Evolution multi-task based probability point cloud registration method and system

Abstract

The application provides an evolutionary multi-task-based probability point cloud registration method, which combines an evolutionary multi-task optimization algorithm, constructs a first loss function of a global task and a second loss function of a local task; the application combines the multi-task and the point cloud registration task by using the characteristics that the multi-task optimization algorithm simultaneously optimizes multiple related tasks, optimizes the transformation parameters of the point cloud data by using the population optimization algorithm corresponding to the two tasks to obtain the optimal transformation parameters of the two loss functions at the minimum value, and transforms the preprocessed point cloud data by using the optimal transformation parameters of the first loss function at the minimum value. The application establishes an efficient and robust registration scheme and designs a reasonable evolutionary multi-task point cloud registration framework, and realizes a point cloud registration technology with high precision, high robustness and no influence of an initial pose.

Description

Technical Field

[0001] This invention belongs to the field of point cloud processing technology, specifically relating to a probabilistic point cloud registration method and system based on evolutionary multi-task. Background Technology

[0002] Point cloud data, a collection of rich discrete 3D points, is currently the most popular data format for representing the 3D physical world. Due to limitations in field of view and 3D scanning equipment, acquired point clouds typically only cover a local portion of an object or scene. Therefore, there is an urgent need to develop point cloud registration methods, aiming to determine a set of rigid transformation parameters that allow the source point cloud to be accurately aligned with the target point cloud to obtain complete object or scene information. Point cloud registration, as a crucial preliminary task in many fields, provides the foundation for various practical applications such as autonomous driving, virtual reality and augmented reality, simultaneous localization and mapping (SLAM), and 3D reconstruction. Especially in the field of autonomous driving, vehicles rely on sensors such as LiDAR to acquire 3D point cloud data of the surrounding environment, and the accuracy of point cloud registration directly affects the vehicle's perception and decision-making capabilities. Furthermore, in virtual reality and augmented reality applications, accurate point cloud registration can provide a more realistic and immersive experience. Simultaneously, for applications such as robot navigation and map building, 3D point cloud registration is fundamental to achieving accurate perception and mapping.

[0003] Currently, more and more point cloud registration methods are emerging, which can be mainly divided into three categories: traditional registration methods, probabilistic registration methods, and evolutionary registration methods.

[0004] 1) A representative traditional registration method is point cloud registration based on the Iterative Closest Point (ICP) algorithm (Besl PJ, McKay NDA Method for Registration of 3-D Shapes[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 1992, 14(2): 239-256.). It models point cloud registration as a least squares problem and solves it iteratively using Singular Value Decomposition (SVD) or quaternions. However, ICP only handles the case where two point clouds completely overlap, resulting in limited performance when dealing with real-world registration problems. Then, the Trimmed Iterative Closest Point (TrICP) algorithm (Chetverikov D, Stepanov D, Krsek P. Robust Euclidean alignment of 3D point sets: the trimmed iterative closestpoint algorithm[J]. Image and Vision Computing, 2005, 23(3): 299-309.) was designed to handle outliers caused by partial overlap. In each iteration, the overlap rate between the two point clouds is estimated and updated, and outlier removal is achieved by truncation. Although ICP-based algorithms are popular, these methods can only guarantee convergence to a local optimum, which makes them very sensitive to the initial pose of the point cloud pairs to be registered. Feature descriptor-based registration methods first extract features from the point clouds to be registered, then estimate the correspondence between points based on the extracted features, and finally use SVD to solve for the transformation matrix. Rusu et al. (Rusu RB, Marton ZC, Blodow N, et al. Persistent point feature histograms for 3D point clouds[C] / / Proc 10th IntConf Intel Autonomous Syst(IAS-10). Baden-Baden, Germany, 2008:119-128.) designed the PFH feature descriptor by considering the local curvature of the surface where the point is located.Subsequently, Rusu et al. (Rusu RB, Blodow N, Beetz M. Fast point feature histograms (FPFH) for 3D registration [C] / / 2009 IEEE international conference on robotics and automation. Kobe, Japan: IEEE, 2009: 3212-3217.) improved upon PFH by proposing the FPFH feature descriptor, which reduces computational overhead while maintaining the effectiveness of the feature descriptor. Salti et al. (Salti S, Tombari F, Di Stefano L. SHOT: Unique signatures of histograms for surface and texture description [J]. Computer Vision and Image Understanding, 2014, 125: 251-264.) proposed the SHOT descriptor, which not only considers the local neighborhood structure of points but also integrates color information to improve feature discriminativeness. However, these methods are less accurate than local registration methods and rely on methods to remove outliers.

[0005] 2) Most probabilistic registration methods are point cloud registration methods based on Gaussian Mixture Models (GMMs). These methods assume that the points in the point cloud are generated by one or more Gaussian mixture models, allowing the probability of each point's occurrence to be calculated according to certain rules. The Expectation Maximization (EM) algorithm is then used to solve for the probability. The Normal Distributions Transform (NDT) algorithm (Biber P, Strabeer W. The normal distributions transform: A new approach to laser scanmatching [C] / / Proceedings 2003 IEEE / RSJ International Conference on Intelligent Robots and Systems (IROS2003) (Cat.No.03CH37453). Las Vegas, NV, USA: IEEE, 2003:2743-2748.) simulates the distribution of all points by combining Gaussian distributions within the cells, meaning that a point-to-point correspondence does not need to be established. Then, the scores calculated from the distributions and the transform are iteratively optimized using Newton's algorithm to match the two point clouds. The Coherent Point Drift (CPD) algorithm proposed by Myronenko et al. (Myronenko A, Song X. Point set registration: Coherent pointdrift[J].IEEE Transactions on Pattern Analysis and Machine Intelligence, 2010, 32(12):2262-2275.) constructs a Gaussian mixture model with each point in the source point cloud as the center of the Gaussian model, and achieves registration by maximizing the probability of the target point cloud.

[0006] 3) Evolutionary registration methods utilize the global search characteristics of evolutionary algorithms to find the optimal transformation parameters. They have attracted widespread attention because they do not require the objective function to have mathematical properties (such as continuity and differentiability). Silva et al. (Silva L, Bellon ORP, Boyer K L. Precision range image registration using arobust surface interpenetration measure and enhanced genetic algorithms[J]. IEEE transactions on pattern analysis and machine intelligence, 2005, 27(5):762-776.) proposed a hybrid genetic algorithm (GA) for registration, where the fitness function is the objective function using a designed surface interpenetration measure (SIM). Since the objective function based on SIM cannot be optimized by numerical optimization methods, evolutionary computation methods are naturally chosen to solve for the global optimum. Li et al. (Li T, Pan Q, Gao L, et al. Differential evolution algorithm-based rangeimage registration for free-form surface parts quality inspection[J]. Swarm and Evolutionary Computation, 2017, 36: 106-123.) designed a novel point descriptor for establishing correspondences in evaluation and used a modified differential evolution algorithm to search for transformations. Zhu et al. (Zhu J, Zhu L, Li Z, et al. Automatic multi-view registration of unordered range scans without feature extraction[J]. Neurocomputing, 2016, 171: 1444-1453.) applied GA to provide the initial pose for fine registration of TRICP, solving the local convergence problem of the ICP-based algorithm.

[0007] Traditional registration methods are highly sensitive to the estimation of the initial registration transformation. Inaccurate initial estimates can cause the algorithm to fail to converge to the correct solution and easily get trapped in local optima, failing to find the global optimum. This is particularly pronounced in complex scenarios. Furthermore, point cloud data often contains noise and outliers, and traditional methods are sensitive to these interfering factors, easily leading to erroneous registration results.

[0008] Probabilistic registration methods involve establishing a Gaussian mixture model based on the point cloud distribution and then solving this model using Expectation-Maximization (EM) or Variational Bayesian estimation. Due to the inherent limitations of EM, it is a locally optimal algorithm and cannot guarantee convergence to the global optimum. This means that the transformation parameters obtained by solving the probabilistic point cloud registration problem using EM are not optimal and cannot meet the accuracy requirements of practical applications. Variational Bayesian estimation requires making assumptions about the prior probability distributions of each parameter and variable. These assumptions differ from the true distributions of the parameters, and this process is subjective, leading to inaccurate estimation of the point cloud data. Furthermore, variational Bayesian estimation requires the Mean-Field condition, which assumes that the variational distribution is a completely decomposable mean-field form, i.e., a strong independence assumption. This assumption ignores the correlation between variables, but in practical applications, the correlation between variables cannot be ignored; otherwise, it will reduce the accuracy of the variational Bayesian inference approximation. Additionally, converting the point cloud into a probability density model ignores its rich local geometric features. To simplify problem solving, they typically construct GMMs with isotropic covariance, which causes GMM components to only penalize point-to-point distances between two point clouds, limiting the model's representational power.

[0009] Existing evolutionary registration methods typically utilize evolutionary algorithms to search for the optimal transformation parameters for point cloud registration. However, these methods often process each registration task independently during optimization, without fully considering knowledge sharing between different registration tasks. This leads to insufficient interaction of point cloud information, making the methods prone to getting trapped in local optima. Summary of the Invention

[0010] To address the aforementioned problems in the existing technology, this invention provides a probabilistic point cloud registration method and system based on evolutionary multi-task. The technical problem to be solved by this invention is achieved through the following technical solution:

[0011] In a first aspect, the present invention provides a probabilistic point cloud registration method based on evolutionary multi-task, comprising:

[0012] S100: Acquire point cloud data and preprocess the point cloud data to obtain preprocessed point cloud data;

[0013] S200, construct the first loss function for the global task and the second loss function for the local task;

[0014] S300, Set the population optimization algorithm corresponding to the global task and the local task, as well as the individual representation, search space and initialization parameters in the population optimization algorithm, and initialize the population optimization algorithm;

[0015] S400, the transformation parameters of the point cloud data are optimized by using the population optimization algorithm corresponding to the global task and the transformation parameters of the point cloud data are optimized by using the population optimization algorithm corresponding to the local task, so as to obtain the optimal transformation parameters of the first loss function at the minimum value and the optimal transformation parameters of the second loss function at the minimum value respectively.

[0016] S500, the preprocessed point cloud data is transformed using the optimal transformation parameters of the first loss function at its minimum value.

[0017] Secondly, the present invention provides a probabilistic point cloud registration system based on evolutionary multi-task, comprising:

[0018] The acquisition module is configured to acquire point cloud data and preprocess the point cloud data to obtain preprocessed point cloud data.

[0019] The building module is configured to construct the first loss function for the global task and the second loss function for the local task;

[0020] The setting module is configured to set the population optimization algorithm corresponding to the global task and the local task, as well as the individual representation, search space and initialization parameters in the population optimization algorithm, and initialize the population optimization algorithm.

[0021] The optimization module is configured to optimize the transformation parameters of the point cloud data using the population optimization algorithm corresponding to the global task and the population optimization algorithm corresponding to the local task, so as to obtain the optimal transformation parameters of the first loss function at the minimum value and the optimal transformation parameters of the second loss function at the minimum value, respectively.

[0022] The transformation module is configured to transform the preprocessed point cloud data using the optimal transformation parameters of the first loss function at its minimum value.

[0023] Beneficial effects:

[0024] This invention addresses the limitations of most registration methods, which either convert point clouds into probability density models while ignoring rich local geometric information or focus only on local geometry within the point cloud neighborhood while neglecting global structural information, thus restricting their representational capabilities. It proposes an evolutionary multi-task-based probabilistic point cloud registration method. This method combines an evolutionary multi-task optimization algorithm to construct a first loss function for the global task and a second loss function for the local task. Leveraging the collaborative mechanism of multiple tasks, this invention integrates multi-tasks with the point cloud registration task. It uses population optimization algorithms corresponding to the two tasks to optimize the transformation parameters of the point cloud data, obtaining the optimal transformation parameters at the minimum of the two loss functions. The optimal transformation parameters at the minimum of the first loss function are then used to transform the preprocessed point cloud data. This invention establishes an efficient and robust registration scheme and designs a reasonable evolutionary multi-task point cloud registration framework, achieving high-precision, high-robustness, and point cloud registration technology unaffected by initial pose. The method will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating a probabilistic point cloud registration method based on evolutionary multi-task provided by the present invention.

[0026] Figure 2 This is a schematic diagram illustrating the difference between the adaptive symmetric point-to-surface GMM and the nearest neighbor correspondence loss provided by this invention;

[0027] Figure 3 This is a flowchart of the population optimization algorithm provided by the present invention for solving the optimal transformation parameters;

[0028] Figure 4 This is a schematic diagram of the final registration visualization results provided by the present invention with σ values ​​of 0.01, 0.03, and 0.05;

[0029] Figure 5 These are registration visualizations of various methods provided by this invention on different outlier data;

[0030] Figure 6 This is a visual comparison diagram of registration on a real-world Arch scene provided by the present invention;

[0031] Figure 7 This invention provides a registration visualization diagram in a real-world Office environment.

[0032] Figure 8 This is a registration visualization diagram on a real-world Courtyard scene provided by the present invention. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0034] Before introducing the solution of this invention, a probabilistic point cloud registration method based on evolutionary multi-task will be described in detail.

[0035] This invention proposes a probabilistic point cloud registration method based on evolutionary multi-task. This method addresses the limitations of most registration methods, which either convert point clouds into probability density models while ignoring the rich local geometric information of the point cloud, or focus only on the local geometry within the point cloud's neighborhood while neglecting global structural information, thus restricting their representational capabilities. This method combines evolutionary multi-task optimization algorithms, leveraging the collaborative mechanism of multiple tasks to simultaneously optimize several related tasks. It integrates evolutionary multi-task with point cloud registration, establishing an efficient and robust registration model and designing a reasonable evolutionary multi-task point cloud registration framework, achieving high-precision, highly robust, and point cloud registration technology unaffected by initial pose.

[0036] The technical details of the present invention are described in detail below.

[0037] Firstly, such as Figure 1 As shown, this invention provides a probabilistic point cloud registration method based on evolutionary multi-task, comprising:

[0038] S100: Acquire point cloud data and preprocess the point cloud data to obtain preprocessed point cloud data;

[0039] This step involves downsampling and decentralizing the point cloud data to obtain preprocessed point cloud data.

[0040] It is worth noting that point cloud data typically contains a large number of points. To reduce computational costs, improve efficiency, and decrease storage requirements, point clouds need to be downsampled. Point cloud downsampling methods include random downsampling, voxel grid downsampling, and farthest point downsampling. This invention can employ farthest point sampling, which selects points that typically have the highest distinctness, thus effectively reducing redundant information while preserving the main structure of the point cloud.

[0041] Decentralization transforms the coordinate system of a point cloud into a coordinate system relative to the centroid of the point cloud, which helps improve the robustness of point cloud processing algorithms to initial estimation errors. By subtracting the centroid of the point cloud, sensitivity to initial estimation is reduced, mitigating the impact of initial estimation errors on the registration algorithm and thus simplifying the complexity of the registration problem. First, the centroid of the point cloud, i.e., the average of the coordinates of all points, is calculated. Then, the point cloud is decentralized by subtracting the centroid from the coordinates of each point.

[0042] S200, construct the first loss function for the global task and the second loss function for the local task;

[0043] It is worth noting that, to address the issues of poor robustness and susceptibility to local optima in point cloud registration methods, a multi-task construction approach specific to point cloud registration is proposed. This approach primarily involves constructing two related tasks to form a multi-task scenario. These two tasks focus on different solution spaces: Task 1, called the global task Tg, mainly focuses on acquiring global structural information, while Task 2, as an auxiliary task, is called the local task T. l It focuses on capturing more accurate local geometric information.

[0044] The global task is modeled using an improved Gaussian mixture model. The local task is constructed using an adaptive robust model that focuses on local geometric information, supplementing it with more accurate local feature information. It further explores narrow regions of the solution space by utilizing local surface geometry.

[0045] The loss function constructed by the global task focuses on capturing global structural information through an improved GMM. This makes the loss function more accurate and complex, with many local optima in the parameter space, making it prone to getting trapped in local optima. Therefore, by using a local task as an auxiliary task, which mainly focuses on more prominent point-plane relationships to more accurately capture the local surface geometry of the point cloud, this approach is relatively simple and helps the global task escape local optima.

[0046] Through the effective implicit synergistic effect of evolutionary multi-tasks, the two tasks interact with each other. The global task can provide the local task with good point cloud distribution information (global structural information) to help obtain more accurate local geometric information. At the same time, the local task can supplement the global task with richer and more accurate local geometric feature information to help make more accurate point cloud distribution estimation.

[0047] In one specific embodiment of the present invention, S200 includes:

[0048] S210, Construct a multi-task environment; the multi-task environment includes two related tasks, the two related tasks focus on different solution spaces, the two related tasks are divided into task one and task two, task one is a global task used to obtain global structure information, and task two is a local task used to capture local set information.

[0049] S220, using an adaptive symmetric point-to-surface Gaussian mixture model, construct the first loss function for the global task; this step S220 includes: S221, using a Gaussian mixture model to perform a rigid transformation on the source point cloud, and treating the transformed source point cloud as the observation value of the GMM; S222, given the GMM and the observation value, construct the first loss function for the global task.

[0050] The loss function of the global task Tg is modeled using a Gaussian mixture model. An adaptive symmetric point-to-surface Gaussian mixture model (GMM) is proposed. By improving the anisotropic covariance, it achieves adaptive fusion of point-to-point penalty and point-to-surface penalty, which can focus on richer global structural information of the entire point cloud and is closer to the real probability distribution. Figure 2 This demonstrates the difference between the adaptive symmetric point-to-surface GMM and the nearest neighbor correspondence loss.

[0051] This invention formulates the registration problem of two point clouds as a maximum likelihood estimation problem. This is achieved by fixing the target point cloud and aligning the source point cloud to it. Therefore, the GMM is built upon the target point cloud. Yes, regarding source cloud A rigid transformation is performed, and the transformed source point cloud is considered as the observation value of the GMM. The optimization objective is to find the optimal transformation (rotation matrix R and translation vector t) that best suits the observation value to the GMM.

[0052] S222 of the present invention includes:

[0053] S2221, Obtain the probability density function of the Gaussian Mixture Model (GMM), expressed as:

[0054]

[0055] θ j =(μ j ,Σ ij ),μ j =y j

[0056] Where w0∈[0,1) is x i The outlier weights are denoted by U(N), where U(N) represents the outlier probability density as a uniform distribution. M and N are the number of points after downsampling the target point cloud and the source point cloud, respectively. In addition to the M Gaussian components centered at each point in the target set, an outlier component is also added to the GMM. For the corresponding component density function, μ j ,Σ ij These are the corresponding mean and covariance matrices, respectively;

[0057] S2222, using the probability density function to design the specific form of the corresponding component density function, the key novelty of this invention lies in... The design incorporates point-to-point and symmetric point-to-surface penalties into probabilistic modeling, expressed as:

[0058]

[0059]

[0060] Where d represents the dimension of the random variable, which is d = 3 in point cloud registration, x i It is a source point cloud, y j It is the target point cloud, σ 2 It is the covariance multiplier, α ij R is the penalty coefficient, and R is the rotation matrix. The surface normals of the source and target point clouds are respectively, and I is the identity matrix; the identity matrix I is penalized by x. i and y i Point-to-point distance between them; The first item Punishment x i and y i In y i The distance from the symmetrical point between the local planes defined at the given location to the surface.

[0061] These two matrices benefit from each other, resulting in an anisotropic covariance matrix. Due to the complex distribution of point cloud data, a GMM with an anisotropic covariance matrix can achieve more accurate distribution estimation and adapt more flexibly to various directional and shape features in the data. Penalty coefficient α ij An adaptive setting adds varying degrees of symmetry point-to-surface penalty based on the flatness of the local surface. This is achieved by calculating x. i and y i This is achieved by considering the curvature of the surrounding area. Specifically, when considering the geometric features of different regions in the point cloud, large symmetry point-to-surface penalties are mainly applied to regions where the local surface can be well approximated by a plane; when dealing with regions with large curvature or when effective normal estimation cannot be obtained due to noise interference, the robustness of the algorithm is improved by appropriately reducing the degree of application of the symmetry point-to-surface penalty. In other words, the smaller the curvature, the flatter the local surface, and α... ij The larger.

[0062] S2223 uses an improved logistic function to update the penalty coefficient α. ij , is represented as:

[0063]

[0064] Where, α max It is the maximum value of the penalty coefficient, and k is the slope of the logistic function, which controls the effect of curvature on the penalty coefficient α. ij The sensitivity of adjustment, C is x i and y i The average local curvature at a given point is calculated using the eigenvalues ​​of the covariance matrix of the neighboring points.

[0065] S2224, Given the GMM and observations, construct the first loss function for the global task using the probability density function.

[0066] This invention uses source point cloud x i After rigid transformation, given the GMM and observations, the negative log-likelihood function (first loss function) is expressed as:

[0067]

[0068] Through further derivation, we can obtain the following formulas for GMM components:

[0069]

[0070] Where t represents the translation vector.

[0071] It can be seen more intuitively that the first loss function combines point-to-point penalties and symmetric point-to-surface penalties. By combining the advantages of both metrics, it can achieve stronger representational power and a wider convergence region. The goal of this invention is to find the optimal rigid transformation that minimizes L(Θ), which is a challenging optimization problem. Therefore, we consider using the global search characteristics of evolutionary computation to find the optimal transformation parameters.

[0072] S230, the residual weights are adaptively adjusted using the M-estimators function to construct the second loss function for the local task.

[0073] When severely disturbed by noise and outliers, the estimation of local geometric information (such as normals and curvature) is easily affected, leading to local optima. Constructing an auxiliary task using a robust model that focuses on local geometric information holds promise for improving registration accuracy while maintaining robustness. Local task T l Constructed from an adaptive robust model that focuses on local geometric information, it can supplement more accurate local feature information by further mining narrow regions of the solution space by utilizing local surface geometric information.

[0074] S230 of the present invention includes:

[0075] S231, adaptive residual weights are obtained by adaptively adjusting the residual weights using a class of M-estimators functions. The class of M-estimators functions is expressed as follows:

[0076]

[0077] Where r is the residual, a is a parameter controlling the shape of the function, which starts from 0 and decreases by 0.05 in each generation until it reaches -2, and remains at -2 in subsequent generations; β is a parameter automatically set according to the density of the point cloud to be registered, so the value of β is different in different registration tasks. resolution is the average resolution of the two point clouds in the task;

[0078] S232, the second loss function for the local task is constructed using adaptive residual weights, expressed as:

[0079]

[0080] Wherein, residual r i (R,t) and r j (R,t) are as follows:

[0081]

[0082]

[0083]

[0084]

[0085] S300, Set the population optimization algorithm corresponding to the global task and the local task, as well as the individual representation, search space and initialization parameters in the population optimization algorithm, and initialize the population optimization algorithm;

[0086] In this step, the individual representations in the population optimization algorithms for both the global and local tasks are encoded as six-dimensional vectors, denoted as [θ]. x ,θ y ,θ z ,t x ,t y ,t z The first three dimensions [θ] x ,θ y ,θ z [] represents the rotation angle along each x, y, z axis, and the last three dimensions [t] x ,t y ,t z ] represents translation along each axis. The rotation matrix R and the translation vector t can be derived from [θ]. x ,θ y ,θ z ] and [t x ,t y ,t z ]Calculated;

[0087] The search space is set as follows: based on the rotation angle range along the x, y, and z axes, respectively, [-π, π]. The translation vector t has boundaries along all three axes, and its range is set to [-π, π], where bound is the longest side of the bounding box of the source point cloud and the target point cloud after decentralization.

[0088] Initialization parameters include population size pop and maximum number of iterations G. max Within the search space, randomly generate pop individuals to initialize the population optimization algorithm. Then, evaluate the initialized population after decoding and assign skill factors (indicating which task they perform better on) based on their fitness values.

[0089] S400, the transformation parameters of the point cloud data are optimized by using the population optimization algorithm corresponding to the global task and the transformation parameters of the point cloud data are optimized by using the population optimization algorithm corresponding to the local task, so as to obtain the optimal transformation parameters of the first loss function at the minimum value and the optimal transformation parameters of the second loss function at the minimum value respectively.

[0090] refer to Figure 3 In one specific embodiment of the present invention, S400 includes:

[0091] S410 decodes individuals from the two population optimization algorithms of the current generation to decode individuals into 4×4 rigid transformation matrices containing rotation matrix R and translation vector t;

[0092] S420: The first loss function is used to evaluate the 4×4 rigid transformation matrix corresponding to the global task, and the first loss function is used to evaluate the 4×4 rigid transformation matrix corresponding to the local task to obtain the evaluation results of each individual in the two populations.

[0093] This step decodes and evaluates the fitness of individuals within the subpopulation. In task T... g And Task T l In the subpopulation, individuals are decoded to generate a 4×4 rigid transformation matrix containing rotation and translation transformations. Then, task T is used respectively. g And Task T l The objective function is used to evaluate the fitness of the subpopulations.

[0094] S430, Based on the evaluation results, calculate the first adaptive crossover probability of the global task learning knowledge from the local task in the Gth generation and the second adaptive crossover probability of the local task learning knowledge from the global task in the Gth generation.

[0095] This invention completes the modeling of the 3D point cloud registration problem, namely the first loss function and the second loss function. Designing a knowledge-sharing mechanism for evolutionary multi-task optimization is also indispensable for achieving high-quality solutions. On the one hand, in evolutionary multi-task optimization, effective knowledge transfer between tasks can alleviate the situation where a single task easily gets trapped in local optima, improving the quality of the final result. On the other hand, effective knowledge transfer can help tasks converge faster, reducing the occurrence of situations where the optimal value has not yet been reached due to insufficient iterations. In this invention, considering that when the global task has fast convergence, it indicates that a promising region is being explored, self-sharing within the task should be strengthened, and knowledge transfer between tasks should be reduced; otherwise, more accurate local geometric information should be supplemented to obtain better global structural information. Similarly, when a local task has good convergence, its own knowledge should be used for a more thorough search; otherwise, more accurate point cloud distribution information should be provided to help the optimization process of the local task. To reasonably control the probability of self-evolution and knowledge transfer, empirical knowledge during the evolution process is used to estimate the convergence trend of the current task. The current convergence trend of the task is estimated by comparing the optimal objective function value of the current population with the objective function values ​​of the previous two generations.

[0096] In one specific embodiment of the present invention, S430 includes:

[0097] S431, for the current generation G of the population optimization algorithm, based on the evaluation results, calculate the maximum function value of the task in the current generation and the previous two generations, denoted as follows: and

[0098] S432, Calculate the difference between the optimal function values ​​of the three generations based on the optimal function values ​​of the three generations;

[0099] The difference in the optimal function values ​​among the three generations is expressed as follows:

[0100]

[0101] S433, using the difference value of the global task, calculate the first adaptive probability of the global task learning knowledge from the local task in the Gth generation, and using the difference value of the local task, calculate the second adaptive probability of the local task learning knowledge from the global task in the Gth generation.

[0102] Among them, global task T g The first adaptive probability of learning knowledge from the local task in generation G is:

[0103]

[0104] Local task T l The second adaptive probability for learning knowledge from the global task in generation G is:

[0105]

[0106] S440, generate a random number between 0 and 1, determine whether the random number is less than the first adaptive crossover probability and the second adaptive crossover probability. If it is, perform crossover and mutation operations between the global task and the population corresponding to the global task to update the current generation population and obtain two updated populations. Otherwise, perform crossover within the corresponding task population.

[0107] This step uses the AMP (amp) to determine whether to perform self-evolution with a specific evolutionary algorithm or knowledge transfer. Knowledge transfer involves transferring promising individuals and adding new individuals through minor mutations in those promising individuals. A dynamic knowledge-sharing mechanism is designed to adaptively adjust the knowledge sharing between global and local tasks during the evolutionary process based on the registration results. This mechanism enables the effective fusion of global structural information and local geometric information, ensuring the final registration result reaches its optimal level.

[0108] S450: Evaluate each individual in the two updated populations to obtain the evaluation result of each individual in the two populations. Select the top-ranked individuals from the two updated populations to form the next generation population of the updated population.

[0109] For each task, the current individual and its offspring individuals are selected as elites based on their fitness values. The best individuals are popped to form the next generation population, and the currently found optimal individuals are updated.

[0110] S460, repeat S410 to S450 until the termination condition is met to obtain the optimal population, and select the best individual from the optimal population;

[0111] During population iteration, determine if the termination condition is met. If the current iteration number is less than or equal to the preset maximum iteration number G. max If the result is positive, the process continues to the next evolutionary stage; otherwise, the iteration loop terminates and task T is output. g The optimal solution found in the search is used as the final point cloud transformation parameters.

[0112] S470, determine the optimal transformation parameters corresponding to the optimal individual of the global task and the optimal transformation parameters corresponding to the optimal individual of the local task.

[0113] S500, the preprocessed point cloud data is transformed using the optimal transformation parameters of the first loss function at its minimum value.

[0114] Secondly, the present invention provides a probabilistic point cloud registration system based on evolutionary multi-task, comprising:

[0115] The acquisition module is configured to acquire point cloud data and preprocess the point cloud data to obtain preprocessed point cloud data.

[0116] The building module is configured to construct the first loss function for the global task and the second loss function for the local task;

[0117] The setting module is configured to set the population optimization algorithm corresponding to the global task and the local task, as well as the individual representation, search space and initialization parameters in the population optimization algorithm, and initialize the population optimization algorithm.

[0118] The optimization module is configured to optimize the transformation parameters of the point cloud data using the population optimization algorithm corresponding to the global task and the population optimization algorithm corresponding to the local task, so as to obtain the optimal transformation parameters of the first loss function at the minimum value and the optimal transformation parameters of the second loss function at the minimum value, respectively.

[0119] The transformation module is configured to transform the preprocessed point cloud data using the optimal transformation parameters of the first loss function at its minimum value.

[0120] The beneficial effects of the present invention will be verified through experiments below.

[0121] The solution of this invention is implemented and verified on various public datasets. For point cloud registration datasets, experiments are conducted on synthetic object datasets (EPFL statue dataset, FGR dataset) and real scene datasets (ETH laser registration dataset), mainly focusing on noisy data registration experiments, outlier data registration experiments, and real scene registration experiments.

[0122] Example 1

[0123] Experiments were conducted using the method proposed in this invention on point clouds with different levels of Gaussian noise. To obtain the Gaussian noise point cloud, the original point cloud was first normalized so that the coordinates of each point on the x, y, and z axes were all within the range [-1, 1]. Then, Gaussian noise with a mean of zero and a standard deviation of σ was directly added to the normalized coordinates of each point. Finally, the contaminated point cloud was restored to its original scale through a reverse normalization operation. For more comprehensive verification, this invention set σ from smallest to largest to 0.01, 0.02, 0.03, 0.04, and 0.05, obtaining point clouds to be registered with different levels of noise contamination.

[0124] In this simulation, the population size for each task is set to pop = 100, and the maximum number of iterations is G. maxThe probability parameter rmp for knowledge sharing is set to 60, and its initial value is 0.3. A genetic algorithm is used as the self-evolutionary mechanism, with the simulated binary crossover parameter Mu = 10 and the Gaussian mutation standard deviation Sigma = 0.02. The data registration results proposed in this patent for noise pollution are shown in the table below:

[0125] Registration results under different levels of noise pollution.

[0126]

[0127] To more intuitively illustrate the registration effect of this method, the final registration visualization results for σ values ​​of 0.01, 0.03, and 0.05 are shown below. Figure 4 As shown.

[0128] Example 2

[0129] To highlight the robustness of the proposed method for point cloud data with high outlier rates, in the EPFL dataset, the first 60% of points in the complete model are used as the source point cloud, and the last 60% as the target point cloud, forming a partially overlapping pair of point clouds to be registered. To simulate outliers, η·N random points are uniformly generated from the bounding box of the source set, where η is set to 1%, 10%, 40%, 70%, 100%, 125%, and 150%, respectively. The results are then compared with eight advanced registration methods:

[0130] (1)ICP: Besl PJ, McKay N DA method for registration of 3-D shapes[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1992, 14(02): 239-256.

[0131] (2)TrICP: Chetverikov D, Stepanov D, Krsek P. Robust Euclidean alignment of 3D point sets: the trimmed iterative closest point algorithm [J]. Image andVision Computing, 2005, 23(3): 299-309.

[0132] (3)RSICP:Li J,Hu Q,Zhang Y,et al.Robust symmetric iterative closestpoint[J].ISPRS Journal of Photogrammetry and Remote Sensing,2022,185:219-231.

[0133] (4)FCPSR:Liao Q,Sun D,Andreasson H.Point set registration for 3Drange scans using fuzzy cluster-based metric and efficient globaloptimization[J].IEEE transactions on pattern analysis and machineintelligence,2020,43(9):3229-3246.

[0134] (5)LSGCPD:Liu W,Wu H,Chirikjian G S.LSG-CPD:Coherent point drift withlocal surface geometry for point cloud registration[C] / / Proceedings of theIEEE / CVF International Conference on Computer Vision.2021:15293-15302.

[0135] (6)EDAVB:Cao H,He Q,Wang H,et al.An estimation of distributionalgorithm based on variational bayesian for point-set registration[J].IEEETransactions on Evolutionary Computation,2021,26(5):926-940.

[0136] (7)EMFPCR: Wu Y, Ding H, Gong M, et al. Evolutionary multiform optimization with two-stage bidirectional knowledge transfer strategy for point cloud registration[J]. IEEE Transactions on Evolutionary Computation, 2022.

[0137] (8)EMTR-SSC: Wu Y, Gong P, Gong M, et al. Evolutionary multitasking with solution space cutting for point cloud registration[J]. IEEE Transactions on Emerging Topics in Computational Intelligence, 2023.

[0138] In this simulation, the population size for each task is set to pop = 100, and the maximum number of iterations is G. max The probability parameter rmp for knowledge sharing was set to 60 and initialized to 0.3. A genetic algorithm was used for self-evolution, with the simulated binary crossover parameter Mu = 10 and the Gaussian mutation standard deviation Sigma = 0.02. The registration results of this method and the compared method on data containing different outliers are shown in the table below:

[0139]

[0140] To more intuitively demonstrate the significant advantages of the method of this invention in processing outlier data, this invention visualizes the registration results with outlier rates of 1%, 70%, and 150%, as shown below. Figure 5 As shown.

[0141] exist Figure 5 The right half of the image is the registration map, and the left half is its corresponding logarithmic scale color-coded map. The more colors below the color bar, the better the registration effect, and vice versa. The method proposed in this invention can effectively handle the point cloud registration problem with high outliers and has good robustness.

[0142] Example 3

[0143] Compared to object-scale point cloud registration, scene-scale point cloud registration is more challenging. Due to the sheer volume of point clouds, many objects can be captured within the same cloud. Furthermore, point densities vary significantly across different regions of the scene, and the surfaces of different objects also differ considerably.

[0144] In this simulation, the population size for each task is set to pop = 100, and the maximum number of iterations is G. max The probability parameter rmp for knowledge sharing was set to 60 and initialized to 0.3. A genetic algorithm was used for self-evolution, with the simulated binary crossover exponent Mu = 10 and the Gaussian mutation standard deviation Sigma = 0.02. Experiments were then conducted on three scenes—Arch, Office, and Courtyard—from the real-world scene dataset (ETH laser registration dataset). The registration visualization results are shown below. Figures 6-8 As shown, they are all point clouds with millions of points.

[0145] This invention addresses the limitation of probabilistic registration methods by neglecting rich local geometric features, thus restricting the model's representational ability. It establishes an efficient and robust point cloud registration scheme that adaptively combines point-to-point penalties with symmetric point-to-surface penalties. Furthermore, to address the problem of local optima caused by inaccurate local geometric information under severe noise and outlier interference, a dynamic evolutionary multi-task registration framework is designed. This invention achieves a novel point cloud registration technique by utilizing evolutionary multi-tasks. It better considers both global structural and local geometric information, exhibits excellent robustness to noise, outliers, and partial overlap, and can be successfully applied to large-scene point cloud registration problems.

[0146] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0147] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0148] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A probabilistic point cloud registration method based on evolutionary multi-task, characterized in that, include: S100: Acquire point cloud data and preprocess the point cloud data to obtain preprocessed point cloud data; S200, construct the first loss function for the global task and the second loss function for the local task; S300, Set the population optimization algorithm corresponding to the global task and the local task, as well as the individual representation, search space and initialization parameters in the population optimization algorithm, and initialize the population optimization algorithm; S400, the transformation parameters of the point cloud data are optimized by using the population optimization algorithm corresponding to the global task and the transformation parameters of the point cloud data are optimized by using the population optimization algorithm corresponding to the local task, so as to obtain the optimal transformation parameters of the first loss function at the minimum value and the optimal transformation parameters of the second loss function at the minimum value respectively. S500, the preprocessed point cloud data is transformed using the optimal transformation parameters of the first loss function at its minimum value; S200 includes: S210, Construct a multi-task environment; the multi-task environment includes two related tasks, the two related tasks focus on different solution spaces, the two related tasks are divided into task one and task two, task one is a global task used to obtain global structure information, and task two is a local task used to capture local set information. S220, using an adaptive symmetric point-to-surface Gaussian mixture model, constructs the first loss function for the global task; S230, The residual weights are adaptively adjusted using the M-estimators function to construct the second loss function for the local task; S220 includes: S221, a rigid transformation of the source point cloud is performed using a Gaussian mixture model, and the transformed source point cloud is regarded as the observation value of the GMM; S222, Given the GMM and the observations, construct the first loss function for the global task; S222 includes: S2221, Obtain the probability density function of the Gaussian Mixture Model (GMM), expressed as: in, for x i Outlier weights, U ( N ) indicates that the outlier probability density is uniformly distributed. M , N These represent the number of points after downsampling the target point cloud and the source point cloud, respectively, excluding those centered at each point in the target set. M In addition to the Gaussian components, the GMM also includes an outlier component. For the corresponding component density function, These are the corresponding mean and covariance matrices, respectively; S2222, using the probability density function, design the specific form of the corresponding component density function, expressed as follows: in, The dimension of a random variable is represented by the number of points in point cloud registration. , It is a source point cloud. It is a target point cloud. It is the covariance multiplier. It is the penalty coefficient. It is a rotation matrix. These are the surface normals of the source point cloud and the target point cloud, respectively. It is the identity matrix; The first item punish x i and y i exist y i The distance from the symmetric point between the local planes defined at the given location to the plane, and the penalty of the identity matrix I. x i and y i Point-to-point distance between them; S2223 utilizes an improved logistic function to update the penalty coefficient. , represented as: in, It is the maximum value of the penalty coefficient. k It is the slope of the logistic function, which controls the effect of curvature on the penalty coefficient. Adjusting sensitivity, yes x i and y i The average local curvature at that location; S2224, Given the GMM and observations, and using the probability density function to construct the first loss function for the global task, it is expressed as: in, Represents the translation vector; S230 includes: S231, adaptive residual weights are obtained by adaptively adjusting the residual weights using a class of M-estimators functions. The class of M-estimators functions is expressed as follows: in, It is a residual. These are parameters that control the shape of the function. Starting from 0, decrease by 0.05 in each generation until it reaches -2, and remain at -2 in subsequent generations; These parameters are automatically set based on the density of the point cloud to be registered. , The average resolution of the two point clouds in the task; S232, the second loss function for the local task is constructed using adaptive residual weights, expressed as: Among them, residual and They are as follows: 。 2. The probabilistic point cloud registration method based on evolutionary multi-task as described in claim 1, characterized in that, S100 includes: Preprocessed point cloud data is obtained by downsampling and decentralizing the point cloud data.

3. The probabilistic point cloud registration method based on evolutionary multi-task as described in claim 2, characterized in that, In the S300 population optimization algorithms for both global and local tasks, individual representations are encoded using six-dimensional vectors. The first three dimensions This represents the rotation angle along each of the x, y, and z axes, the last three dimensions. Representing translation along each axis, the rotation matrix R and the translation vector t can be derived from... and Calculated; The search space is set as follows: based on the rotation angle range along the x, y, and z axes respectively. , and The translation vector t has boundaries along all three axes, and its range is set to... ,in bound The longest edge of the bounding box between the source point cloud and the target point cloud after decentralization; Initialization parameters include population size pop and maximum number of iterations G max ; Randomly generated within the search space. pop Each individual is used to complete the initialization of the population optimization algorithm.

4. The probabilistic point cloud registration method based on evolutionary multi-task as described in claim 3, characterized in that, The S400 includes: S410 decodes individuals from the two population optimization algorithms of the current generation to decode individuals into 4×4 rigid transformation matrices containing rotation matrix R and translation vector t; S420: The first loss function is used to evaluate the 4×4 rigid transformation matrix corresponding to the global task, and the first loss function is used to evaluate the 4×4 rigid transformation matrix corresponding to the local task to obtain the evaluation results of each individual in the two populations. S430, based on the evaluation results, calculate the global task from the [missing information]. G The first adaptive crossover probability of local task learning knowledge and the local task from the first G The second adaptive crossover probability for learning knowledge on a global task; S440, generate a random number between 0 and 1, and determine whether the random number is less than the first adaptive crossover probability and the second adaptive crossover probability. If it is, perform crossover and mutation operations between the global task and the population corresponding to the global task to update the current generation population and obtain two updated populations. Otherwise, perform crossover within the corresponding task population. S450: Evaluate each individual in both updated populations to obtain the evaluation result for each individual in both populations, and select the individual with the highest evaluation result from each updated population. pop Each individual constitutes the next generation of the updated population; S460, repeat S410 to S450 until the termination condition is met to obtain the optimal population, and select the best individual from the optimal population; S470, determine the optimal transformation parameters corresponding to the optimal individual of the global task and the optimal transformation parameters corresponding to the optimal individual of the local task.

5. The probabilistic point cloud registration method based on evolutionary multi-task as described in claim 4, characterized in that, The S430 includes: S431, for the current generation of population optimization algorithms G Based on the evaluation results, the maximum function values ​​of the task in the current generation and the two previous generations are calculated and denoted as follows: , and ; S432, Calculate the difference between the optimal function values ​​of the three generations based on the optimal function values ​​of the three generations; The difference in the optimal function values ​​among the three generations is expressed as follows: ， S433, using the difference value of the global task, calculate the global task from the first... G The first adaptive probability of learning knowledge from local tasks and the calculation of the difference value of local tasks from the first... G The second adaptive probability for learning knowledge on behalf of the global task; The global task T g From the G The first adaptive probability for learning knowledge on a local task is: Local task T l From the G The second adaptive probability for learning knowledge on behalf of the global task is: 。 6. A probabilistic point cloud registration system based on evolutionary multi-task, characterized in that, include: The acquisition module is configured to acquire point cloud data and preprocess the point cloud data to obtain preprocessed point cloud data. The building module is configured to construct the first loss function for the global task and the second loss function for the local task; The setting module is configured to set the population optimization algorithm corresponding to the global task and the local task, as well as the individual representation, search space and initialization parameters in the population optimization algorithm, and initialize the population optimization algorithm. The optimization module is configured to optimize the transformation parameters of the point cloud data using the population optimization algorithm corresponding to the global task and the population optimization algorithm corresponding to the local task, so as to obtain the optimal transformation parameters of the first loss function at the minimum value and the optimal transformation parameters of the second loss function at the minimum value, respectively. The transformation module is configured to transform the preprocessed point cloud data using the optimal transformation parameters of the first loss function at its minimum value. S200 includes: S210, Construct a multi-task environment; the multi-task environment includes two related tasks, the two related tasks focus on different solution spaces, the two related tasks are divided into task one and task two, task one is a global task used to obtain global structure information, and task two is a local task used to capture local set information. S220, using an adaptive symmetric point-to-surface Gaussian mixture model, constructs the first loss function for the global task; S230, The residual weights are adaptively adjusted using the M-estimators function to construct the second loss function for the local task; S220 includes: S221, a rigid transformation of the source point cloud is performed using a Gaussian mixture model, and the transformed source point cloud is regarded as the observation value of the GMM; S222, Given the GMM and the observations, construct the first loss function for the global task; S222 includes: S2221, Obtain the probability density function of the Gaussian Mixture Model (GMM), expressed as: in, for x i Outlier weights, U ( N ) indicates that the outlier probability density is uniformly distributed. M , N These represent the number of points after downsampling the target point cloud and the source point cloud, respectively, excluding those centered at each point in the target set. M In addition to the Gaussian components, the GMM also includes an outlier component. For the corresponding component density function, These are the corresponding mean and covariance matrices, respectively; S2222, using the probability density function, design the specific form of the corresponding component density function, expressed as follows: in, The dimension of a random variable is represented by the number of points in point cloud registration. , It is a source point cloud. It is a target point cloud. It is the covariance multiplier. It is the penalty coefficient. It is a rotation matrix. These are the surface normals of the source point cloud and the target point cloud, respectively. It is the identity matrix; The first item punish x i and y i exist y i The distance from the symmetric point between the local planes defined at the given location to the plane, and the penalty of the identity matrix I. x i and y i Point-to-point distance between them; S2223 utilizes an improved logistic function to update the penalty coefficient. , represented as: in, It is the maximum value of the penalty coefficient. k It is the slope of the logistic function, which controls the effect of curvature on the penalty coefficient. Adjusting sensitivity, yes x i and y i The average local curvature at that location; S2224, Given the GMM and observations, and using the probability density function to construct the first loss function for the global task, it is expressed as: in, Represents the translation vector; S230 includes: S231, adaptive residual weights are obtained by adaptively adjusting the residual weights using a class of M-estimators functions. The class of M-estimators functions is expressed as follows: in, It is a residual. These are parameters that control the shape of the function. Starting from 0, decrease by 0.05 in each generation until it reaches -2, and remain at -2 in subsequent generations; These parameters are automatically set based on the density of the point cloud to be registered. , The average resolution of the two point clouds in the task; S232, the second loss function for the local task is constructed using adaptive residual weights, expressed as: Among them, residual and They are as follows: 。

Images