Binocular image distortion automatic correction method and system

Abstract

The invention relates to the technical field of general image data processing, in particular to a binocular image distortion automatic correction method and system, and the method comprises the steps: obtaining a synchronous binocular image sequence; establishing a cross-frame and cross-view feature point space-time corresponding relation; constructing a unified distortion model; based on the spatial corresponding feature point pair and the time corresponding feature point pair, jointly quantifying a geometric consistency error and a luminosity consistency error; iteratively optimizing the unified distortion model parameters according to the space-time consistency error; and according to the optimized unified distortion model parameters, correcting the synchronous binocular image sequence in real time, and generating an undistorted output binocular image sequence. The method does not need a calibration board or manual intervention, can dynamically compensate distortion caused by environment change, and improves the geometric and luminosity consistency of a binocular image.

Description

Technical Field

[0001] This invention relates to the field of general image data processing technology, and in particular to a method and system for automatic correction of binocular image distortion. Background Technology

[0002] Binocular images serve as the fundamental data source for 3D reconstruction and depth estimation, and their imaging quality directly determines the accuracy and robustness of subsequent visual tasks. However, due to the inherent physical characteristics of optical lenses, binocular cameras equipped with optical lenses inevitably introduce geometric distortions during actual imaging, commonly including radial distortion, tangential distortion, and non-coplanar deviation between the left and right images.

[0003] Existing methods for correcting geometric distortion are mostly based on static calibration parameters. When the camera is subjected to dynamic disturbances such as temperature changes, mechanical vibrations, or structural minor changes caused by long-term use, the preset parameters are difficult to accurately reflect the current imaging state, resulting in the accumulation of residual correction errors. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic correction method and system for binocular image distortion, in order to solve the technical problems of existing binocular vision relying on offline calibration, being unable to adapt in real time to parameter drift and geometric distortion caused by changes in the physical environment.

[0005] One aspect of the present invention provides an automatic binocular image distortion correction method, comprising: Acquire a synchronized stereo image sequence, which consists of a left camera image sequence and a right camera image sequence; Establishing a spatiotemporal correspondence between feature points across frames and viewpoints involves detecting feature points in the current frame of the left camera image sequence and tracking the feature points in the current frame of the right camera image sequence and the next frame of the left camera image sequence, thereby obtaining spatially corresponding feature point pairs and temporally corresponding feature point pairs. A unified distortion model is constructed, which is characterized by a set of unified distortion parameters. The unified distortion parameter set includes the internal parameter matrix elements of the left and right cameras, radial and tangential distortion coefficients, rotation matrix and translation vector elements between the cameras, as well as vignetting correction coefficients and color transformation matrix elements used to describe luminance changes. Based on the spatially corresponding feature point pairs and the temporally corresponding feature point pairs, the geometric consistency error and the photometric consistency error are jointly quantized; the geometric consistency error is determined by calculating the deviation between the spatially corresponding feature point pairs and the geometric constraints; the photometric consistency error is determined by calculating the brightness residual between the temporally corresponding feature point pairs after photometric correction by the current unified distortion model. Based on the spatiotemporal consistency error, the parameters of the unified distortion model are iteratively optimized; Based on the optimized unified distortion model parameters, the synchronous stereo image sequence is corrected in real time to generate a distortion-free output stereo image sequence with consistent geometry and luminosity.

[0006] In some embodiments, acquiring a synchronized stereo image sequence includes: generating a periodic hardware trigger pulse signal through an external synchronization signal generator, and simultaneously inputting the hardware trigger pulse signal to the external trigger input pins of the left and right cameras to ensure that the exposure start times of the two camera image sensors are perfectly aligned.

[0007] In some embodiments, establishing the spatiotemporal correspondence of feature points across frames and viewpoints includes: An accelerated segmentation test feature detection algorithm is used to extract corner points as feature points in the current frame image of the left camera image sequence; The Canard-Lucas-Tommasi tracking algorithm based on pyramid optical flow is used to match the feature points in the current frame of the right camera image sequence to establish a spatial correspondence, and simultaneously track the feature points in the next frame of the left camera image sequence to establish a temporal correspondence.

[0008] In some embodiments, the specific process of combining quantization geometric consistency error and photometric consistency error is as follows: For any spatially corresponding feature point pair, its normalized coordinates in the left and right images are pl and pr, respectively. Its geometric consistency error is defined as the expression |prT·F·pl|, where the fundamental matrix F is calculated from the intrinsic and extrinsic parameters in the current unified distortion model. For any temporally corresponding feature point pair, its image coordinates in the current frame and the next frame are pt and p{t+1}, respectively. Its photometric consistency error is defined as the sum of the absolute values ​​of the pixel brightness differences between the image block centered at pt and the image block centered at p{t+1} after correction by the current unified distortion model.

[0009] In some embodiments, the iterative optimization of the unified distortion model parameters includes: weighting and summing the sum of squares of the geometric consistency errors of all spatially corresponding feature point pairs with the sum of squares of the photometric consistency errors of all temporally corresponding feature point pairs to construct a cost function; The cost function is optimized using the Levenberg-Marquardt nonlinear least squares method. In each iteration, the Jacobian matrix of the cost function with respect to all uniform distortion parameters is calculated, and the update equation is solved to obtain the parameter increment. The current parameter value is updated with the parameter increment until the magnitude of the parameter increment is lower than the preset convergence criterion.

[0010] In some embodiments, the step of real-time correction of the synchronized stereo image sequence based on optimized unified distortion model parameters includes: Based on the optimized internal parameters and distortion coefficients, a lookup table from distorted image coordinates to distortion-free image coordinates is pre-calculated for each camera. During correction, the geometrically distortion-free image is quickly generated using this lookup table and bilinear interpolation. Based on the optimized vignetting effect model parameters and color correction matrix, pixel-by-pixel brightness compensation and color space transformation are performed on images without geometric distortion, and finally, the corrected binocular image pairs are output.

[0011] In some embodiments, the lookup table is regenerated only when the change in the uniform distortion parameter exceeds a preset threshold, in order to maintain the real-time performance of the correction operation.

[0012] Another aspect of the present invention provides a binocular image distortion automatic correction system, comprising: The acquisition module is used to acquire a synchronized stereo image sequence, which consists of a left camera image sequence and a right camera image sequence, and the exposure acquisition of each frame in the left camera image sequence is strictly synchronized with the corresponding frame in the right camera image sequence in time. The spatiotemporal feature module is used to establish the spatiotemporal correspondence of feature points across frames and viewpoints. This module detects feature points in the current frame of the left camera image sequence and tracks the feature points in the current frame of the right camera image sequence and the next frame of the left camera image sequence, thereby obtaining spatially corresponding feature point pairs and temporally corresponding feature point pairs. An error module is used to jointly quantize geometric consistency error and photometric consistency error based on the spatially corresponding feature point pairs and the temporally corresponding feature point pairs; the geometric consistency error is determined by calculating the deviation between the spatially corresponding feature point pairs and the geometric constraints; the photometric consistency error is determined by calculating the brightness residual between the temporally corresponding feature point pairs after photometric correction by the current unified distortion model. The optimization module iteratively optimizes the parameters of the unified distortion model. This module constructs a cost function with the weighted sum of the geometric consistency error and photometric consistency error as the objective, and uses a nonlinear optimization algorithm to continuously adjust the unified distortion parameters until the cost function converges to below a preset threshold. The correction module is used to correct the synchronous stereo image sequence in real time based on the optimized unified distortion model parameters, and generate a distortion-free output stereo image sequence.

[0013] In some embodiments, the acquisition module, the spatiotemporal feature module, the error module, the optimization module, and the correction module are interconnected via a high-speed data cable.

[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention relies on information from natural scene image sequences for self-correction, requiring no manual intervention, no special calibration board, or offline calibration process; 2. This invention, through online optimization, can compensate in real time for drift in camera internal and external parameters caused by temperature changes, mechanical vibration, or component aging; 3. This invention improves the correction accuracy through joint optimization of geometry and photometry, resulting in improved quality in both three-dimensional structural consistency and appearance consistency of the generated corrected image; 4. This invention uses geometric consistency and photometric consistency as monitoring signals, which has extremely strong robustness. Attached Figure Description

[0015] To more clearly illustrate the technical solutions 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 of the binocular image distortion automatic correction method of the present invention; Figure 2 This is a schematic diagram of the binocular image distortion automatic correction system of the present invention. Detailed Implementation

[0017] The following will be based on embodiments of the present invention. Figures 1-2 The technical solutions in the embodiments of the present invention will be clearly and completely described together. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0018] Example 1 This embodiment provides an automatic correction method for binocular image distortion, including: acquiring a synchronized binocular image sequence, which consists of a left camera image sequence and a right camera image sequence; establishing a spatiotemporal correspondence of feature points across frames and viewpoints, including detecting feature points in the current frame of the left camera image sequence and tracking feature points in the current frame of the right camera image sequence and the next frame of the left camera image sequence, thereby obtaining spatially corresponding feature point pairs and temporally corresponding feature point pairs; and constructing a unified distortion model, which is parameterized by a set of unified distortion parameters, including the intrinsic parameter matrix elements of the left and right cameras, radial and tangential distortion... The system includes variable coefficients, rotation matrices and translation vectors between cameras, vignetting correction coefficients and color transformation matrix elements to describe luminance changes; it jointly quantizes geometric consistency error and luminance consistency error based on spatially and temporally corresponding feature point pairs; geometric consistency error is determined by calculating the deviation between spatially corresponding feature point pairs and geometric constraints; luminance consistency error is determined by calculating the luminance residual between temporally corresponding feature point pairs after luminance correction using the current unified distortion model; based on the spatiotemporal consistency error, iteratively optimizes the unified distortion model parameters; and based on the optimized unified distortion model parameters, iteratively corrects the synchronous stereo image sequence to generate a distortion-free output stereo image sequence.

[0019] Specifically as follows: First, a synchronized stereo image sequence is acquired. The synchronized stereo image sequence consists of a left camera image sequence and a right camera image sequence, and the exposure acquisition of each frame in the left camera image sequence is strictly synchronized with the exposure acquisition of the corresponding frame in the right camera image sequence.

[0020] To ensure time synchronization, an external synchronization signal generator is used to produce periodic hardware trigger pulse signals. These pulse signals are simultaneously input to the external trigger input pins of both the left and right cameras. Upon receiving the same trigger pulse, the image sensors of both cameras synchronously initiate exposure operations, ensuring that the left and right image frames are perfectly aligned in physical time. This process eliminates motion blur differences caused by inter-frame time offsets, providing a reliable data foundation for subsequent feature point tracking and consistency error calculation.

[0021] Next, after obtaining the synchronized stereo image sequence, the spatiotemporal correspondence of feature points across frames and viewpoints is established. Feature points are detected in the current frame of the left camera image sequence. The feature point detection adopts an accelerated segmentation test feature detection algorithm. The detection algorithm quickly identifies corner points with significant local contrast by comparing the brightness values ​​of pixels with multiple sampling points in their neighborhood.

[0022] During the detection process, a minimum response threshold is set to filter weak response points, and non-maximum suppression is applied to ensure that feature points are evenly distributed in space. Subsequently, feature points are matched in the current frame of the right camera image sequence to establish spatial correspondence; and feature points are tracked in the next frame of the left camera image sequence to establish temporal correspondence.

[0023] Both matching and tracking employ the Canard-Lucas-Tommasi tracking algorithm based on pyramid optical flow. The Canard-Lucas-Tommasi tracking algorithm constructs a multi-scale image pyramid, performing large-scale displacement estimation at the coarse-scale layer and sub-pixel accuracy optimization at the fine-scale layer.

[0024] For spatial matching, the optimal matching point is searched on the corresponding epipolar line of the current frame of the left image, using the feature point coordinates in the current frame of the left image as the initial value; for temporal tracking, the feature point position of the previous frame is used as the initial value, and optical flow tracing is performed in the next frame of the left image.

[0025] All successfully matched or tracked feature points are recorded with their precise coordinates in the left and right images and adjacent frames, forming two sets of correspondences: one set of spatially corresponding feature point pairs and the other set of temporally corresponding feature point pairs. If a feature point is lost in any target image, it is removed from the correspondence set to ensure the reliability of the data for subsequent error calculations.

[0026] Following the establishment of the spatiotemporal correspondence of feature points, a unified distortion model is constructed. This unified distortion model comprehensively describes the geometric and photometric distortion characteristics of the binocular camera system and is parameterized by a set of unified distortion parameters.

[0027] The unified distortion parameter set includes geometric parameters and photometric parameters. The geometric parameters include: the intrinsic parameter matrix K of the left camera. l Its element is focal length f xl f yl and the principal point coordinates c xl c yl Right camera internal parameter matrix K r Its element is focal length f xr f yr and principal point coordinates c xr c yr Radial distortion coefficient k1 of the left camera l k2 l k3 l and tangential distortion coefficient p1 l p2 l Radial distortion coefficient k1 of the right camera r k2 r k3 rand tangential distortion coefficient p1 r p2 r ; and a rotation matrix R and a translation vector t describing the pose of the right camera relative to the left camera coordinate system, where R is represented by three Euler angles or equivalent nine matrix elements, and t is represented by three translation components t x t y t z express.

[0028] The photometric parameters include: left camera vignetting effect model parameter a0 l a1 l a2 l The model is defined as V l(r) =a0 l + a1 l r + a2 l r², where r is the normalized radial distance from the pixel to the principal point; the vignetting effect model parameter a0 of the right camera. r a1 r a2 r Its model form is the same as that of the left camera; and a 3x3 color correction matrix M. rgb This is used to transform the RGB response of the right camera to the color space of the left camera to align the color deviations caused by the differences in the spectral responses of the two sensors.

[0029] Then, after the unified distortion model is constructed, the geometric consistency error and photometric consistency error are jointly quantized based on spatially and temporally corresponding feature point pairs. For the geometric consistency error, for each spatially corresponding feature point pair, let its pixel coordinates in the current frame of the left image be u. l p is obtained by back-projecting the current geometric model to the normalized camera coordinate system. l Similarly, the corresponding point u in the right image r Back projection yields p r .

[0030] The fundamental matrix F is calculated from the current intrinsic and extrinsic parameters, and its expression is F = K. r ^{-T} [t]x R K l ^{-1}, where [t]x is the antisymmetric matrix of the translation vector t, and the geometric consistency error e geo Defined as |p r ^T F p l| This value should theoretically be zero, but in practice it reflects the degree of violation of epipolar geometry constraints. For photometric consistency error, for each time-corresponding feature point pair, let its coordinates in the current frame of the left image be p. t The coordinates in the next frame are p{t+1}. t Extract an n×n image patch I centered on the image. t Extract the corresponding image patch I centered at p{t+1} {t+1} First, regarding I t and I {t+1} Correction is performed using the current photometric model: For each pixel, a vignetting compensation factor is calculated based on its distance to the principal point, and the color correction matrix M is applied. rgb (Applicable only to the right image; the left image is considered the reference). The corrected image patch is denoted as Î. t and Î {t+1}。 Photometric consistency error e photo Defined as ∑|Î t(i,j) - Î {t+1}(i,j) | represents the sum of the absolute values ​​of the pixel brightness differences between two image patches. e represents the sum of the values ​​of all feature point pairs. geo and e photo The values ​​are summed separately to form the total geometric error term E. geo With photometric error term E photo .

[0031] Then, based on the geometric consistency error and photometric consistency error, the parameters of the unified distortion model are iteratively optimized by constructing a joint cost function J, whose expression is J = w geo E geo + w photo E photo , where w geo With w photo These are preset weighting coefficients used to balance the relative importance of geometric and photometric constraints.

[0032] Furthermore, to minimize J, the Levenberg-Marquardt nonlinear least squares method is employed. In each iteration, the Jacobian matrix of J with respect to all uniform distortion parameters is calculated. Each row of this matrix corresponds to the partial derivative of an error term with respect to each parameter. These partial derivatives are obtained through numerical differentiation or analytical derivation. For example, for the focal length f... xl The partial derivative can be obtained through a small perturbation f xl And recalculate E geo With E photo To estimate. Further, solve the update equation (Jacobian^T) Jacobian + λ I) Δθ = -Jacobian^T `residual` is a function of the damping factor, I is the identity matrix, `residual` is the current error vector, and `Δθ` is the parameter increment. The current parameter value is updated using `Δθ`: θ new = θ old + Δθ.

[0033] After updating, recalculate J. If the decrease in J or the norm of Δθ is lower than the preset convergence threshold, terminate the iteration; otherwise, continue to the next round of optimization.

[0034] Finally, the synchronized binocular image sequence is corrected in real time to generate a distortion-free output binocular image sequence. The correction process is divided into two stages: geometric correction and photometric correction.

[0035] For geometric correction, based on the optimized intrinsic parameters and distortion coefficients, a lookup table from distorted image coordinates to undistorted image coordinates is pre-computed for each camera. This lookup table covers the entire image plane, and the corresponding position (u',v') of each output pixel position (u,v) is calculated in the original image through the inverse mapping of the distortion model.

[0036] Since (u', v') are usually non-integer coordinates, bilinear interpolation is used during correction to synthesize the output pixel value by weighting the brightness values ​​of the four integer pixels surrounding (u', v'). The lookup table is regenerated after the uniform distortion parameters are updated, but to avoid frequent recalculation affecting real-time performance, updates are only triggered when parameter changes exceed a certain threshold.

[0037] For photometric correction, first calculate the radial distance r from the pixel to the principal point, substitute it into the vignetting model V(r) to obtain the compensation gain, and multiply the pixel brightness by this gain; then apply the color correction matrix M to the right camera image. rgb A linear transformation of the RGB channels is performed while the left camera image remains unchanged. Ultimately, the corrected left and right images satisfy strict epipolar constraints in terms of geometric structure and exhibit highly consistent brightness distribution and color characteristics in terms of photometric performance. They can be directly used for high-precision stereo matching or 3D reconstruction.

[0038] Example 2 Based on the same inventive concept as the binocular image distortion automatic correction method in Embodiment 1 above, such as Figure 2As shown, the present invention also provides an automatic binocular image distortion correction system, which includes: an acquisition module for acquiring a synchronized binocular image sequence, the synchronized binocular image sequence consisting of a left camera image sequence and a right camera image sequence, wherein the exposure acquisition of each frame in the left camera image sequence and the corresponding frame in the right camera image sequence are strictly synchronized in time; a spatiotemporal feature module for establishing a spatiotemporal correspondence between feature points across frames and viewpoints, this module detecting feature points in the current frame of the left camera image sequence and tracking feature points in the current frame of the right camera image sequence and the next frame of the left camera image sequence, thereby obtaining spatially corresponding feature point pairs and temporally corresponding feature point pairs; and an error module for... The system jointly quantizes geometric consistency error and photometric consistency error based on feature point pairs and time-corresponding feature point pairs. Geometric consistency error is determined by calculating the deviation between spatially corresponding feature point pairs and geometric constraints. Photometric consistency error is determined by calculating the brightness residual between time-corresponding feature point pairs after photometric correction using the current unified distortion model. An optimization module iteratively optimizes the unified distortion model parameters. This module constructs a cost function with the weighted sum of geometric consistency error and photometric consistency error as the objective and uses a nonlinear optimization algorithm to continuously adjust the unified distortion parameters until the cost function converges to below a preset threshold. A correction module is used to correct the synchronized stereo image sequence in real time based on the optimized unified distortion model parameters, generating a distortion-free output stereo image sequence. The acquisition module, spatiotemporal feature module, error module, optimization module, and correction module are interconnected via high-speed data cables.

[0039] Understandably, the acquisition module consists of two high-frame-rate CMOS image sensors, a matching lens assembly, and a hardware synchronization trigger circuit, responsible for outputting a strictly time-aligned raw binocular image stream. The spatiotemporal feature module internally stores the hardware logic for accelerated segmentation test feature detection and pyramid optical flow tracking, enabling parallel processing of multiple frames of images in a pipeline manner and real-time output of the spatiotemporal correspondence of feature points. The error module is a dedicated digital signal processor that receives the feature point coordinates from the spatiotemporal feature module, performs quantization of geometric and photometric errors through parallel computation, and outputs an error vector. The optimization module is a multi-core central processing unit that runs the Levenberg-Marquardt optimization algorithm, receives the error vector, calculates the parameter increment, and sends the updated parameters back to the system's storage location. The correction module is implemented as a graphics processing unit, whose texture unit and shader program are configured to perform geometric lookup table mapping and photometric compensation operations, enabling full-frame correction of high-resolution binocular images at a rate of tens of frames per second.

[0040] Throughout the entire operation, the binocular image distortion automatic correction system of this embodiment has an anomaly handling mechanism. If the number of feature points is lower than the preset threshold, it indicates that the scene texture is lacking. At this time, parameter optimization is paused, the current model parameters are maintained, and a low confidence warning is issued. If the cost function diverges during the optimization process, it is reverted to the previous stable parameter set and the learning rate is reduced to try again.

[0041] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0042] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for automatic correction of binocular image distortion, characterized in that, include: Acquire a synchronized stereo image sequence, which consists of a left camera image sequence and a right camera image sequence; Establishing a spatiotemporal correspondence between feature points across frames and viewpoints involves detecting feature points in the current frame of the left camera image sequence and tracking the feature points in the current frame of the right camera image sequence and the next frame of the left camera image sequence, thereby obtaining spatially corresponding feature point pairs and temporally corresponding feature point pairs. A unified distortion model is constructed, which is characterized by a set of unified distortion parameters. The unified distortion parameter set includes the internal parameter matrix elements of the left and right cameras, radial and tangential distortion coefficients, rotation matrix and translation vector elements between the cameras, as well as vignetting correction coefficients and color transformation matrix elements used to describe luminance changes. Based on the spatially corresponding feature point pairs and the temporally corresponding feature point pairs, the geometric consistency error and the photometric consistency error are jointly quantized; the geometric consistency error is determined by calculating the deviation between the spatially corresponding feature point pairs and the geometric constraints; the photometric consistency error is determined by calculating the brightness residual between the temporally corresponding feature point pairs after photometric correction by the current unified distortion model. Based on the spatiotemporal consistency error, the parameters of the unified distortion model are iteratively optimized; Based on the optimized unified distortion model parameters, the synchronous stereo image sequence is corrected in real time to generate a distortion-free output stereo image sequence.

2. The method according to claim 1, characterized in that, The acquisition of the synchronized stereo image sequence includes: generating a periodic hardware trigger pulse signal through an external synchronization signal generator, and simultaneously inputting the hardware trigger pulse signal to the external trigger input pins of the left and right cameras to ensure that the exposure start times of the two camera image sensors are perfectly aligned.

3. The method according to claim 1, characterized in that, The establishment of the spatiotemporal correspondence of feature points across frames and viewpoints includes: An accelerated segmentation test feature detection algorithm is used to extract corner points as feature points in the current frame image of the left camera image sequence; The Canard-Lucas-Tommasi tracking algorithm based on pyramid optical flow is used to match the feature points in the current frame of the right camera image sequence to establish a spatial correspondence, and simultaneously track the feature points in the next frame of the left camera image sequence to establish a temporal correspondence.

4. The method according to claim 1, characterized in that, The specific process of combining quantization geometric consistency error and photometric consistency error is as follows: For any pair of corresponding feature points in space, their normalized coordinates in the left and right images are respectively p l and p r Its geometric consistency error is defined by the expression |p r T ·F·p l | where the fundamental matrix F is calculated from the intrinsic and extrinsic parameters in the current unified distortion model; for any pair of feature points corresponding to a given time, their image coordinates in the current frame and the next frame are respectively p t and p {t+1} Its photometric consistency error is defined as p t Image blocks centered on p {t+1} It is the sum of the absolute values ​​of the pixel brightness differences between the central image blocks after correction by the current unified distortion model.

5. The method according to claim 4, characterized in that, The iterative optimization of the unified distortion model parameters includes: weighting and summing the sum of squares of the geometric consistency errors of all spatially corresponding feature point pairs with the sum of squares of the photometric consistency errors of all temporally corresponding feature point pairs to construct a cost function; The cost function is optimized using the Levenberg-Marquardt nonlinear least squares method. In each iteration, the Jacobian matrix of the cost function with respect to all uniform distortion parameters is calculated, and the update equation is solved to obtain the parameter increment. The current parameter value is updated with the parameter increment until the magnitude of the parameter increment is lower than the preset convergence criterion.

6. The method according to claim 5, characterized in that, The step of real-time correction of the synchronized stereo image sequence based on the optimized unified distortion model parameters includes: Based on the optimized internal parameters and distortion coefficients, a lookup table from distorted image coordinates to distortion-free image coordinates is pre-calculated for each camera. During correction, the geometrically distortion-free image is quickly generated using this lookup table and bilinear interpolation. Based on the optimized vignetting effect model parameters and color correction matrix, pixel-by-pixel brightness compensation and color space transformation are performed on images without geometric distortion, and finally, the corrected binocular image pairs are output.

7. The method according to claim 6, characterized in that, The lookup table is regenerated only when the change in the uniform distortion parameter exceeds a preset threshold, in order to maintain the real-time performance of the correction operation.

8. A binocular image distortion automatic correction system, characterized in that, include: The acquisition module is used to acquire a synchronized stereo image sequence, which consists of a left camera image sequence and a right camera image sequence, and the exposure acquisition of each frame in the left camera image sequence is strictly synchronized with the corresponding frame in the right camera image sequence in time. The spatiotemporal feature module is used to establish the spatiotemporal correspondence of feature points across frames and viewpoints. This module detects feature points in the current frame of the left camera image sequence and tracks the feature points in the current frame of the right camera image sequence and the next frame of the left camera image sequence, thereby obtaining spatially corresponding feature point pairs and temporally corresponding feature point pairs. An error module is used to jointly quantize geometric consistency error and photometric consistency error based on the spatially corresponding feature point pairs and the temporally corresponding feature point pairs; the geometric consistency error is determined by calculating the deviation between the spatially corresponding feature point pairs and the geometric constraints; the photometric consistency error is determined by calculating the brightness residual between the temporally corresponding feature point pairs after photometric correction by the current unified distortion model. The optimization module iteratively optimizes the parameters of the unified distortion model. This module constructs a cost function with the weighted sum of the geometric consistency error and photometric consistency error as the objective, and uses a nonlinear optimization algorithm to continuously adjust the unified distortion parameters until the cost function converges to below a preset threshold. The correction module is used to correct the synchronous stereo image sequence in real time based on the optimized unified distortion model parameters, and generate a distortion-free output stereo image sequence.

9. The system according to claim 8, characterized in that, The acquisition module, the spatiotemporal feature module, the error module, the optimization module, and the correction module are interconnected via a high-speed data cable.

Images