Geometric feature-based camera extrinsic parameter automatic calibration method

Abstract

The invention relates to the technical field of camera external parameter calibration, and discloses a geometric feature-based camera external parameter automatic calibration method, which comprises the following steps of: 1, acquiring a UTM point set and an image point set, and forming a corresponding distance matrix; constructing a cost matrix according to the distance matrix; optimal point pair matching is found, and a feature point set is formed; step 2, dividing the feature point set into a matched point set and an unmatched point set; forming a main region by using the matched point set; performing region division in the unmatched point set to form a sub-region set; 3, forming a main homography matrix according to the main region; a sub homography matrix is constructed for the unmatched point set in the sub region; and step 4, carrying out fine tuning optimization on the homography matrix in each area until an optimization threshold value is met, and obtaining a plurality of homography matrixes for external parameter calibration. According to the invention, automatic calibration is realized, the risk of mismatching is reduced, the processing efficiency and the result consistency are improved, and a high-precision calibration result can still be maintained in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of camera extrinsic parameter calibration, and specifically relates to an automatic calibration method for camera extrinsic parameters based on geometric features. Background Art

[0002] At present, with the booming development of intelligent transportation systems, the accurate calibration of roadside devices is the core foundation for ensuring the efficient operation of the system. Especially the calibration work of cameras, whose accuracy directly affects whether the intelligent transportation system can function accurately and effectively. As a key device for the intelligent transportation system to obtain traffic information, the calibration task of the camera aims to accurately match the image data captured by the camera with the geographical information in the real world, so as to realize the real-time monitoring and in-depth analysis of traffic conditions.

[0003] Currently, monocular cameras play an important role in the field of target recognition and positioning in intelligent transportation systems. For accurate target positioning, the extrinsic parameter calibration of the camera is an essential and important link. By comprehensively using the intrinsic and extrinsic parameters of the camera, the pixel coordinates in the image can be accurately mapped to the real-world coordinate system, and then the accurate coordinate information of the target in the real world can be obtained. This accurate coordinate information is of crucial significance for key applications such as traffic monitoring, autonomous driving, and intelligent traffic management, and is the key prerequisite for realizing the intelligence and automation of intelligent transportation systems.

[0004] However, most of the current traditional camera calibration methods still rely on manual parameter setting. The process is not only cumbersome and time-consuming, but also very likely to introduce various errors due to human factors during the operation. The existence of these errors, on the one hand, seriously affects the efficiency of camera calibration, making it difficult for the calibration work to quickly adapt to the rapid development needs of intelligent transportation systems; on the other hand, in a complex and changeable traffic environment, the introduced errors may lead to serious accuracy problems, which in turn have a negative impact on the overall traffic management effect and cannot meet the strict requirements of intelligent transportation systems for high-precision data. Therefore, with the rapid development of intelligent transportation and autonomous driving technologies, the demand for high-precision and automatic camera calibration solutions is becoming increasingly urgent. The existing traditional calibration methods are difficult to meet the continuously improving performance requirements of intelligent transportation systems. Summary of the Invention

[0005] The present invention aims to provide an automatic calibration method for camera extrinsic parameters based on geometric features to solve the problems that the existing camera calibration methods are cumbersome and time-consuming, have low automation efficiency, and are prone to introducing errors due to various factors, affecting the calibration efficiency and accuracy.

[0006] To achieve the above object, the present invention adopts the following technical solution. An automatic calibration method for camera extrinsic parameters based on geometric features includes the following steps.

[0007] Step 1: Obtain the UTM point set and the image point set, and respectively form the corresponding distance matrices D utm and D img ; According to the distance matrices D utm and D img construct a cost matrix; find the optimal point pair matching within the cost matrix to form a feature point set;

[0008] Step 2: Divide the feature point set into a matched point set and an unmatched point set according to the determination method; perform regional division in the matched point set to form a main region; perform regional division in the unmatched point set to form a sub-region set;

[0009] Step 3: Form multiple main homography matrices in the main region; and construct sub-homography matrices for the unmatched point sets in the sub-regions according to the calibration method;

[0010] Step 4: Fine-tune and optimize the homography matrices within each region until the set optimization threshold is met, and obtain multiple homography matrices for external parameter calibration.

[0011] The principle and advantages of this solution are:

[0012] When faced with the problems of high false matching rate, large error, and cumbersome operation and large workload in the existing calibration methods, the commonly considered optimization method is to optimize and improve the matching rate of the constructed mapping matrix and continuously adjust the matching strategy. However, it has not been found that the root cause of the problems brought by the existing algorithms lies in the lack of an effective geometric feature matching algorithm, and the relative positions and geometric features of the points are not fully utilized for optimization, but continuous efforts are made to optimize on a constructed mapping matrix. At the same time, when generating the mapping matrix, an algorithm for automatically fine-tuning coordinates is not adopted, resulting in an upper limit to the optimization effect. Therefore, even if the matching or calibration point positions are continuously calibrated in the existing technology, it is still difficult to meet the high-precision calibration requirements, thus leading to the limitations of the existing technology.

[0013] In this solution, by calculating the relative distance matrix between the UTM point set and the image point set, a geometric relationship model between points is established. The distance matrix provides a solid data foundation for subsequent matching, making the matching process more targeted. The best matching pairs are identified in the distance matrix to ensure a high similarity in geometric characteristics of the selected points and reduce the risk of false matching. Secondly, the feature point set is divided to ensure the effectiveness and accuracy of subsequent processing. For the matching points in the divided point sets, a main homography matrix is generated. For the unmatched point sets, the centers of each sub-region are identified, and several matching points closest to the center are determined to form a sub-homography matrix, thus completing region division and calibration, forming multiple mapping matrices in different environments, and effectively improving the applicability of the calibration process in complex environments. Finally, the pixel coordinates are finely adjusted to generate a new homography matrix and calculate the average error of each point. Through continuous iterative fine-tuning until the average error reaches the set threshold, the final optimized homography matrix is obtained to ensure the stability and accuracy of the system. Applying this solution has the following advantages:

[0014] 1. High-precision calibration: Through geometric feature matching and optimization algorithms, the accuracy of the calibration results is significantly improved, and the false matching rate is reduced.

[0015] 2. High degree of automation: This solution reduces the need for human intervention and adopts a fully automated processing flow, effectively reducing errors and the risk of mistakes, and improving operation efficiency.

[0016] 3. Strong adaptability: Through region division and calibration, this solution generates multiple homography matrices, which can effectively handle different geometric characteristics in complex scenarios, select a more suitable homography matrix for calibration, with higher accuracy, and ensure the reliability of the calibration accuracy in various environments.

[0017] 4. Strong noise suppression ability: Optimize the results of the preliminary matching, reduce the influence of noise and outliers, and effectively improve the stability of the system.

[0018] 5. High integration: This solution can be seamlessly integrated into existing intelligent transportation or vehicle-road cooperation systems, has strong compatibility and scalability, can work in coordination with different sensors and platforms, and greatly simplifies the system integration process. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the process framework of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0020] The following is a more detailed description through specific embodiments:

[0021] Embodiment 1

[0022] An automatic calibration method for the external parameters of a camera based on geometric features in this embodiment significantly reduces manual intervention through the automatic calibration method, reduces the risk of incorrect matching, improves the processing efficiency and the consistency of results, and performs regional calibration according to the geometric characteristics of different scenarios to ensure high-precision calibration results can still be maintained in complex environments. In this embodiment, the automatic calibration method is as shown in the appendix Figure 1 and includes the following steps:

[0023] S1. Obtain the UTM point set and the image point set, and respectively form the corresponding distance matrices D utm and D img ; construct a cost matrix according to the distance matrices D utm and D img ; find the optimal point pair matching within the cost matrix to form a feature point set.

[0024] In this embodiment, first, several positioning points are selected in a selected area according to a set method, and the UTM coordinate data and image coordinate data of the positioning points are respectively obtained, and the UTM point set can be expressed as P = {p 1 , p 2 , …, p n}, and the image point set can be expressed as Q = {q 1 , q 2 , …, q m}. In this embodiment, the set method is to set the required number of acquisition points and the corresponding point density according to the actual scene conditions in the delimited area, and use the method of average point selection to select several positioning points in the area according to the set number and density to ensure that the selected positioning points can cover the selected area, avoid omission, improve the selection efficiency at the same time, and can meet the description requirements of the terrain in the current area to adapt to different scenarios, make the selection method more general, and ensure that the obtained point set can meet the requirements for quickly constructing calibration accuracy.

[0025] In the existing algorithm, when collecting the coordinate information of feature points, it is necessary to obtain and save the corresponding image coordinates and UTM coordinates at the same time. If there is an incorrect match in the collected coordinate information, it will be very difficult to correctly associate the image coordinates and UTM coordinates of the feature points. This situation often requires re-collecting data, wasting time and resources, and resulting in a high incorrect matching rate.

[0026] After obtaining the corresponding point set data, calculate the distances between each point in each point set and other points respectively through the distance formula, and thus obtain the two-dimensional distance matrices D utm and D img respectively. To establish the geometric relationship between points, it provides a solid data basis for subsequent matching and makes the matching process more targeted.

[0027] In this embodiment, taking the UTM point set as an example, the distance formula can be expressed as

[0028]

[0029] In the formula, p i,x represents the x - coordinate value of the i - th point in the UTM point set, and p j,x represents the x - coordinate value of the j - th point; p i,y represents the y - coordinate value of the i - th point; p j,y represents the y - coordinate value of the j - th point.

[0030] The two - dimensional distance matrix of the UTM point set calculated through the distance formula is D utm , its dimension is N×N, and each element represents the Euclidean distance between the i - th UTM point and the j - th UTM point. Specifically, D utm can be expressed as

[0031]

[0032] Similarly, the two - dimensional distance matrix of the image point set calculated through the distance formula is D img , its dimension is M×M, and each element represents the Euclidean distance between the i - th image point and the j - th image point. Specifically, D img can be expressed as

[0033]

[0034] The matrices calculated through the distance formula are N×N and M×M distance matrices respectively, to represent the geometric relationship between each pair of points. Then, the similarity of the two sets of geometric relationships is used for matching analysis to construct a cost matrix for quantifying the matching cost between different point sets. This provides a reliable data basis for subsequent matching, ensuring that the system can make full use of geometric characteristics for optimization.

[0035] In this embodiment, the cost matrix is constructed by calculating the difference between the two distance matrices D utm and D img , and is expressed as:

[0036] costmatrix[i,j]=‖D utm [:,i]-D img [j,:]‖ 2 .

[0037] Then, the Hungarian algorithm is used to minimize the cost matrix and find the best point pair matching in the cost matrix. By calculating the average error E of each pair of matches, the optimal point pair matching is found to form a feature point set, ensuring that the selected point pairs have high similarity in geometric characteristics, thereby reducing the possibility of false matches. In this embodiment, the calculation expression of the average error is

[0038]

[0039] In the formula, k is the number of matching pairs, and i j and j j are the indexes of the j-th pair of matches respectively. By optimizing the matching pairs, the uncertainty in subsequent processing can be reduced, and the accuracy of the final calibration can be improved.

[0040] When establishing the mapping relationship, it is crucial to accurately select the coordinate points in the image and the UTM coordinate points. In the existing algorithms, when collecting UTM coordinates at the road end, a small deviation in position may cause significant errors. In addition, when obtaining the pixel coordinates of the feature points, due to the inevitable deviation in the manual selection process, an error of several pixels may lead to a significant deviation in the overall effect. These errors not only reduce the stability and reliability of the system, but may also accumulate in the subsequent processing process, further increasing the system error. Therefore, the method of relying on manually making the calibration matrix in the existing algorithms is not only inefficient, but also difficult to ensure consistency, resulting in an increase in the overall error.

[0041] In the face of the generated errors, in many existing technologies, it still relies on manual adjustment and verification, and it is necessary to manually check and select the coordinate information of the feature points (the coordinates of the feature points in the image and the coordinates in the UTM). This adjustment method not only increases the complexity of the operation, but also introduces the risk of potential human errors. After matching the coordinate points, it is also necessary to finely adjust the coordinate information of the feature points to eliminate the errors in manually collecting coordinates. This is not only time-consuming and laborious, but also difficult to ensure the consistency of accuracy, thus limiting the scalability and adaptability of the system.

[0042] S2. Divide the feature point set into a matched point set (matching point set) and an unmatched point set (non-matching point set) according to the determination method; form the main area with the matched point set; perform area division in the unmatched point set to form a sub-area set.

[0043] In this embodiment, before dividing the matched point set and the unmatched point set, it is also necessary to optimize the preliminary matching result by using the RANSAC (Random Sample Consensus) algorithm to reduce the influence of noise and false matches, form a feature point set, and then divide the feature point set according to the determination method.

[0044] Among them, the determination method includes the following sub-steps:

[0045] S2.1. Determine the initial region range, and classify the set of feature points within the initial region range into the matched point set; the rest are candidate points.

[0046] In this embodiment, the method for determining the initial region range is to first calculate the center C and radius R of the set of feature points, and the calculation formulas are expressed as

[0047]

[0048] In the formula, N is the number of already matched points.

[0049] Then, the region defined with the center C as the center point and 1 / 2R as the radius is determined as the initial region range. The set of feature points within the initial region range constitutes the matched point set, and the set of feature points not within the initial region range is the candidate points. By calculating the center and radius of the set of feature points, the matched point set and candidate points are divided based on this, ensuring the effectiveness and accuracy of subsequent processing.

[0050] S2.2. According to the set sorting method, add the nearest candidate points to the matched point set one by one to generate a mapping matrix, and calculate the average error E of the current mapping matrix new .

[0051] In this embodiment, the sorting method is to sort according to the distance from the candidate points to the determined center point from small to large. Starting from the candidate point with the closest distance, add the candidate points to the matched point set one by one to form the latest mapping matrix one by one. And during this process, calculate the average error E of the currently generated latest mapping matrix new for each one.

[0052] S2.3. Compare the calculated average error E new with the set threshold T. When E new > T, classify the current candidate point into the unmatched point set; otherwise, classify the current candidate point into the matched point set.

[0053] After the division of the matched point set is completed, form the main region with the matched point set. And perform region division on the formed unmatched point set. By calculating the angle θ between adjacent points. In this embodiment, the adjacent point angle θ is calculated with the center of the circle as the origin, and the calculation method of its adjacent angle θ is:

[0054] In the formula, C y represents the y coordinate value of the center point C, and C x represents the x coordinate value of the center point C.

[0055] Then, those with θ < 45° are grouped into the same region, and a set of sub-regions is formed in this way for more detailed calibration.

[0056] S3. Based on the main region, form the main homography matrix H; and construct sub-homography matrices for the unmatched point sets in the sub-regions according to the calibration method.

[0057] In this embodiment, the calibration method is to identify the center of each sub-region and select the three closest matched points to the center, denoted as M = {m 1 , m 2 , m 3}. In this embodiment, the calculation of the sub-homography matrix is based on the DLT (Direct Linear Transformation) method, and its form can be expressed as

[0058]

[0059] When the number of points in the sub-region = 1, that is, there is only one unmatched point, then form a point set with this point and the center of the circle to generate a sub-homography matrix; when the number of points in the sub-region > 1, then select the two closest matched points in M and the sub-region to form a point set together to generate a sub-homography matrix; finally, obtain the sub-homography matrices of each sub-region to ensure that each region can be independently optimized based on its characteristics.

[0060] S4. Fine-tune and optimize the homography matrices within each region until the set optimization threshold is met, and obtain multiple homography matrices for external parameter calibration.

[0061] To improve the calibration accuracy, in this embodiment, the deep reinforcement learning (DDPG, Deep Deterministic Policy Gradient) algorithm is used to automatically fine-tune the pixel coordinates, thereby optimizing the homography matrix and generating a new homography matrix. DDPG gradually adjusts the pixel coordinates through the interactive learning between the agent and the environment, making the calibration result tend to be optimal. Stop when the average error meets the set threshold through continuous fine-tuning, and finally obtain the optimized homography matrix. In actual application, according to the position of the calibration object, select the most suitable homography matrix for calibration to improve the calibration accuracy and matching degree.

[0062] In this embodiment, the fine-tuning and optimization include the following sub-steps:

[0063] S4.1. Coordinate transformation.

[0064] During the calibration process, the pixel points in the image are transformed into points in the real-world coordinate system through the constructed homography matrix H. Given a pixel point in the image represented as p img = (x img , y img), through the homography matrix H, the transformed real-world coordinates are represented as p real =(x real ,y real ), and its calculation method is as follows:

[0065]

[0066] In the formula, (x img ,y img ) is the pixel point in the image; H is the currently selected homography matrix; (x real ,y real ) are the transformed real-world coordinates. By continuously adjusting the pixel coordinates, the optimized H should be able to more accurately transform the image points into the real-world coordinate system.

[0067] S4.2, Calculate the average error.

[0068] In each fine-tuning process, based on the current pixel coordinates, a new homography matrix H is generated new , and the average error ò between the transformed real-world coordinates of each point and the actual predicted world coordinates is calculated avg , which is expressed as

[0069]

[0070] In the formula, P is the number of points to be optimized, p real is the real-world coordinate point obtained by transforming through the homography matrix H, and p pred is the coordinate point predicted by the current homography matrix. By calculating the error of each point and taking its average value, the overall error of the current model is obtained, which can reflect the accuracy of the current calibration result. The smaller the average error, the more accurate the current calibration result, thus ensuring the accuracy of the result.

[0071] S4.3, Algorithm optimization.

[0072] In this embodiment, the DDPG algorithm plays a key role in this process, and the accuracy of the algorithm needs to be optimized to improve the accuracy of the algorithm. It includes selecting an adjustment action for the current pixel coordinates according to the current calibration state information; calculating the adjusted average error, and giving a reward strategy according to the error change to optimize the selection of the adjustment action, and optimizing the homography matrix until the average error reaches the limit value and then terminating.

[0073] Specifically, in the DDPG algorithm, the state information consists of the current pixel coordinates and the corresponding homography matrix H. This state reflects the accuracy of the current calibration and the environment of the system. Through the state information, the agent can evaluate the difference between the current calibration result and the target.

[0074] Based on the state information, an adjustment action is selected. In the actor network of DDPG, an action is an adjustment to the current pixel coordinates. Since DDPG is applicable to continuous action spaces, the actor network will output a continuous value, usually in the range [-1,1], indicating the amount of change in the pixel coordinates. The magnitude of each adjustment is determined by the actor network, and through interaction with the current state, it decides how to adjust the position of the pixel point.

[0075] After each adjustment, the system calculates a new average error ò avg , and rewards are given according to the change of error. In this embodiment, the reward function is set as:

[0076] R(tad avg )=- avg ,

[0077] The reward function penalizes larger error values, encouraging the system to choose a pixel adjustment strategy that can reduce the error. By optimizing the reward, the agent will continuously adjust the pixel coordinates in the direction of lower error, thereby optimizing the homography matrix.

[0078] At the same time, in this embodiment, the quality of each action is evaluated through the critic network, and the Q value of each adjustment action is calculated. The calculation formula of the Q value is expressed as

[0079] Q(s t ,a t )=r t +γ·Q(s t+1 ,a t+1 );

[0080] In the formula, r t is the current reward, γ is the discount factor, Q(s t+1 ,a t+1 ) is the Q value of the next state and action; based on the feedback of the Q value, the actor network updates the adjustment strategy and adjusts the selection of pixel coordinates, so that the intelligent agent can gradually converge to an optimal solution in the process of continuous adjustment.

[0081] In this embodiment, the termination condition of the DDPG algorithm is to pass multiple iterations until the average error is avg Reaching the preset threshold threshold , that is, the error is small enough. At this point, the system stops adjusting and finally obtains the optimized homography matrix H final The calibration process ends when the following conditions are met:

[0082] tadpole avg ≤ threshold .

[0083] Through this process, the DDPG algorithm continuously optimizes the pixel coordinates based on the feedback information and adjusts the strategy in multiple iterations, ultimately enabling the accuracy of the homography matrix to meet the requirements.

[0084] Since the existing technologies usually neglect the geometric relationships between points, lack geometric feature analysis, and cannot effectively capture the relative positions of feature points, their improvement in calibration accuracy is limited. Moreover, the existing technologies often rely on manual intervention and lack an automated processing flow, resulting in low efficiency and easy introduction of human errors, leading to a high error rate. Also, due to the large amount of intervention work and cumbersome operation process, only one mapping matrix is set, assuming that it can meet the calibration requirements and ensure the unity of the mapping matrix. However, in the face of complex environmental changes, a single mapping matrix cannot be flexibly adjusted, restricting its adaptability in diverse application scenarios. Additionally, for the noise and interference in sensor data, the existing technologies also lack effective filtering and correction mechanisms, resulting in the instability of the calibration results.

[0085] In this embodiment, the root cause of the problem is creatively discovered, and the conventional view that a single mapping matrix with unity is required is breakthroughly abandoned. Instead, it is found that in the actual calibration process, most calibration objects are only in local positions and do not cover the entire area. Therefore, complete unity is not necessary, and more importantly, the targeting and accuracy of local features are required. Thus, this embodiment breaks the conventional thinking and breakthroughly establishes multiple mapping matrices for selection according to different environments.

[0086] In this embodiment, the main homography matrix is generated for the set of matched points; for the set of unmatched points, the centers of each sub-region are identified, and several matched points closest to the center are determined to form sub-homography matrices. This sub-region strategy improves the adaptability of the calibration process in complex environments, can select a more suitable homography matrix for calibration according to different scenarios, thereby reducing errors and mismatches, and can better meet the calibration requirements of more complex environments while improving the calibration accuracy. Especially under complex road conditions, such as uneven or sloped roads, the existing single mapping matrix often cannot effectively adapt, resulting in a decrease in local calibration accuracy. However, this embodiment adopts a sub-region calibration strategy to handle the geometric characteristics of different regions, thereby ensuring the accuracy and reliability of the calibration results.

[0087] Meanwhile, the pixel coordinates are finely adjusted through the Deep Deterministic Policy Gradient (DDPG) algorithm to generate a new homography matrix and calculate the average error of each point. Through continuous iterative fine-tuning until the average error reaches the set threshold, the finally optimized homography matrix is obtained to ensure the stability and accuracy of the system.

[0088] The above are only embodiments of the present invention, and common general technical solutions and / or characteristics in the solutions are not described in detail herein. It should be noted that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicability of the patent. The protection scope claimed in this application shall be subject to the content of its claims, and the specific implementation manners and the like recorded in the specification can be used to interpret the content of the claims.

Claims

1. A method for automatic calibration of camera extrinsic parameters based on geometric features, characterized in that: The following steps are included: Step 1: Get the UTM point set and the image point set, and form the corresponding distance matrix D respectively utm and D img ; According to the distance matrix D utm and D img Construct a cost matrix; find the optimal point pair matching in the cost matrix to form a feature point set; Step 2: Divide the feature point set into a matched point set and an unmatched point set according to the determination method; form a main region with the matched point set; perform region division in the unmatched point set to form a sub-region set; Step 3: Form the main homography matrix according to the main area; And construct a sub-homography matrix for the unmatched point set in the sub-region according to the calibration method; Step 4: Fine-tune and optimize the homography matrix in each area until it meets the set optimization threshold, and obtain multiple homography matrices for external parameter calibration.

2. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 1, characterized in that: In step 2, the determination method includes the following sub-steps: Step 2.1, determine the initial area range, classify the feature point set within the initial area range as the matched point set; the rest are candidate points; Step 2.2: According to the set sorting method, add the nearest candidate points one by one to the matched point set to generate a mapping matrix, and calculate the average error E of the current mapping matrix. new ; Step 2.3, the average error E new Compared with the set threshold T, when E new >T, it is included in the unmatched point set, otherwise it is included in the matched point set.

3. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 1, characterized in that: In step 2, the unmatched point set is divided into regions, the angle θ of adjacent points is calculated, and those with θ < 45° are classified into the same region to form a sub-region set.

4. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 1, characterized in that: In step 3, the calibration method is to identify the center of each sub-region and select the three matched points closest to the center; when the number of points in the sub-region = 1, the point and the center constitute a point set to generate a sub-homography matrix; when the number of points in the sub-region is greater than 1, the two nearest points are selected to form a point set together with the sub-region to generate a sub-homography matrix; and the sub-homography matrix of each sub-region is obtained.

5. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 2, characterized in that: In step 1, it also includes selecting a number of positioning points in the selected area according to a set method, and correspondingly obtaining the UTM coordinate data and image coordinate data of the positioning points to form a UTM point set P = {p1, p2, ..., p n } and image point set Q = {q1,q2,…,q m }; and through the formula Calculate the distance between each point in each point set and other points respectively, and get the distance matrix D utm and D img ; In the formula, p i,x represents the x-coordinate value of the i-th point in the UTM point set, p j,x represents the x-coordinate value of the jth point; p i,y represents the y coordinate value of the i-th point; p j,y Represents the y-coordinate value of the j-th point.

6. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 5, characterized in that: The cost matrix is ​​constructed by calculating the difference between two distance matrices, expressed as: costmatrix[i,j]=‖D utm [:,i]-D img [j,:]‖2。 7. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 1, characterized in that: In step 4, fine-tuning optimization includes the following sub-steps, Step 4.1, coordinate transformation: transform the pixels in the image into real-world coordinates through the constructed homography matrix, expressed as In the formula, (x img ,y img ) is the pixel point in the image; H is the currently selected homography matrix; (x real ,y real ) is the transformed real world coordinate; Step 4.2, calculate the average error: Generate a new homography matrix H based on the current pixel coordinate adjustment new , calculate the average error between the transformed real world coordinates of each point and the actual predicted world coordinatesò avg , expressed as Where P is the number of points to be optimized, p real is the real world coordinate point obtained by transforming the homography matrix H, p pred is the coordinate point predicted by the current homography matrix; Step 4.3, algorithm optimization: According to the current calibration status information, select the adjustment action for the current pixel coordinates; calculate the average error after adjustment, and give a reward strategy according to the error change to optimize the selection of adjustment actions and optimize the homography matrix until the average error reaches the limit.

8. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 5, characterized in that: In step 2.1, the initial region range is the center C and radius R of the feature point set, which can be expressed as Where N is the number of matched points; The area defined by center C as the center point and 1 / 2R as the radius is the initial area range.

9. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 1, characterized in that: In step 1, the Hungarian algorithm is also used to minimize the cost matrix and calculate the average error E of each pair of matches to find the optimal pair matching. The average error expression is: In the formula, k is the number of matching pairs, i j and j j are the indices of the j-th matching pair respectively.

10. The method for automatic calibration of camera extrinsic parameters based on geometric features according to claim 7, characterized in that: In step 4.3, the Q value of each adjustment action is calculated, which is expressed as Q(s t ,a t )=r t +γ·Q(s t+1 ,a t+1 ); In the formula, r t is the current reward, γ is the discount factor, Q(s t+1 ,a t+1 ) is the Q value of the next state and action; the strategy is updated and adjusted based on the feedback of the Q value.

Images