An Improved Epipolar Constraint-Based Dynamic Object Detection Method

Abstract

This invention relates to an improved epipolar-constrained dynamic object detection method. The algorithm includes the following steps: S1: image acquisition and preprocessing; S2: estimation of the fundamental matrix; S3: calculation of the epipolar line and the distance from the feature point to the epipolar line; S4: introduction of the maximum a posteriori probability method; S5: calculation of the posterior probability and determination of dynamics; S6: removal of dynamic objects. This algorithm fully utilizes the prior semantic information already captured by the system that was not used by the original fixed threshold algorithm, improving the recognition rate of dynamic and static objects, thereby improving the system's positioning accuracy. At the same time, due to its universality, this invention avoids the tedious work of constantly manually adjusting the threshold when dealing with different scenarios.

Description

Technical Field

[0001] This invention belongs to the field of dynamic SLAM technology, specifically relating to an improved epipolar constraint dynamic object detection method. Background Technology

[0002] In the fields of computer vision and image processing, dynamic feature point removal is a crucial step in many applications, especially in scenarios such as moving object detection, video surveillance, 3D reconstruction, and robot visual navigation. Its core objective is to accurately identify and remove feature points corresponding to dynamic objects from image sequences, thereby preserving feature points in the static background and providing a reliable data foundation for subsequent visual tasks.

[0003] Traditional dynamic feature point culling methods are primarily based on the epipolar constraint principle. The epipolar constraint is a fundamental concept in stereo vision, describing the geometric relationship between a point and its corresponding point in two viewpoints. Specifically, when a feature point is matched in two images, it must satisfy the constraint condition of being located on an epipolar line. In dynamic object culling scenarios, it is typically assumed that feature points in the static background strictly satisfy the epipolar constraint, while feature points in dynamic objects deviate from the epipolar line due to motion. Therefore, by calculating the distance from the feature point to the epipolar line and comparing it with a preset fixed threshold, it is determined whether the point is dynamic. If the distance from the feature point to the epipolar line exceeds the threshold, the point is considered a dynamic object; otherwise, it is considered part of the static background.

[0004] This fixed threshold method based on epipolar constraints can achieve certain results in some simple scenarios. For example, when the moving objects are slow, the scene is relatively simple, and the background is relatively static, some dynamic feature points can be removed relatively accurately by setting a reasonable threshold. However, as application scenarios become more complex and the requirements for the accuracy of dynamic object detection increase, the limitations of this method gradually become apparent. Specifically, it has the following limitations:

[0005] 1) Lack of adaptability

[0006] Fixed thresholds cannot adapt to dynamic changes in different scenarios. For example, when dynamic objects move at high speeds or when the scene is highly complex, fixed thresholds may not accurately distinguish between dynamic and static feature points. Furthermore, when the direction of motion of a dynamic object is close to the epipolar direction, judging solely by the vertical distance difference can lead to misjudgment.

[0007] 2) Insufficient universality

[0008] Different dynamic scenarios require different threshold settings. For example, the motion patterns and speeds of the same object in an indoor environment may be drastically different from those in an outdoor scene. A fixed threshold is difficult to maintain consistent performance across multiple scenarios. This requires researchers to manually adjust the threshold for each specific scenario, increasing the complexity and difficulty of application of the method.

[0009] 3) Limitations of polar constraints

[0010] The epipolar constraint only considers the distance from the feature point to the epipolar line, ignoring other factors that may affect the detection of dynamic objects. This single criterion can easily lead to misjudgments in complex scenes, reducing the accuracy and robustness of dynamic object removal. Summary of the Invention

[0011] This invention addresses the limitations of fixed thresholds in epipolar constraint algorithms by proposing a maximum a posteriori probability method. Instead of the original fixed empirical threshold, it calculates the maximum a posteriori probability of each pixel's distance to different semantic information categories (people, trees, cars, etc.). If the static probability is less than the dynamic probability, the point is considered dynamic; otherwise, it is retained as a valid static point. Compared to fixed threshold algorithms, this invention fully utilizes prior semantic information already captured by the system that the original algorithm did not use, improving the recognition rate of dynamic and static objects and thus enhancing the system's positioning accuracy. Furthermore, due to its universality, this invention avoids the tedious work of constantly adjusting the threshold manually for different scenarios.

[0012] To achieve the above objectives, the technical solution of the present invention is as follows: an improved epipolar-constrained dynamic object detection method, the method comprising the following steps:

[0013] S1: Image acquisition and preprocessing

[0014] S2: Estimate the fundamental matrix.

[0015] S3: Calculate the epipolar line and the distance from the feature point to the epipolar line.

[0016] S4: Introduce the maximum a posteriori probability method.

[0017] S5: Calculate the posterior probability and determine the dynamics.

[0018] S6: Dynamic object culling.

[0019] S1: Image acquisition and preprocessing, as detailed below:

[0020] Image acquisition: Two images, I1 and I2, are acquired in consecutive frames using a binocular or monocular camera, ensuring that the time interval between the two images is short enough to capture the motion of moving objects.

[0021] Image preprocessing: The acquired image is preprocessed, including grayscale conversion, denoising and normalization. Denoising can be done using Gaussian filtering or other filtering methods to reduce the impact of image noise on subsequent processing.

[0022] Feature point extraction: Extract feature points from two images. Use algorithms such as SIFT, SURF or ORB to extract feature points and perform matching. Record the matched feature point pairs as (p1, p2).

[0023] Specifically, the two parallelograms represent two frames of images captured by the camera, I1 and I2, respectively. O1 and O2 are the optical centers of the camera that captured these two frames. Points p1 and p2 are the pixels of point P in three-dimensional space on different imaging planes, and these two points are also a pair of matched feature points. Points O1, O2, and P define a plane called the polar plane. The intersections of the polar plane with the two imaging planes I1 and I2 are l1 and l2, respectively, called epipolar lines. The coordinates of points p1 and p2, and their secondary coordinates p′1 and p′2, are then expressed as:

[0024] p1 = [u1, v1], p2 = [u2, v2]

[0025] p′1=[u1,v1,1], p′2=[u2,v2,1]

[0026] Wherein, S2: estimates the fundamental matrix, as follows.

[0027] Feature point selection: Select a set of feature points (usually 8 or more pairs) from the matched feature point pairs to estimate the fundamental matrix F. The selected feature points should be evenly distributed in the image to improve the robustness of the fundamental matrix estimation.

[0028] From the knowledge of epipolar geometry constraints, we know that the epipolar line l1 represents the possible projection position of a 3D spatial point P onto the imaging plane I1, and the epipolar line l2 represents the possible projection position of a 3D spatial point P onto the imaging plane I2. The epipolar line l2 on the imaging plane I2 can be represented by the equation au + bv + c = 0. The geometric meaning of the fundamental matrix F is the mapping relationship from p1' to the epipolar line l2, that is:

[0029]

[0030] Calculate the fundamental matrix: Estimate the fundamental matrix F using the eight-point method, and let F be:

[0031]

[0032] The fundamental matrix F satisfies:

[0033]

[0034] Optimize the fundamental matrix: Further optimize the fundamental matrix F using nonlinear optimization methods to reduce estimation errors. During the optimization process, the reprojection error can be used as the optimization objective to further improve the accuracy of the fundamental matrix.

[0035] Wherein, S3: calculates the epipolar line and the distance from the feature point to the epipolar line, as detailed below.

[0036] After obtaining the epipolar line l2, the distance D from pixel p2 to the epipolar line is calculated using the point-to-line distance formula, satisfying the following:

[0037]

[0038] In classic dynamic object detection methods, a fixed threshold is set, and point D is compared with this threshold. If D is greater than the threshold, the point is considered a dynamic point; if D is less than or equal to the threshold, it is retained. In this invention, however, this step only requires calculating D, without using a one-size-fits-all approach of comparing it to a fixed threshold.

[0039] S4: Introduces the maximum a posteriori probability method, as detailed below.

[0040] Set the prior probability P(ω) i ): Set dynamic prior probability P(ω1) and static prior probability P(ω2) based on semantic information;

[0041] Set the conditional probability P(X|ω) i Assume that the distance D from the feature point to the epipolar line follows a normal distribution, and set the mean μ and standard deviation σ for dynamic and static cases; where P(ω1|X) represents the probability that the pixel is a dynamic pixel when D is observed to be X, and P(ω2|X) represents the probability that the pixel is a static pixel when D is observed to be X.

[0042] Calculate the conditional probability P(X|ω) i ): The conditional probability is calculated based on the probability density function of the normal distribution and satisfies:

[0043]

[0044] Where P(X|ω1) is the probability of observing D as X when the current pixel is a dynamic pixel, and P(X|ω2) is the probability of observing D as X when the current pixel is a static pixel.

[0045] This step is one of the key steps in this invention. Instead of using a fixed threshold to determine whether a pixel is a dynamic point, this invention sets the following parameters in this step: dynamic prior probability P(ω1), mean μ and standard deviation σ under dynamic and static conditions. This invention sets the parameters according to the following classification of objects that may be identified in the image: one category is animals and objects that people often carry, such as cats, dogs, people, backpacks, etc. These objects have a high dynamic probability, a large mean, and a small standard deviation; another category is objects that exist basically as static objects in daily life, such as buildings, chairs, etc. These objects have a low dynamic probability, a small mean, and a small standard deviation; and the third category is various means of transportation, such as cars, bicycles, etc. The dynamic probability and static probability of these objects are equal, with a large mean and a large standard deviation. To put it simply, this step corresponds to the fixed threshold set in classic dynamic object detection methods. However, the parameters set here have the following advantages over the fixed threshold: Since the system can read the object attributes in the camera, its dynamics can be judged more accurately, thus improving the success rate of dynamic object detection; although more parameters are set than in the original method, once the parameters are set, the system can guarantee the success rate of dynamic object detection in different scenes and environments without adjusting the parameters, while the fixed threshold needs to be adjusted in real time according to the environment to ensure its effectiveness.

[0046] S5: Calculate the posterior probability and determine the dynamic nature, as detailed below.

[0047] Calculate the posterior probability λ: Calculate the dynamic and static posterior probabilities using Bayes' theorem:

[0048]

[0049] Determine the dynamic nature: If λ>1, the feature point is considered a dynamic point; otherwise, the feature point is considered a static point.

[0050] Steps S5 and S4 are complementary and crucial to this invention. Step S4 replaces the fixed threshold parameter of the original method with more stringent and precise parameters. Similar to the original method's comparison of D with a fixed threshold to determine its dynamism, step S5 calculates and divides the posterior probabilities of the pixel's dynamic and static states respectively. If the posterior probability is greater than 1, it indicates a higher probability of dynamic states; conversely, it indicates a higher probability of static states.

[0051] S6: Dynamic object culling, as detailed below.

[0052] Removing dynamic points: Feature points judged as dynamic are removed from subsequent processing, retaining only static feature points for subsequent visual tasks. Removing dynamic points reduces the interference of dynamic objects on subsequent processing and improves the robustness of the system.

[0053] Optimization results: The fundamental matrix F is re-estimated based on the removed feature points, and steps S3 to S5 are repeated to further optimize the accuracy of dynamic object removal. Through iterative optimization, the effect of dynamic object removal is gradually improved.

[0054] Compared with the prior art, the advantages of the present invention are as follows:

[0055] This invention improves the accuracy of dynamic object detection, thereby enhancing the system's localization accuracy. By capturing the system's prior information about objects, this invention uses the maximum a posteriori probability method to improve the system's accuracy in detecting dynamic objects. This invention selects sitting and walking sequences from the TUM dataset for dataset experiments. This experiment verifies whether the improved method can achieve better dynamic localization accuracy than the original method. The comparison metrics mainly include the root mean square error (RMSE), mean, and median of the absolute trajectory error (APE) and relative trajectory error (RPE), as well as the magnitude of the change compared to the original method.

[0056]

[0057] As can be seen from the table, the improved method used in this invention has higher system positioning accuracy for both sitting sequences with low environmental dynamism and walking sequences with high dynamism.

[0058] Another major advantage of this invention is that it directly abandons the cumbersome steps of constantly modifying fixed thresholds for different environments such as indoors and outdoors, and uses the maximum a posteriori probability method, a mathematical method. Although more parameters are set, as long as the pre-setting is successful, theoretically no further changes are needed in the subsequent operation of the system, achieving a one-time solution. Attached Figure Description

[0059] Figure 1 This is an epipolar constraint plot. Detailed Implementation

[0060] Example: A method for detecting dynamic objects in images based on an improved epipolar constraint algorithm, comprising the following steps:

[0061] S1: Image Acquisition and Preprocessing

[0062] Image Acquisition: Acquire two images, I1 and I2, in consecutive frames using a stereo or monocular camera. Ensure the time interval between the two images is short enough to capture the motion of moving objects.

[0063] Image preprocessing: Preprocessing the acquired image includes operations such as grayscale conversion, denoising, and normalization. Denoising can be performed using Gaussian filtering or other filtering methods to reduce the impact of image noise on subsequent processing.

[0064] Feature point extraction: Feature points are extracted from the two images using algorithms such as SIFT, SURF, or ORB, and then matched. The matched feature point pairs are denoted as (p1, p2). When extracting feature points, those with good discriminative power and stability should be selected to improve the accuracy of subsequent calculations.

[0065] Specifically, such as Figure 1 As shown, the two blue parallelograms represent two frames of images captured by the camera, I1 and I2, respectively. O1 and O2 are the optical centers of the camera that captured these two frames. Points p1 and p2 are the pixels of point P in three-dimensional space on different imaging planes, and these two points are also a pair of matched feature points. Points O1, O2, and P define a plane called the polar plane. The intersections of the polar plane with the two imaging planes I1 and I2 are l1 and l2, respectively, called epipolar lines. The coordinates of points p1 and p2, and their secondary coordinates p1' and p'2, are expressed as follows:

[0066] p1 = [u1, v1], p2 = [u2, v2]

[0067] p′1=[u1,v1,1], p′2=[u2,v2,1]

[0068] S2: Estimating the fundamental matrix

[0069] Feature point selection: Select a set of feature points (usually 8 or more pairs) from the matched feature point pairs to estimate the fundamental matrix F. The selected feature points should be evenly distributed in the image to improve the robustness of the fundamental matrix estimation.

[0070] From the knowledge of epipolar geometry constraints, we know that the epipolar line l1 represents the possible projection position of a 3D spatial point P onto the imaging plane I1, and the epipolar line l2 represents the possible projection position of a 3D spatial point P onto the imaging plane I2. The epipolar line l2 on the imaging plane I2 can be represented by the equation au + bv + c = 0. The geometric meaning of the fundamental matrix F is the mapping relationship from p′1 to the epipolar line l2, that is:

[0071]

[0072] Calculate the fundamental matrix: Estimate the fundamental matrix F using the eight-point method. Let F be:

[0073]

[0074] The fundamental matrix F satisfies:

[0075]

[0076] Optimize the fundamental matrix: Further optimize the fundamental matrix F using nonlinear optimization methods to reduce estimation errors. During the optimization process, the reprojection error can be used as the optimization objective to further improve the accuracy of the fundamental matrix.

[0077] S3: Calculate the epipolar line and the distance from the feature point to the epipolar line.

[0078] After obtaining the epipolar line l2, the distance D from pixel p2 to the epipolar line is calculated using the point-to-line distance formula, satisfying the following:

[0079]

[0080] If the distance is greater than the threshold, the point will be considered a dynamic point; if the distance between pixels is less than or equal to the threshold, it will be retained.

[0081] S4: Introducing the Maximum A posteriori probability method

[0082] Set the prior probability P(ω) i Based on semantic information (such as people, trees, cars, etc.), set dynamic prior probabilities P(ω1) and static prior probabilities P(ω2). For example, the dynamic probability of a person is 0.7 and the static probability is 0.3; the static probability of a tree is 0.9 and the dynamic probability is 0.1. These prior probabilities can be adjusted according to the actual application scenario.

[0083] Set the conditional probability P(X|ω) i Assume that the distance D from the feature point to the epipolar line follows a normal distribution, and set the mean μ and standard deviation σ for dynamic and static cases. For example, the mean is larger and the standard deviation is smaller in the dynamic case; the mean is smaller and the standard deviation is smaller in the static case. Here, P(ω1|X) represents the probability that the pixel is a dynamic pixel when D is observed to be X, and P(ω2|X) represents the probability that the pixel is a static pixel when D is observed to be X.

[0084] Calculate the conditional probability P(X|ω) i ): The conditional probability is calculated based on the probability density function of the normal distribution and satisfies:

[0085]

[0086] Where P(X|ω1) is the probability of observing D as X when the current pixel is a dynamic pixel, and P(X|ω2) is the probability of observing D as X when the current pixel is a static pixel.

[0087] S5: Calculate the posterior probability and determine the dynamics.

[0088] Calculate the posterior probability λ: Calculate the dynamic and static posterior probabilities using Bayes' theorem:

[0089]

[0090] Determine the dynamic nature: If λ>1, the feature point is considered a dynamic point; otherwise, the feature point is considered a static point.

[0091] S6: Dynamic Object Removal

[0092] Removing dynamic points: Feature points identified as dynamic are removed from subsequent processing, retaining only static feature points for subsequent visual tasks (such as 3D reconstruction, camera localization, etc.). Removing dynamic points reduces the interference of dynamic objects on subsequent processing and improves the robustness of the system.

[0093] Optimization results: The fundamental matrix F is re-estimated based on the removed feature points, and steps S3 to S5 are repeated to further optimize the accuracy of dynamic object removal. Through iterative optimization, the effect of dynamic object removal can be gradually improved.

[0094] By setting reasonable mean and standard deviation for the normal distribution of D under different semantics and states, and setting empirical probabilities for different states under different semantics, this invention uses the maximum a posteriori probability method to effectively solve the problems caused by fixed thresholds, improve the recognition rate of dynamic and static objects, and thus improve the positioning accuracy of the system.

[0095] Derivation of the maximum a posteriori probability method

[0096] Bayes' theorem:

[0097]

[0098] P(ω i |X) is the posterior probability, representing the probability of event ω given the observed data X. i The probability of occurrence.

[0099] P(X|ω i Let ω be the conditional probability, representing the probability of event ω. i Given that this occurs, the probability of observing data X.

[0100] P(ω i Let ω be the prior probability, representing the probability of event ω before any observed data is available. i The probability of occurrence.

[0101] P(X) is the total probability of observing X, calculated using the law of total probability:

[0102]

[0103] Specifically in the invention, P(ω) i The probability is divided into dynamic probability and static probability, which are the dynamic probability and static probability of a certain semantic information (person, tree, etc.), denoted as P(ω1) and P(ω2). P(X) is the probability that the observed pixel point is within a limit distance D of X. The total probability formula can be written as:

[0104] P(X)=P(X|ω1)P(ω1)+P(X|ω2)P(ω2) (1-8)

[0105] P(X|ω i The probability of observing D as X is divided into the dynamic case P(X|ω1) and the static case P(X|ω2). P(X|ω1) is the probability of observing D as X when the current pixel is a dynamic pixel, and P(X|ω2) is the probability of observing D as X when the current pixel is a static pixel. i The probability of a pixel being a dynamic pixel when D is X is also divided into two cases: P(ω1|X) and P(ω2|X). P(ω1|X) represents the probability that the pixel is a dynamic pixel when D is X, and P(ω2|X) represents the probability that the pixel is a static pixel when D is X. We then present a specific application of the Bayesian formula in this invention:

[0106]

[0107] It is evident that if λ > 1, this pixel should be removed as a dynamic point; otherwise, it should be retained as a static point. Observing the right side of the equation, P(ω1) and P(ω2) are prior probabilities. In several classic semantic information acquired by YOLO, we have given the corresponding empirical dynamic and static probabilities in advance. For example, the dynamic probability of a person is 0.7, and the static probability is 0.3; the static probability of a tree is 0.1, and the dynamic probability is 0.9. This prior information has been determined by the system before this pixel is judged. P(X|ω1) and P(X|ω2) are conditional probabilities. Under uniform illumination, we assume that D follows a normal distribution, and the conditional probability P(X|ω1) is... i It can be calculated using the probability density function of the normal distribution:

[0108]

[0109] Where, μ i Let σ be the mean of the normal distribution that D of the current pixel satisfies under a specific semantic condition in both dynamic (denoted as μ1) and static (denoted as μ2) cases, and let σ be the mean of the distribution. i These parameters are all empirical parameters, representing the corresponding standard deviations. For example, the standard deviation for semantic information such as cars will be set larger than that for semantic information such as people and trees, which have relatively weak uncertainty and are subject to obvious dynamic or static biases.

[0110] By setting reasonable mean and standard deviation for the normal distribution of D under different semantics and states, and setting empirical probabilities for different states under different semantics, this invention uses the maximum a posteriori probability method to effectively solve the problems caused by fixed thresholds, improve the recognition rate of dynamic and static objects, and thus improve the positioning accuracy of the system.

Claims

1. A method for detecting dynamic objects in images based on an improved epipolar constraint algorithm, characterized in that, The method includes the following steps: S1: Image acquisition and preprocessing S2: Estimate the fundamental matrix. S3: Calculate the epipolar line and the distance from the feature point to the epipolar line. S4: Introduce the maximum a posteriori probability method. S5: Calculate the posterior probability and determine the dynamics. S6: Dynamic object culling; S1: Image acquisition and preprocessing, as detailed below: Image acquisition: Acquiring two images in consecutive frames using a stereo or monocular camera. and To ensure that the time interval between the two images is short enough to capture the motion of moving objects; Image preprocessing: The acquired image is preprocessed, including grayscale conversion, denoising and normalization. Denoising uses Gaussian filtering or other filtering methods to reduce the impact of image noise on subsequent processing. Feature point extraction: Feature points are extracted from two images using algorithms such as SIFT, SURF, or ORB. These feature points are then matched, and the matched feature point pairs are denoted as follows: , The two parallelograms represent two images captured by the camera, respectively. and , and These are the optical centers and points of the camera that captured these two images, respectively. and points For three-dimensional space points Pixels on different imaging planes, and these two points are also a pair of matched feature points. , , These three points define a plane called the polar plane, which intersects with the two imaging planes. and The lines of intersection are respectively and , called the polar line, then the point and points The coordinates of the first and second coordinates and They are represented as follows: ； S2: Estimate the fundamental matrix, as follows. Feature point selection: Select a set of feature points from the matched feature point pairs to estimate the fundamental matrix. The selected feature points should be evenly distributed across the image to improve the robustness of the fundamental matrix estimation. From the knowledge of polar geometric constraints, we know that the polar line Representing a point in three-dimensional space In the imaging plane Possible projection positions, and epipolar lines Representing a point in three-dimensional space In the imaging plane Possible projection positions, epipolar lines In the imaging plane It can be expressed by equations To represent, the fundamental matrix The geometric meaning is to the poles The mapping relationship is as follows: Calculate the fundamental matrix: Estimate the fundamental matrix using the eight-point method. ,make for: Fundamental matrix satisfy: Optimize the fundamental matrix: Further optimize the fundamental matrix using nonlinear optimization methods. To reduce estimation errors, during the optimization process, reprojection error is used as the optimization objective to further improve the accuracy of the fundamental matrix; S4: Introduce the maximum a posteriori probability method, as detailed below. Set prior probabilities : Set dynamic prior probabilities based on semantic information and static prior probability ; Set conditional probability Assume the distance from the feature point to the epipolar line. It follows a normal distribution, and we set the mean μ and standard deviation σ for dynamic and static conditions; where, Indicates that when observed for In this case, the probability that this pixel is a dynamic pixel. Indicates that when observed for In the case of [condition], the probability that this pixel is a static pixel; Calculate conditional probability The conditional probability, calculated based on the probability density function of the normal distribution, satisfies: in, When the current pixel is a dynamic pixel, the observed for The probability, When the current pixel is a static pixel, the observed for The probability of.

2. The image dynamic object detection method based on the improved epipolar constraint algorithm according to claim 1, characterized in that, S3: Calculate the epipolar line and the distance from the feature point to the epipolar line, as detailed below. In obtaining the polar line Then, the point-to-line distance formula is used to calculate the pixel point. Distance to the polar line satisfy: 。 3. The image dynamic object detection method based on the improved epipolar constraint algorithm according to claim 1, characterized in that, S5: Calculate the posterior probability and determine the dynamic nature, as detailed below. Calculate the posterior probability Calculate the dynamic and static posterior probabilities using Bayes' theorem: Determine dynamism: If If the value is greater than 1, the feature point is considered a dynamic point; otherwise, the feature point is considered a static point.

4. The image dynamic object detection method based on the improved epipolar constraint algorithm according to claim 1, characterized in that, S6: Dynamic object culling, as detailed below. Removing dynamic points: Feature points judged as dynamic are removed from subsequent processing, retaining only static feature points for subsequent visual tasks. Removing dynamic points reduces the interference of dynamic objects on subsequent processing and improves the robustness of the system. Optimization result: Re-estimate the fundamental matrix based on the removed feature points. Then repeat steps S3 to S5 to further optimize the accuracy of dynamic object removal. Through iterative optimization, the effect of dynamic object removal is gradually improved.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements an image dynamic object detection method based on an improved epipolar constraint algorithm as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement an image dynamic object detection method based on an improved epipolar constraint algorithm as described in any one of claims 1-4.

Images