Camera calibration method based on reward feedback sampling and fixed benchmark

Abstract

This invention discloses a camera calibration method based on reward feedback sampling and a fixed reference, relating to the fields of computer vision and image processing technology. The method includes: Step 1, introducing a fixed reference; Step 2, decomposing the pose transformation from the local coordinate system of the calibration board to the real camera coordinate system, and measuring the pose of the local coordinate system of the calibration board relative to the fixed reference coordinate system using a laser tracker; Step 3, constructing a nonlinear optimization model, randomly selecting a portion from the calibration images and performing preliminary calibration using the minimum reprojection error, repeating this sampling and preliminary calibration process, and summarizing the calibration results for each parameter; Step 4, independently optimizing the preliminary results of each parameter obtained in Step 3 using a reward feedback sampling mechanism to determine the final calibration result. The camera calibration method based on reward feedback sampling and a fixed reference provided by this invention can solve problems such as poor noise resistance, low accuracy, and dependence on feature extraction quality in existing calibration methods.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, and in particular to a camera calibration method based on reward feedback sampling and a fixed benchmark. Background Technology

[0002] Camera calibration is a crucial technique in computer vision and image processing, aiming to determine the camera's intrinsic and extrinsic parameters. These parameters are essential for tasks such as image coordinate to world coordinate transformation and 3D reconstruction. Camera calibration typically includes intrinsic parameter calibration (e.g., focal length, principal point coordinates, radial and tangential distortion coefficients) and extrinsic parameter calibration (determining the camera's position and orientation in the world coordinate system, i.e., the camera's rotation matrix and translation vector). Existing camera calibration methods are mainly divided into three categories: self-calibration methods, active vision calibration methods, and target calibration methods. Target calibration methods use calibration boards with known geometry, achieving a good balance between accuracy and flexibility, and are therefore widely used in practice. However, target calibration methods still face many challenges in practical applications, such as outlier interference caused by feature matching errors and pixel-level coordinate perturbations caused by image acquisition defects, making it difficult to meet high-precision requirements in some cases. High-precision and robust camera calibration remains crucial for advancing vision measurement technology.

[0003] In traditional calibration processes, the calibration board needs to be imaged in multiple poses to estimate intrinsic and extrinsic parameters. However, frequent changes in the calibration board's pose increase the degrees of freedom, thus significantly expanding the dimensionality of the parameter space. This not only increases the optimization difficulty but also makes the estimation results more sensitive to initial values, affecting the final calibration accuracy and robustness.

[0004] Meanwhile, noise is also a major factor contributing to the instability of camera calibration results. RANSAC estimates the model by uniformly sampling the minimum subset and selecting the final solution based on the normal value rate, thus exhibiting robustness to outliers. However, this algorithm ignores the spatial distribution characteristics of the data, limiting its performance. BANSAC introduces a dynamic Bayesian network to update the normal value probability, but some nodes rely on a predefined conditional probability table, limiting the model's flexibility and affecting its adaptability to the real distribution. Summary of the Invention

[0005] The purpose of this invention is to provide a camera calibration method based on reward feedback sampling and a fixed benchmark, which solves the problems of poor noise resistance, low accuracy, and dependence on feature extraction quality in existing calibration methods.

[0006] To achieve the above objectives, this invention provides a camera calibration method based on reward feedback sampling and a fixed benchmark, comprising the following steps:

[0007] Step 1: Introduce a fixed datum into the calibration system to provide a unified fixed datum coordinate system reference;

[0008] Step 2: Use a robotic arm to adjust the position and orientation of the calibration board relative to the camera, performing a pose transformation from the local coordinate system of the calibration board to the real camera coordinate system. Decomposition, and simultaneously using a laser tracker to measure pose transformation, including pose transformation from the local coordinate system of the calibration board to the iTracker 6D attitude smart sensor coordinate system [R] ls |t ls Pose transformation from the iTracker 6D pose smart sensor coordinate system to a fixed reference coordinate system Where i∈{1,2,...,V} represents different placement postures of the calibration board; the iTracker 6D attitude intelligent sensor is an important component of the laser tracker;

[0009] Step 3: Construct a nonlinear optimization model with camera intrinsic parameters, fixed reference extrinsic parameters, and distortion coefficients as optimization objectives. Randomly select a portion of samples from the acquired calibration images and optimize these parameters by minimizing the reprojection error of all calibration points in the sample images to obtain preliminary calibration results. Perform the sampling and preliminary calibration process independently multiple times and summarize the results of each calibration parameter obtained from each sampling.

[0010] Step 4: Use a reward-feedback-driven dynamic Bayesian optimization method to independently optimize the preliminary results of each calibration parameter obtained in Step 3, and determine the final calibration result.

[0011] Furthermore, the decomposition expression in step 2 is:

[0012]

[0013] In the formula, This represents the rotation matrix from the current local coordinate system of the calibration board to the actual camera coordinate system; This represents the translation vector from the current local coordinate system of the calibration board to the real camera coordinate system; 0 T This represents a row vector with three elements, where each element has a value of 0; This represents the homogeneous coordinates of the calibration points on the calibration plate in the local coordinate system; R rc This represents the rotation matrix from the fixed reference coordinate system to the actual camera coordinate system; t rc This represents the translation vector from the fixed reference coordinate system to the actual camera coordinate system; This represents the rotation matrix from the current local coordinate system of the calibration plate to the fixed reference coordinate system; This represents the translation vector from the current local coordinate system of the calibration plate to the fixed reference coordinate system; This represents the rotation matrix from the current iTracker 6D attitude smart sensor coordinate system to the fixed reference coordinate system; R represents the translation vector from the current iTracker 6D attitude smart sensor coordinate system to the fixed reference coordinate system; ls This represents the rotation matrix from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor; t ls This represents the translation vector from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor; This represents the homogeneous coordinates of the calibration point on the current calibration plate in a fixed reference coordinate system.

[0014] The 6V external parameters required for traditional multi-view calibration are reduced to 6, with a theoretical dimensionality reduction rate of 100%. This effectively reduces the dimensionality of external parameters.

[0015] Furthermore, the nonlinear optimization model in step 3 is as follows:

[0016]

[0017] In the formula, Θ={K,R rc ,t rc ,d} gives the set of optimization parameters; i ranges from 1 to V, representing traversing all images from different perspectives; j ranges from 1 to W, representing traversing all calibration points in each image; V represents the number of images from different perspectives acquired; W represents the number of calibration points that the calibration board can identify in each image; Represents the point observed in the i-th image in a fixed reference coordinate system. The corresponding homogeneous pixel coordinates; A(·) represents the nonlinear camera imaging model; K represents the camera intrinsic parameter matrix. Where f x and f y c represents the camera focal length along the x-axis and y-axis of the image, respectively, in pixels; x and c y These represent the coordinates of the principal point of the image on the x-axis and y-axis, respectively; d represents the distortion coefficient. This represents the homogeneous coordinates of the j-th calibration point observed in the i-th image in a fixed reference coordinate system.

[0018] Furthermore, the specific content of step 4 is as follows:

[0019] S41. Model the observed values ​​of each calibration parameter as random variables, and establish a random variable x. n The independent dynamic Bayesian network structure, where n∈{1,...,N}, and N is the number of samplings in step 3;

[0020] S42. In the kth iteration, constraints are applied based on the first-order Markov assumption, and the state of the observed random variable is determined to be normal or abnormal based on the absolute deviation Δ between the observed data value and the current calibration result.

[0021] S43. Perform weighted sampling on each calibration parameter to generate a high-confidence sample set. Based on expected value Estimate calibration parameters and calculate the normal value rate η k A reward feedback mechanism is introduced into the dynamic Bayesian network. A point-level reward function is defined, and the reward value is fed back to the dynamic Bayesian network to update the nodes. and The conditional probability mass function;

[0022] S44. In the k-th iteration, a posterior probability estimation method based on information fusion is adopted, which estimates the probability of historical normal values ​​from the previous iteration. and the current predicted normal value probability We obtain the updated normal probability by weighting.

[0023] S45. Regarding independent preliminary calibration results Based on the above reward feedback-driven sampling strategy, the probability of normal values ​​of each calibration parameter is initially estimated and updated iteratively. Then, high-confidence sample sets of each parameter are gradually generated through weighted sampling. On this basis, the expected value of the sample set is used to estimate the calibration parameters, and the final value of each calibration parameter is determined based on the maximum normal value rate criterion.

[0024] Furthermore, in S42:

[0025] The constraints are:

[0026]

[0027] In the formula, This represents the nth hidden state random variable in the kth iteration; In the k-th iteration, the relationship with The corresponding observed random variable; Represents a sequence of historical hidden states; Indicates a historical observation sequence; Let n represent the nth hidden state random variable in the (k-1)th iteration; all hidden state random variables and observed random variables are defined on a binary state space, where 0 represents an outlier and 1 represents a normal value;

[0028] The decision rule for the value of the hidden state random variable is defined as follows:

[0029]

[0030] In the formula, express The posterior probability of being judged as a normal value in the k-th iteration;

[0031] The decision rule for observing the values ​​of random variables is defined as follows:

[0032]

[0033] In the formula, Δ is the absolute deviation between the observed value and the calibration result in the current iteration k, and ∈ is the preset threshold.

[0034] Furthermore, the point-level reward function expression in S43 is:

[0035]

[0036] In the formula, This indicates that in the k-th iteration, the hidden state random variable... The generated point-level reward value is fed back to the dynamic Bayesian network at the (k+1)th iteration. Represents hidden state random variables The posterior probability of being judged as a normal value in the k-th iteration; Denotes the hidden state random variable in the k-th iteration The absolute deviation between the associated data and the currently calculated calibration value; η k This represents the normal value rate in the k-th iteration.

[0037] Furthermore, in S43 The conditional probability mass function is mainly composed of hidden state random variables. Point-based reward value Decide:

[0038] In the initial iteration, the conditional probability mass function is defined as:

[0039]

[0040] In the formula, Let represent the initial nth hidden state random variable; x represents the specific value of the hidden state random variable, x∈{0,1}; Indicates the first iteration with The corresponding observed random variable; c represents the specific value of the observed random variable, c∈{0,1}; δ c,x Let represent the Kronecker delta function, used to indicate whether c and x are equal; if c = x, then δ c,x The value of η is 1, otherwise it is 0; 0 This represents the normality rate during the initial iteration;

[0041] For k≥2, the conditional probability mass function is defined as:

[0042]

[0043] In the formula, This represents a weighted average of the different values ​​of the random variable J; the variable J is uniformly distributed on the set {1,2,...,k-1}; N represents the total number of samples; This indicates that in the J-th iteration, the hidden state random variable... The generated point-level reward value; This represents the Kronecker delta function, used to indicate the i-th hidden state random variable in the J-th iteration. Does the value of equal the given value x? If but The value is 1 if it is 1, otherwise it is 0. This represents the Kronecker delta function, used to indicate the i-th observed random variable in the (J+1)-th iteration. Does the value of equal the given observed value c? If but The value is 1 if it is 1, otherwise it is 0.

[0044] Furthermore, in S43 The update of the conditional probability mass function is controlled by the initial predefined probability distribution P0. In the initial iteration, the probability distribution P0 is directly inherited. For subsequent iterations k≥2, the probability shift in the log probability space is influenced by the reward signal. Adjustment, The expression for the conditional probability mass function is:

[0045]

[0046] In the formula, This represents the initial predefined probability distribution P0 given And observation In this case, The probability of; Indicates the reward value Indexation;

[0047] In the k-th iteration, given the observed random variable If the observed value c is given, then The posterior probability distribution in the previous iteration k-1 is updated as follows:

[0048]

[0049] Given but The conditional probability mass function is expressed as:

[0050]

[0051] Furthermore, the updated normal value probability in S44 The expression is:

[0052]

[0053] In the formula, μ is used to adjust the pair and The level of trust.

[0054] Therefore, the camera calibration method based on reward feedback sampling and fixed benchmark described above has the following beneficial effects:

[0055] (1) By introducing a fixed reference, the dimension of the parameter space is significantly reduced, the redundancy of the solution is reduced, and the system instability caused by the change of viewpoint is effectively suppressed, thereby improving the overall calibration accuracy.

[0056] (2) Combining the sampling strategy of dynamic Bayesian network based on reward feedback, the reward signal is calculated according to the current sampling result, which drives the dynamic update of the conditional probability quality function and normal value probability of each node in the dynamic Bayesian network, and the estimated result with the highest normal value rate is selected as the final calibration solution, which effectively enhances the robustness to outliers.

[0057] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0058] Figure 1 This is an overall flowchart of the camera calibration method based on reward feedback sampling and fixed benchmark of the present invention;

[0059] Figure 2 This is a schematic diagram of the joint calibration framework according to an embodiment of the present invention;

[0060] Figure 3 x is an embodiment of the present invention n Schematic diagram of dynamic Bayesian network structure

[0061] Figure 4 This is a comparison of the mean absolute error of the intrinsic parameters using different methods at different noise levels.

[0062] Figure 5 This invention compares the mean absolute error of extrinsic parameters from a fixed reference coordinate system to a real camera coordinate system under different noise levels.

[0063] Figure 6 This is a box plot showing the results of multiple internal parameter estimations under actual conditions according to the present invention;

[0064] Figure 7 This is a box plot showing the extrinsic parameter estimation results of the present invention under actual conditions, where the reference coordinate system was repeatedly fixed to the real camera coordinate system. Detailed Implementation

[0065] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0066] Please see Figure 1-7 The camera calibration method based on reward feedback sampling and a fixed benchmark includes the following steps:

[0067] Step 1: Introduce a fixed reference into the calibration system to provide a unified fixed reference coordinate system.

[0068] Step 2: Use a robotic arm to adjust the position and orientation of the calibration board relative to the camera, performing a pose transformation from the local coordinate system of the calibration board to the real camera coordinate system. Decomposition, and simultaneously using a laser tracker to measure pose transformation, including pose transformation from the local coordinate system of the calibration board to the iTracker 6D attitude smart sensor coordinate system [R] ls |t ls Pose transformation from the iTracker 6D pose smart sensor coordinate system to a fixed reference coordinate system Where i∈{1,2,...,V} represents different placement postures of the calibration board; the iTracker 6D attitude intelligent sensor is an important component of the laser tracker.

[0069] The decomposition expression is:

[0070]

[0071] In the formula, This represents the rotation matrix from the current local coordinate system of the calibration board to the actual camera coordinate system; This represents the translation vector from the current local coordinate system of the calibration board to the real camera coordinate system; 0 T This represents a row vector with three elements, where each element has a value of 0; This represents the homogeneous coordinates of the calibration points on the calibration plate in the local coordinate system; R rc This represents the rotation matrix from the fixed reference coordinate system to the actual camera coordinate system; t rc This represents the translation vector from the fixed reference coordinate system to the actual camera coordinate system; This represents the rotation matrix from the current local coordinate system of the calibration plate to the fixed reference coordinate system; This represents the translation vector from the current local coordinate system of the calibration plate to the fixed reference coordinate system; This represents the rotation matrix from the current iTracker 6D attitude smart sensor coordinate system to the fixed reference coordinate system; R represents the translation vector from the current iTracker 6D attitude smart sensor coordinate system to the fixed reference coordinate system; ls This represents the rotation matrix from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor; t ls This represents the translation vector from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor; This represents the homogeneous coordinates of the calibration point on the current calibration plate in a fixed reference coordinate system.

[0072] Step 3: Construct a nonlinear optimization model with camera intrinsic parameters, fixed reference extrinsic parameters, and distortion coefficients as optimization objectives. Randomly select a portion of samples from the acquired calibration images and optimize these parameters by minimizing the reprojection error of all calibration points in the sample images to obtain preliminary calibration results. Perform the sampling and preliminary calibration process independently multiple times and summarize the results of each calibration parameter obtained from each sampling.

[0073] The nonlinear optimization model is as follows:

[0074]

[0075] In the formula, Θ={K,R rc ,t rc ,d} gives the set of optimization parameters; i ranges from 1 to V, representing traversing all images from different perspectives; j ranges from 1 to W, representing traversing all calibration points in each image; V represents the number of images from different perspectives acquired; W represents the number of calibration points that the calibration board can identify in each image; Represents the point observed in the i-th image in a fixed reference coordinate system. The corresponding homogeneous pixel coordinates; A(·) represents the nonlinear camera imaging model; K represents the camera intrinsic parameter matrix. Where f x and f y c represents the camera focal length along the x-axis and y-axis of the image, respectively, in pixels; x and c y These represent the coordinates of the principal point of the image on the x-axis and y-axis, respectively; d represents the distortion coefficient. This represents the homogeneous coordinates of the j-th calibration point observed in the i-th image in a fixed reference coordinate system.

[0076] Step 4: Use a reward-feedback-driven dynamic Bayesian optimization method to independently optimize the preliminary results of each calibration parameter obtained in Step 3, and determine the final calibration result.

[0077] S41. Model the observed values ​​of each calibration parameter as random variables, and establish a random variable x. n The independent dynamic Bayesian network structure is given, where n∈{1,...,N}, and N is the number of samplings in step 3.

[0078] S42. In the k-th iteration, constraints are applied based on the first-order Markov assumption, and the state of the observed random variable is determined as normal or abnormal based on the absolute deviation Δ between the observed data value and the current calibration result.

[0079] The constraints are as follows:

[0080]

[0081] In the formula, This represents the nth hidden state random variable in the kth iteration; In the k-th iteration, the relationship with The corresponding observed random variable; Represents a sequence of historical hidden states; Indicates a historical observation sequence; Let n represent the nth hidden state random variable in the (k-1)th iteration; all hidden state random variables and observed random variables are defined on a binary state space, where 0 represents an outlier and 1 represents a normal value;

[0082] The decision rule for the value of the hidden state random variable is defined as follows:

[0083]

[0084] In the formula, express The posterior probability of being judged as a normal value in the k-th iteration;

[0085] The decision rule for observing the values ​​of random variables is defined as follows:

[0086]

[0087] In the formula, Δ is the absolute deviation between the observed value and the calibration result in the current iteration k, and ∈ is the preset threshold.

[0088] S43. Perform weighted sampling on each calibration parameter to generate a high-confidence sample set. Based on expected value Estimate calibration parameters and calculate the normal value rate η kA reward feedback mechanism is introduced into the dynamic Bayesian network. A point-level reward function is defined, and the reward value is fed back to the dynamic Bayesian network to update the nodes. and The conditional probability mass function is used to optimize the normal value sampling strategy in subsequent iterations.

[0089] The point-level reward function expression is as follows:

[0090]

[0091] In the formula, This indicates that in the k-th iteration, the hidden state random variable... The generated point-level reward value is fed back to the dynamic Bayesian network at the (k+1)th iteration. Represents hidden state random variables The posterior probability of being judged as a normal value in the k-th iteration; Denotes the hidden state random variable in the k-th iteration The absolute deviation between the associated data and the currently calculated calibration value; η k This represents the normal value rate in the k-th iteration.

[0092] in, The conditional probability mass function is mainly composed of hidden state random variables. Point-based reward value Decide:

[0093] In the initial iteration, the conditional probability mass function is defined as:

[0094]

[0095] In the formula, Let represent the nth hidden state random variable in the initial state; x represents the specific value of the hidden state random variable, x∈{0,1}; Indicates the first iteration with The corresponding observed random variable; c represents the specific value of the observed random variable, c∈{0,1}; δ c,x Let represent the Kronecker delta function, used to indicate whether c and x are equal; if c = x, then δ c,x The value of η is 1, otherwise it is 0; 0 This represents the normality rate during the initial iteration;

[0096] For k≥2, the conditional probability mass function is defined as:

[0097]

[0098] In the formula, This represents a weighted average of the different values ​​of the random variable J; the variable J is uniformly distributed on the set {1,2,...,k-1}; N represents the total number of samples; This indicates that in the J-th iteration, the hidden state random variable... The generated point-level reward value; This represents the Kronecker delta function, used to indicate the i-th hidden state random variable in the J-th iteration. Does the value of equal the given value x? If but The value is 1 if it is 1, otherwise it is 0. This represents the Kronecker delta function, used to indicate the i-th observed random variable in the (J+1)-th iteration. Does the value of equal the given observed value c? If but The value is 1 if it is 1, otherwise it is 0.

[0099] in, The update of the conditional probability mass function is controlled by the initial predefined probability distribution P0. In the initial iteration, the probability distribution P0 is directly inherited. For subsequent iterations k≥2, the probability shift in the log probability space is influenced by the reward signal. Adjustment, The expression for the conditional probability mass function is:

[0100]

[0101] In the formula, This represents the initial predefined probability distribution P0 given And observation In this case, The probability of; Indicates the reward value Indexation;

[0102] In the k-th iteration, given the observed random variable If the observed value c is given, then The posterior probability distribution in the previous iteration k-1 is updated as follows:

[0103]

[0104] Given but The conditional probability mass function is expressed as:

[0105]

[0106] S44. In the k-th iteration, a posterior probability estimation method based on information fusion is adopted, which estimates the probability of historical normal values ​​from the previous iteration. and the current predicted normal value probability We obtain the updated normal probability by weighting. The expression is:

[0107]

[0108] In the formula, μ is used to adjust the pair and The level of trust.

[0109] S45. Regarding independent preliminary calibration results Based on the above reward feedback-driven sampling strategy, the probability of normal values ​​of each calibration parameter is initially estimated and updated iteratively. Then, high-confidence sample sets of each parameter are gradually generated through weighted sampling. On this basis, the expected value of the sample set is used to estimate the calibration parameters, and the final value of each calibration parameter is determined based on the maximum normal value rate criterion.

[0110] Simulation Experiment

[0111] To evaluate the effectiveness of the proposed calibration method, the calibration method proposed in this invention was compared and analyzed under different uniformly distributed pixel noise levels. Specifically, the camera's intrinsic parameter matrix was set as follows: Radial distortion coefficients k1 = -0.1, k2 = 0.1. This applies to pose transformation from the local coordinate system of the calibration plate to the fixed reference coordinate system. Uniform noise of [-0.5°, 0.5°] and [-0.02mm, 0.02mm] was added, where i∈{1,2,...,V} represents different orientations of the calibration board within a range of 300mm to 600mm from the camera. The noise amplitude gradually increased from 0.1 pixels to 1.5 pixels in increments of 0.1, generating a total of 15 different noise levels. At each noise level, noise was uniformly sampled from the interval [-ξ,+ξ], ξ∈{0.1,0.2,...,1.5} and added independently to the pixel coordinates. 100 independent experiments were conducted for each pixel noise level. The calibration accuracy was evaluated by calculating the mean absolute error between the estimated parameters and the true values. The calibration results of this invention, from the intrinsic parameters and the fixed reference coordinate system to the extrinsic parameters of the real camera coordinate system, show that the reward feedback-driven sampling strategy combined with a fixed reference significantly reduces the dimensionality of the parameter space, decreases the intrinsic parameter estimation bias, and further improves the calibration accuracy and robustness. When the pixel error exceeds 1.0 pixel, although the deviation between the radial distortion coefficient k1 calculated by this method and the true value is slightly higher than that of Zhang Zhengyou's calibration method, the deviation is within a very small range. This shows that this method can still provide high calibration accuracy under high pixel error conditions.

[0112] Real experiment

[0113] From 95 captured images, 100 subsets were uniformly sampled, each containing 50 poses. Compared to Zhang Zhengyou's calibration method, the calibration results calculated using the method proposed in this invention are more concentrated, have fewer outliers, and are more consistent, indicating higher robustness under actual calibration conditions.

[0114] Therefore, this invention employs the aforementioned camera calibration method based on reward feedback sampling and a fixed benchmark. By introducing a fixed benchmark, the dimensionality of the parameter space is significantly reduced, redundancy in the solution is decreased, and system instability caused by changes in viewing angle is effectively suppressed, thereby improving the overall calibration accuracy. Simultaneously, by combining a dynamic Bayesian network sampling strategy based on reward feedback, a reward signal is calculated based on the current sampling results, driving the dynamic update of the conditional probability quality function and normal value probability of each node in the dynamic Bayesian network. The estimated result with the highest normal value rate is selected as the final calibration solution, effectively enhancing the robustness to outliers.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A camera calibration method based on reward feedback sampling and a fixed benchmark, characterized in that, Includes the following steps: Step 1: Introduce a fixed datum into the calibration system to provide a unified fixed datum coordinate system reference; Step 2: Use a robotic arm to adjust the position and orientation of the calibration board relative to the camera, performing a pose transformation from the local coordinate system of the calibration board to the real camera coordinate system. Decomposition was performed, and pose transformation was measured using a laser tracker, including pose transformation from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor. Pose transformation from the iTracker6D pose smart sensor coordinate system to a fixed reference coordinate system ,in To accommodate different placement orientations of the calibration board; the iTracker 6D attitude intelligent sensor is an important component of the laser tracker; This represents the rotation matrix from the current local coordinate system of the calibration board to the actual camera coordinate system; This represents the translation vector from the current local coordinate system of the calibration board to the actual camera coordinate system; This represents the rotation matrix from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor; This represents the translation vector from the local coordinate system of the calibration board to the coordinate system of the iTracker 6D attitude smart sensor; This represents the rotation matrix from the current iTracker 6D attitude smart sensor coordinate system to the fixed reference coordinate system; This represents the translation vector from the current iTracker 6D attitude smart sensor coordinate system to the fixed reference coordinate system; Step 3: Construct a nonlinear optimization model with camera intrinsic parameters, fixed reference extrinsic parameters, and distortion coefficients as optimization objectives. Randomly select a portion of samples from the acquired calibration images and optimize these parameters by minimizing the reprojection error of all calibration points in the sample images to obtain preliminary calibration results. Perform the sampling and preliminary calibration process independently multiple times and summarize the results of each calibration parameter obtained from each sampling. Step 4: Use a reward-feedback-driven dynamic Bayesian optimization method to independently optimize the preliminary results of each calibration parameter obtained in Step 3, and determine the final calibration result. S41. Model the observed values ​​of each calibration parameter as random variables, and establish random variables. Independent dynamic Bayesian network structure, in which , This refers to the number of samples taken in step 3. S42, in the In this iteration, constraints are applied based on the first-order Markov assumption, and the absolute deviation between the observed data values ​​and the current calibration results is used. Determine whether the observed random variable is a normal or outlier value; S43. Perform weighted sampling on each calibration parameter to generate a high-confidence sample set. According to expected value Estimate calibration parameters and calculate the normal value rate. A reward feedback mechanism is introduced into the dynamic Bayesian network. A point-level reward function is defined, and the reward value is fed back to the dynamic Bayesian network to update the nodes. and The conditional probability mass function; S44, in the In this iteration, a posterior probability estimation method based on information fusion is adopted, which estimates the probability of historical normal values ​​from the previous iteration. and the current predicted normal value probability We obtain the updated normal probability by weighting. ; S45. Regarding independent preliminary calibration results Based on the above reward feedback-driven sampling strategy, the probability of normal values ​​of each calibration parameter is initially estimated and updated iteratively. Then, high-confidence sample sets of each parameter are gradually generated through weighted sampling. On this basis, the expected value of the sample set is used to estimate the calibration parameters, and the final value of each calibration parameter is determined based on the maximum normal value rate criterion.

2. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 1, characterized in that, The decomposition expression in step 2 is: ； In the formula, This represents a row vector with three elements, where each element has a value of 0; This represents the homogeneous coordinates of the calibration points on the calibration plate in the local coordinate system. Represents the rotation matrix from the fixed reference coordinate system to the actual camera coordinate system; This represents the translation vector from the fixed reference coordinate system to the actual camera coordinate system; This represents the rotation matrix from the current local coordinate system of the calibration plate to the fixed reference coordinate system; This represents the translation vector from the current local coordinate system of the calibration plate to the fixed reference coordinate system; This represents the homogeneous coordinates of the calibration point on the current calibration plate in a fixed reference coordinate system.

3. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 2, characterized in that, The nonlinear optimization model in step 3 is as follows: ； In the formula, Provide the set of optimization parameters; From 1 to This indicates that all images from different perspectives are traversed; From 1 to This means traversing all the calibration points in each image; This indicates the number of images collected from different perspectives. This indicates the number of calibration points that the calibration board can identify in each image; Indicates the first Points observed in the fixed reference coordinate system in the image The corresponding homogeneous pixel coordinates; This represents a nonlinear camera imaging model; Represents the camera intrinsic parameter matrix ,in and Representing along the image shaft and The camera focal length of the axis, in pixels; and These represent the principal points of the image at... shaft and Coordinates on the axis; Indicates the distortion coefficient; Indicates the first The first image observed The homogeneous coordinates of each calibration point in a fixed reference coordinate system.

4. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 3, characterized in that, In S42: The constraints are: ； ； In the formula, Indicates the first In the nth iteration One hidden state random variable; Indicates the first In the next iteration and The corresponding observed random variable; Represents a sequence of historical hidden states; Indicates a historical observation sequence; Indicates the first In the nth iteration There are 1 hidden state random variables; all hidden state random variables and observed random variables are defined on a binary state space, where 0 represents an outlier and 1 represents a normal value; The decision rule for the value of the hidden state random variable is defined as follows: In the formula, express In the The posterior probability of being judged as a normal value in the next iteration; The decision rule for observing the values ​​of random variables is defined as follows: In the formula, For the current iteration The absolute deviation between the observed value and the calibration result This is a preset threshold.

5. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 4, characterized in that, The point-level reward function expression in S43 is: ； In the formula, Indicates the first In the next iteration, the hidden state random variable is... The generated point-level reward value, in the th... Feedback is provided to the dynamic Bayesian network during the next iteration; Represents hidden state random variables In the The posterior probability of being judged as a normal value in the next iteration; Indicates the first In the next iteration, the hidden state random variable The absolute deviation between the associated data and the currently calculated calibration value; Indicates the first The rate of normal values ​​in the next iteration.

6. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 5, characterized in that: S43 The conditional probability mass function is determined by the hidden state random variables. Point-based reward value Decide: In the initial iteration, the conditional probability mass function is defined as: ； In the formula, Indicates the initial number of... One hidden state random variable; This represents the specific values ​​that the hidden state random variable can take. ; Indicates the first iteration with The corresponding observed random variable; This indicates the specific values ​​that the observed random variable can take. ; Kronek Functions used to indicate and Are they equal? ​​If ,but The value is 1 if it is 1, otherwise it is 0. This represents the normality rate during the initial iteration; for The conditional probability mass function is defined as: ； In the formula, Represents the relationship between random variables The variable is weighted and averaged according to its different values; In the set Evenly distributed on the top; Indicates the total number of samples; Indicates the first In the next iteration, the hidden state random variable is... The generated point-level reward value; Kronek Function, used to indicate the first In the nth iteration Hidden state random variables Is the value equal to the given value? ;like ,but The value is 1 if it is 1, otherwise it is 0. Kronek Function, used to indicate the first In the nth iteration Observed random variables Does the value of equal the given observed value? ;if ,but The value is 1 if it is 1, otherwise it is 0.

7. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 6, characterized in that: S43 The update of the conditional probability mass function is subject to the initial predefined probability distribution. Control, probability distribution during the initial iteration Direct inheritance, for subsequent iterations The probability shift in the logarithmic probability space is affected by the reward signal. Adjustment, The expression for the conditional probability mass function is: ； In the formula, Represents the initial predefined probability distribution Given And observation In this case, The probability of; Indicates the reward value Indexation; In the In each iteration, given the observed random variable Observations ,but In the previous iteration The posterior probability distribution in the data is updated as follows: ； Given ,but The conditional probability mass function is expressed as: 。 8. The camera calibration method based on reward feedback sampling and fixed benchmark according to claim 7, characterized in that, The updated normal value probability in S44 The expression is: ； In the formula, Used to adjust and The level of trust.

Images