce image distortion, ghosting and structural discontinuity in scenes with large parallax and complex depth changes. This invention provides a large parallax full-scene image stitching method and system for a ring camera array, which can achieve fully online stable stitching in dynamic scenes without the need for precise calibration.

[0007] The objective of this invention is achieved through the following technical solution: a method for stitching large parallax full-scene images using a circular camera array, the method comprising:

[0008] Acquire multi-camera image data and preprocess it. Construct a homography matrix for each camera with rotation components removed to obtain a simplified homography matrix.

[0009] The normalized corner coordinates of the projected image are calculated using a simplified homography matrix.

[0010] Based on the normalized corner coordinates, calculate the projection area boundary of the image on the panoramic canvas and construct the translation matrix. Calculate the local homography matrix and use the local homography matrix to transform the image to obtain the transformed image.

[0011] A basic weight matrix is ​​constructed, a perspective transformation is performed using a local homography matrix, and the transformed image is weighted and fused using the weight matrix after perspective transformation to achieve content consistency. The fusion is performed iteratively for each camera to obtain a complete panoramic image containing information from all cameras.

[0012] Furthermore, the construction of the simplified homography matrix includes: removing the rotation component of the homography matrix, retaining only the translation transformation of the camera relative to the panoramic canvas in the horizontal and vertical directions, for the first... For each camera, a simplified homography matrix is ​​constructed as a 3x3 matrix, with the first row being... The second line The third line ,in For camera Relative to the horizontal translation of the panoramic canvas, For camera The amount of vertical translation relative to the panoramic canvas.

[0013] Furthermore, the step of calculating the normalized corner coordinates after image projection using the simplified homography matrix includes:

[0014] Construct a homogeneous coordinate matrix for the corner points of the image, apply the simplified homography matrix to the corner point coordinate matrix to obtain the projected corner point coordinates, and normalize the projected corner point coordinates to obtain the normalized corner point coordinates after image projection.

[0015] Furthermore, the calculation of the projection region boundary of the image on the panoramic canvas and the construction of the translation matrix include:

[0016] Construct the translation matrix as a 3x3 matrix, with the first row being... The second line The third line ,in and These are the minimum boundary coordinates of the projected region;

[0017] The local homography matrix is ​​equal to the product of the translation matrix and the simplified homography matrix.

[0018] Furthermore, the construction of the basic weight matrix includes:

[0019] Generate one-dimensional arrays in the row and column directions: forming a weight distribution from the edge to the center and back to the edge. The weight values ​​increase linearly from 0 to 1 from the starting edge to the center, and decrease linearly from 1 to 0 from the center to the ending edge.

[0020] The exponential method is used to perform an outer product operation on the row-direction weight array and the column-direction weight array and then perform an exponential operation to obtain the spatial basic weight matrix. The exponent is used in the exponential operation to control the steepness of the weight distribution. A larger exponent value makes the weights in the central region of the image more concentrated and the weights in the edge region decay rapidly.

[0021] Furthermore, the weighted fusion of the transformed image using the weight matrix after perspective transformation to ensure content consistency includes:

[0022] Based on existing panoramic images and images after the newly introduced local homography transformation, the pixel differences between the two images are converted into weights using the Gaussian kernel function to obtain content consistency weight factors.

[0023] The weights of the newly introduced image are calculated by multiplying the weight matrix after perspective transformation with the content consistency weights.

[0024] The normalized weights for this fusion are dynamically calculated by using the global cumulative weights maintained for the panoramic image as the numerator and the sum of the global cumulative weights and the weights of the newly introduced image as the denominator.

[0025] Based on the normalized weights, the newly introduced image is superimposed onto the current global panoramic image to obtain the updated panoramic image; after the current image fusion is completed, the global cumulative weight matrix is ​​updated according to the weights of the newly introduced image.

[0026] Furthermore, the multi-camera system adopts a ring array configuration, with six cameras symmetrically distributed in a regular hexagon. The optical axis of each camera radiates outward from the center of the array, and the angle between adjacent optical axes is 60 degrees.

[0027] According to another aspect of the specification, a system for implementing the parallax full-scene image reconstruction and stitching method of the circular camera array is also provided. The system includes an image acquisition unit, an image preprocessing unit, a homography matrix construction unit, a local transformation unit, a weight generation unit, and an image fusion unit connected in sequence.

[0028] The image acquisition unit is used to acquire several real-time image data captured by multiple camera devices;

[0029] The image preprocessing unit is used to preprocess several real-time image data to obtain several preprocessed real-time image data.

[0030] The homography matrix construction unit is used to construct a simplified homography matrix that contains only translation components;

[0031] The local transformation unit is used to calculate the local homography matrix and transform the image;

[0032] The weight generation unit is used to generate a two-dimensional weight matrix;

[0033] The image fusion unit is used to perform content-consistency-based weighted fusion to obtain a complete panoramic image.

[0034] According to another aspect of the specification, a large parallax full-scene image stitching device for a ring camera array is also provided, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the large parallax full-scene image stitching method for a ring camera array.

[0035] According to another aspect of the specification, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, implements the large parallax full-scene image stitching method of a ring camera array.

[0036] The beneficial effects of this invention are:

[0037] By constraining the degrees of freedom of the general homography model, the impact of large parallax and complex depth variations on the geometric consistency of splicing is reduced, effectively avoiding global structural deformation.

[0038] Stable stitching is achieved without the need for high-precision camera calibration and 3D reconstruction, significantly reducing the complexity of system deployment and use;

[0039] By using a pixel-level weighted fusion mechanism, unavoidable projection errors are transformed into a smooth transition process, reducing ghosting and discontinuity.

[0040] Suitable for fully online processing in dynamic scenes, meeting the requirements of real-time vision systems for continuity and stability;

[0041] The camera array configuration is universal, does not depend on a specific platform, and has good application scalability and engineering practicality. Attached Figure Description

[0042] Figure 1 A flowchart illustrating the method provided in this embodiment of the invention;

[0043] Figure 2 The hardware configuration of the ring-shaped multi-camera array provided in the embodiments of the present invention;

[0044] Figure 3 This is a schematic diagram of continuous weighted fusion based on content consistency provided in an embodiment of the present invention;

[0045] Figure 4 This is a scene stitching diagram provided in an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of the apparatus provided in an embodiment of the present invention. Detailed Implementation

[0047] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0048] like Figure 1 As shown, this invention employs a ring-shaped multi-camera array configuration. Six cameras are symmetrically distributed in a regular hexagon on the top of the mobile robot platform. The optical axis of each camera radiates outward from the center of the array, with an angle of 60 degrees between adjacent optical axes. The camera spacing is 13.5 cm, a spacing that balances baseline length and parallax effect, being small enough to avoid excessive parallax from nearby objects, yet large enough to cover information across the entire scene. All cameras use a unified intrinsic parameter and distortion model, and the precise extrinsic parameters of each camera relative to the array center are obtained through offline calibration.

[0049] like Figure 2 As shown, the overall process of the method of the present invention includes the following steps. First, several real-time image data captured by multiple camera devices are acquired. Then, a simplified homography matrix is ​​constructed. Next, local homography transformation is performed. Then, a weight matrix is ​​generated. Finally, weighted fusion is performed to obtain the panoramic output.

[0050] The method of the present invention will be described in detail below.

[0051] Step 1: Image Input and Preprocessing

[0052] Collect several real-time image data points captured by multiple camera devices, and record the first... Each camera at any time The captured image is ,in For camera indexing. Image preprocessing is performed on several real-time image data sets to obtain several preprocessed real-time image data sets. The image preprocessing includes Gaussian denoising, image enhancement, and size normalization.

[0053] Step 2: Simplify the construction of the homography matrix

[0054] like Figure 3 As shown, a traditional complete homography matrix contains multiple transformation components such as rotation, translation, and perspective, which can lead to image distortion in scenes with large parallax. This invention proposes a simplified homography matrix, removing the rotation component and retaining only the translation transformation, thereby avoiding image distortion caused by rotation and maintaining the geometric integrity of the image.

[0055] 2.1 For the first For each camera, an improved simplified homography matrix is ​​constructed. for

[0056]

[0057] In the formula, —For the camera The amount of horizontal translation relative to the panoramic canvas;

[0058] — This represents the vertical translation of camera k relative to the panoramic canvas;

[0059] 2.2 For the six cameras, construct the corresponding simplified homography matrices. Taking cameras 0 to 5 as examples, the translation parameters in the simplified homography matrices of each camera are determined based on the offline calibration results.

[0060] Step 3, Local homography transformation

[0061] 3.1 Calculate the homogeneous coordinates of the image corners. Assume the image size is the width. and height Construct the homogeneous coordinate matrix of image corner points for

[0062]

[0063] 3.2 Calculate the coordinates of the projected corner points. The simplified homography matrix will be used. Applying this to the corner coordinate matrix yields the projected corner coordinates.

[0064]

[0065] 3.3 Normalize the projected corner coordinates to obtain normalized corner coordinates.

[0066]

[0067] In the formula, —For the first projection corner points Coordinate components;

[0068] — For the first projection corner points Coordinate components;

[0069] — For the first projection Homogeneous coordinate components of each corner point;

[0070] 3.4 Calculate the boundary of the projected area of ​​the image on the panoramic canvas based on the normalized corner coordinates. Let the size of the panoramic canvas be its width. and height Calculate the coordinates of the minimum bounding rectangle of the projected region. , , , .

[0071] 3.5 Constructing the translation matrix Used for local coordinate transformation

[0072]

[0073] 3.6 Calculate the local homography matrix

[0074]

[0075] 3.7 The image is transformed using the local homography matrix to obtain the transformed image. .

[0076] After the aforementioned local homography transformation, although image distortion is avoided, geometric inconsistencies and ghosting artifacts caused by parallax still exist in the overlapping areas of adjacent camera images. To eliminate these artifacts, weighted fusion processing is required.

[0077] Step 4: Weight Matrix Generation

[0078] 4.1 Generate a one-dimensional weight array. For a length of... A one-dimensional array is used to generate a weight distribution from the edge to the center and back to the edge, with weight values ​​increasing from 0 to 1 and then decreasing to 0. Let the index be... One-dimensional weights The calculation method is as follows: when the index When the array length is less than half, the weight value increases linearly from 0 to 1; when the index... When the weight value is greater than or equal to half the array length, the weight value decreases linearly from 1 to 0.

[0079] 4.2 Generating the Two-Dimensional Spatial Fundamental Weight Matrix. Using the exponential method, the row-direction weight array and the column-direction weight array are outer-producted and then raised to the power of the product to obtain the spatial fundamental weight matrix. The calculation formula is as follows:

[0080]

[0081] In the formula, —A one-dimensional weight array in the horizontal direction;

[0082] — A one-dimensional weight array in the vertical direction;

[0083] — For exponential parameters;

[0084] Exponential parameters Its function is to control the steepness of the weight distribution. A larger exponent value makes the weights more concentrated in the central region of the image, while the weights in the edge regions decay rapidly, thus achieving a smoother transition effect.

[0085] 4.3 Local coordinate transformation of spatial weights. This spatial fundamental weight matrix... This is used to characterize the spatial reliability of pixels within a single camera image. Before applying it to the panoramic canvas, the local homography matrix obtained in section 3.6 is used. Performing the same perspective transformation on it yields the transformed spatial weight matrix. So that in step five, the transformed image can be compared with the original image. Simultaneously participate in converged computing.

[0086] Step 5: Weighted fusion based on content consistency

[0087] This invention employs an iterative weighted fusion strategy. Assume that in the panoramic coordinate system, the system globally maintains an existing panoramic image. and the corresponding global cumulative weight (During system initialization, the pixel values ​​of both are set to 0). For the newly introduced... Image after camera transformation and their corresponding transformed spatial weights (As obtained from 4.3), perform a single fusion iteration according to the following sub-steps:

[0088] 5.1 Calculate the pixel-level content consistency weight factor. For any pixel coordinate on the panoramic image... Assume there is an existing panoramic image. The image after the newly introduced local homography transformation is A content consistency weighting factor is constructed based on pixel content differences. The calculation formula is as follows:

[0089]

[0090] In the formula, —Parameters for controlling the weighted response scale;

[0091] — The Euclidean distance between the new image and the reference image at this pixel.

[0092] 5.2 Calculate the normalized fusion weights. Combine this with the global cumulative weights. Spatial weights of the current new image and content consistency weighting factor Calculate the normalized weights for this fusion. :

[0093]

[0094] In the formula, —A small constant to prevent division by zero.

[0095] At this point, if it is the first fusion ( ), due to molecules Calculated It is always 0.

[0096] 5.3 Perform weighted fusion operation. Based on the normalized weights , will new image Overlay onto the current global panoramic image In the process, the updated panoramic image is obtained. :

[0097]

[0098] Then Overwrite the value to the global panoramic image In the middle. During the first fusion, due to New image It will be written to the panoramic canvas with a weight of 1 without loss.

[0099] 5.4 Update the global cumulative weight matrix. After completing the current image fusion, the global cumulative weight matrix must be updated according to the new weights introduced this time, for the next image (the first image). The fusion of multiple cameras provides a historical reference. The update formula is:

[0100]

[0101] The calculated Overwrite to global maintenance Complete one iteration and wait for the next image input from the camera.

[0102] Step 6: Iteratively fuse all camera images

[0103] Steps two through five are performed sequentially on the images from the six cameras, from camera 0 to camera 5. After processing each camera image, it is fused into the current panoramic image, and the corresponding accumulated weights are updated. After six iterations of fusion, a complete panoramic image containing information from all cameras is obtained. .

[0104] like Figure 4 The image shown is a scene stitching effect obtained using the method of this invention. It can be seen that multiple camera images are seamlessly stitched into a complete panoramic image, with smooth and natural transitions between adjacent images, without obvious stitching seams or ghosting. The image as a whole maintains good geometric consistency, without distortion caused by rotation transformations. Overlapping areas achieve smooth transitions through weighted fusion, effectively eliminating artifacts caused by parallax.

[0105] The method of this invention avoids image distortion by simplifying the homography matrix, reduces global error accumulation by constructing a local projection domain, achieves smooth transition of overlapping areas by exponential weighted fusion, and effectively eliminates artifacts caused by parallax by dynamic weight adjustment based on content consistency, ultimately obtaining a high-quality panoramic stitched image.

[0106] Corresponding to the aforementioned embodiment of a large parallax full-scene image stitching method for a ring camera array, the present invention also provides an embodiment of a large parallax full-scene image stitching device for a ring camera array.

[0107] See Figure 5 The present invention provides a large parallax full-scene image stitching device for a ring camera array, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a large parallax full-scene image stitching method for a ring camera array as described in the above embodiment.

[0108] The embodiment of the large parallax full-scene image stitching device for a ring camera array provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is an example of a large parallax full-scene image stitching device for a ring camera array provided by the present invention. (Except for...) Figure 5 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0109] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0110] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0111] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a large parallax full-scene image stitching method for a ring camera array as described in the above embodiments.

[0112] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0113] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the large parallax full-scene image stitching method of a ring camera array.

[0114] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0115] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for stitching large parallax full-scene images using a circular camera array, characterized in that, The method includes: Acquire multi-camera image data and preprocess it. Construct a homography matrix for each camera with rotation components removed to obtain a simplified homography matrix. The normalized corner coordinates of the projected image are calculated using a simplified homography matrix. Based on the normalized corner coordinates, calculate the projection area boundary of the image on the panoramic canvas and construct the translation matrix. Calculate the local homography matrix and use the local homography matrix to transform the image to obtain the transformed image. A basic weight matrix is ​​constructed, a perspective transformation is performed using a local homography matrix, and the transformed image is weighted and fused using the weight matrix after perspective transformation to achieve content consistency. The fusion is performed iteratively for each camera to obtain a complete panoramic image containing information from all cameras.

2. The method for large parallax full-scene image stitching using a circular camera array according to claim 1, characterized in that, The construction of the simplified homography matrix includes: removing the rotation component of the homography matrix, retaining only the translation transformation of the camera relative to the panoramic canvas in the horizontal and vertical directions, for the first... For each camera, a simplified homography matrix is ​​constructed as a 3x3 matrix, with the first row being... The second line The third line ,in For camera Relative to the horizontal translation of the panoramic canvas, For camera The amount of vertical translation relative to the panoramic canvas.

3. The method for large parallax full-scene image stitching using a circular camera array according to claim 1, characterized in that, The calculation of normalized corner coordinates after image projection using the simplified homography matrix includes: Construct a homogeneous coordinate matrix for the corner points of the image, apply the simplified homography matrix to the corner point coordinate matrix to obtain the projected corner point coordinates, and normalize the projected corner point coordinates to obtain the normalized corner point coordinates after image projection.

4. The method for large parallax full-scene image stitching using a circular camera array according to claim 1, characterized in that, The calculation of the projection region boundary of the image on the panoramic canvas and the construction of the translation matrix include: Construct the translation matrix as a 3x3 matrix, with the first row being... The second line The third line ,in and These are the minimum boundary coordinates of the projected region; The local homography matrix is ​​equal to the product of the translation matrix and the simplified homography matrix.

5. The method for large parallax full-scene image stitching using a circular camera array according to claim 1, characterized in that, The construction of the basic weight matrix includes: Generate one-dimensional arrays in the row and column directions: forming a weight distribution from the edge to the center and back to the edge. The weight values ​​increase linearly from 0 to 1 from the starting edge to the center, and decrease linearly from 1 to 0 from the center to the ending edge. The exponential method is used to perform an outer product operation on the row-direction weight array and the column-direction weight array and then perform an exponential operation to obtain the spatial basic weight matrix. The exponent is used in the exponential operation to control the steepness of the weight distribution. A larger exponent value makes the weights in the central region of the image more concentrated and the weights in the edge region decay rapidly.

6. The method for large parallax full-scene image stitching using a circular camera array according to claim 1, characterized in that, The weighted fusion of the transformed image using the weight matrix after perspective transformation for content consistency includes: Based on existing panoramic images and images after the newly introduced local homography transformation, the pixel differences between the two images are converted into weights using the Gaussian kernel function to obtain content consistency weight factors. The weights of the newly introduced image are calculated by multiplying the weight matrix after perspective transformation with the content consistency weights. The normalized weights for this fusion are dynamically calculated by using the global cumulative weights maintained for the panoramic image as the numerator and the sum of the global cumulative weights and the weights of the newly introduced image as the denominator. Based on the normalized weights, the newly introduced image is superimposed onto the current global panoramic image to obtain the updated panoramic image; after the current image fusion is completed, the global cumulative weight matrix is ​​updated according to the weights of the newly introduced image.

7. The method for large parallax full-scene image stitching using a circular camera array according to claim 1, characterized in that, The multi-camera system is configured in a ring array, with six cameras symmetrically distributed in a regular hexagon. The optical axis of each camera radiates outward from the center of the array, and the angle between adjacent optical axes is 60 degrees.

8. A system for implementing the parallax full-scene image reconstruction and stitching method of a circular camera array as described in any one of claims 1 to 7, characterized in that, The system includes an image acquisition unit, an image preprocessing unit, a homography matrix construction unit, a local transformation unit, a weight generation unit, and an image fusion unit connected in sequence. The image acquisition unit is used to acquire several real-time image data captured by multiple camera devices; The image preprocessing unit is used to preprocess several real-time image data to obtain several preprocessed real-time image data. The homography matrix construction unit is used to construct a simplified homography matrix that contains only translation components; The local transformation unit is used to calculate the local homography matrix and transform the image; The weight generation unit is used to generate a two-dimensional weight matrix; The image fusion unit is used to perform content-consistency-based weighted fusion to obtain a complete panoramic image.

9. A large parallax full-scene image stitching device for a circular camera array, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a large parallax full-scene image stitching method for a ring camera array as described in any one of claims 1-7.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a large parallax full-scene image stitching method for a ring camera array as described in any one of claims 1-7.

Images