Probability multi-scale-based missile loading vehicle tail end relative attitude estimation method and system

Abstract

The invention provides a probabilistic multi-scale-based missile loading vehicle tail end relative attitude estimation method and system, and belongs to the technical field of aviation equipment automation. Comprising the following steps: firstly, performing coordinate system conversion and noise suppression processing on acquired point cloud data, and constructing a probability multi-scale consistency model; then, on the basis of weighted features obtained by the probability multi-scale consistency model, density self-adaptive multi-scale feature fusion is executed; and finally, according to the multi-scale features after feature fusion, performing dual-stage pose estimation. Through dynamic noise suppression, component-level multi-scale feature modeling and a real-time optimization strategy, millimeter-level and high-reliability alignment between the end effector of the missile loading vehicle and the missile loading frame is achieved.

Description

Technical Field

[0001] This invention belongs to the field of aviation equipment automation technology, and in particular relates to a method and system for estimating the terminal relative attitude of a missile-carrying vehicle based on probabilistic multi-scale. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] During aircraft missile loading and assembly, precise alignment between the missile hooks and the missile racks on the loading vehicle is crucial. However, traditional methods rely primarily on manual adjustment using a level, which suffers from low accuracy, slow response, and susceptibility to human error, failing to meet the demands of modern, high-precision assembly. While existing technologies have incorporated visual sensors and point cloud registration algorithms to assist attitude estimation, they still face numerous challenges in practical applications, leading to difficulties in ensuring assembly efficiency and quality.

[0004] Specifically, the existing technology has the following main drawbacks: (1) Poor adaptability to dynamic interference. The assembly workshop is generally subject to complex environmental factors such as mechanical vibration and changes in lighting (such as strong light reflection from the overhead lights in the engine room), which can easily cause significant noise in the point cloud data collected by the vision sensor, resulting in large attitude estimation errors of the registration algorithm under such interference.

[0005] (2) Insufficient utilization of key component features. Components such as missile racks and missile hooks have complex multi-scale geometric characteristics. Existing deep learning models cannot effectively integrate global contour and local detail information, resulting in limited accuracy of component-level attitude estimation.

[0006] To address the aforementioned issues, there is an urgent need for a method for estimating the relative attitude at the end of the armature loading vehicle, in order to achieve rapid, stable, and precise alignment during the armature loading assembly process. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a method and system for estimating the relative attitude of the end effector of a missile loading vehicle based on probabilistic multi-scale. Through dynamic noise suppression, component-level multi-scale feature modeling and real-time optimization strategies, millimeter-level, highly reliable alignment between the end effector of the missile loading vehicle and the missile loading rack is achieved.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a method for estimating the relative attitude of a missile-carrying vehicle at the end of its trajectory based on probabilistic multi-scale methods.

[0009] A probabilistic multi-scale method for estimating the terminal relative attitude of a missile-carrying vehicle includes: Acquire point cloud data of the end effector and the bomb rack of the bomb-loading vehicle, and perform coordinate system transformation and noise suppression processing; A probabilistic multi-scale consistency model is constructed based on point cloud data after noise suppression. Based on the weighted features obtained from the probabilistic multi-scale consistency model, density-adaptive multi-scale feature fusion is performed. Specifically, the matching constraint strength is dynamically adjusted according to the local density distribution of the point cloud; hierarchical features at high, medium and low scales are fused, and distance constraints are integrated through maximization operations. Based on the multi-scale features after feature fusion, a two-stage pose estimation is performed. Specifically, candidate pose sets are filtered by matching probability thresholds, and the six-degree-of-freedom pose parameters are iteratively optimized by combining component priority. The pose estimation results corresponding to the six-degree-of-freedom pose parameters are sent to the robotic arm control unit to achieve alignment between the end effector of the bomb-loading vehicle and the bomb rack.

[0010] Furthermore, the noise suppression process includes: using a statistical outlier filter to remove noise points that deviate from a predetermined standard deviation threshold of the local mean.

[0011] Furthermore, a probabilistic multi-scale consistency model is constructed, including: probabilistically modeling point cloud features using a Gaussian mixture model, calculating the confidence level of each feature, and dynamically allocating attention weights based on the obtained confidence levels.

[0012] Furthermore, feature extraction of the Gaussian mixture model is achieved through an encoder, which employs a hierarchical feature extraction strategy, specifically including an input layer, a shallow feature extraction layer, a middle feature extraction layer, and a deep feature extraction layer.

[0013] Furthermore, the fusion of hierarchical features at different scales is performed through a feature pyramid network. The execution logic of the feature pyramid network includes: a bottom-up path for extracting high, medium, and low-level feature maps; a top-down path for transmitting high-level semantic features through upsampling; and a lateral connection path for fusing features from encoders and decoders at the same scale.

[0014] Furthermore, the output of the feature pyramid network is reconstructed by a decoder. The decoder is based on the principle of symmetry, and each decoding layer establishes a skip connection with the corresponding encoding layer. The skip connection fusion method includes three strategies: feature concatenation, element addition, and gated fusion. The optimal fusion method is selected through adaptive weight learning.

[0015] Furthermore, the candidate pose set is selected by matching probability threshold as the first stage of the two-stage pose estimation, which is achieved by sparse convolution under the parallelized RANSAC framework, that is, based on coarse matching geometric constraints, multiple candidate relative pose assumptions are generated; the six-degree-of-freedom pose parameters are iteratively optimized by component priority as the second stage of the two-stage pose estimation, which is achieved by using a nonlinear least squares optimization method.

[0016] The second aspect of the present invention provides a probabilistic multi-scale system for estimating the relative attitude of a missile-carrying vehicle at the end of its trajectory.

[0017] A probabilistic multi-scale system for estimating the terminal relative attitude of a missile-carrying vehicle includes: The point cloud data acquisition module is configured to acquire point cloud data of the end effector and the bomb rack of the bomb loading vehicle, and perform coordinate system transformation and noise suppression processing. The probabilistic multi-scale consistency modeling module is configured to: construct a probabilistic multi-scale consistency model based on noise-suppressed point cloud data; The feature fusion module is configured to perform density-adaptive multi-scale feature fusion based on the weighted features obtained from the probabilistic multi-scale consistency model. Specifically, it dynamically adjusts the matching constraint strength according to the local density distribution of the point cloud; it fuses hierarchical features at high, medium, and low scales and integrates distance constraints through a maximization operation. The pose estimation module is configured to perform two-stage pose estimation based on the multi-scale features after feature fusion. Specifically, it filters the candidate pose set by matching probability thresholds and iteratively optimizes the six-degree-of-freedom pose parameters by combining component priority. The pose estimation results corresponding to the six-degree-of-freedom pose parameters are sent to the robotic arm control unit to achieve alignment between the end effector of the bomb carrier and the bomb rack. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps in the probabilistic multi-scale-based method for estimating the terminal relative attitude of a missile-carrying vehicle as described in the first aspect of the present invention.

[0018] The fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the probabilistic multi-scale-based method for estimating the terminal relative attitude of a missile-carrying vehicle as described in the first aspect of the present invention.

[0019] The above one or more technical solutions have the following beneficial effects: (1) In the process of sequentially attenuating high, medium and low frequency vibrations, this invention combines a statistical outlier filter to remove noise points that deviate from the predetermined standard deviation threshold of the local mean, thereby achieving dynamic noise suppression of point cloud data. At the same time, by using a Gaussian mixture model to probabilistically model the point cloud features, calculate the feature confidence and dynamically allocate attention weights, the interference of low-confidence noise features can be effectively suppressed. The above measures significantly improve the system's adaptability to dynamic interference such as mechanical vibration and illumination changes, and reduce attitude estimation errors in complex environments.

[0020] (2) This invention constructs a probabilistic multi-scale consistency model, which integrates features at different scales (high, medium, and low) through a feature pyramid network. It extracts multi-scale feature maps from a bottom-up path, transmits high-level semantic features from a top-down path, and integrates features of the same scale through lateral connections. Furthermore, it dynamically adjusts the matching constraint strength using a density adaptive mechanism to maximize the integration of distance constraints. This strategy effectively integrates the global contour and local details of components such as missile racks and missile hooks, fully utilizing their complex multi-scale geometric characteristics to significantly improve component-level attitude estimation accuracy and achieve millimeter-level, highly reliable alignment between the end effector of the missile rack and the missile rack.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a structural flowchart of the PMSC-Net network in Embodiment 1 of the present invention.

[0024] Figure 2 This is a schematic diagram of the bimodal probability density distribution of the feature space in Embodiment 1 of the present invention.

[0025] Figure 3 This is a schematic diagram illustrating the result of the multi-scale feature fusion process in Embodiment 1 of the present invention; wherein, Figure 3 (a) in the diagram is a schematic diagram of geometric detail encoding. Figure 3 (b) in the diagram is a schematic diagram of the component structure modeling. Figure 3 (c) in the diagram is a schematic representation of the global topology. Figure 3 (d) in the figure is a schematic diagram of cross-scale feature fusion. Detailed Implementation

[0026] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0027] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0028] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0029] Example 1 This embodiment discloses a method for estimating the relative attitude of a missile-carrying vehicle at the end of its trajectory based on probabilistic multi-scale methods.

[0030] A probabilistic multi-scale method for estimating the terminal relative attitude of a missile-carrying vehicle includes: Step S1: Obtain point cloud data of the end effector and the bomb rack of the bomb-loading vehicle, and perform coordinate system transformation and noise suppression processing; Step S2: Construct a probabilistic multi-scale consistency model based on the noise-suppressed point cloud data; Step S3: Based on the weighted features obtained from the probabilistic multi-scale consistency model, perform density-adaptive multi-scale feature fusion. Specifically, dynamically adjust the matching constraint strength according to the local density distribution of the point cloud; fuse hierarchical features at high, medium, and low scales, and integrate distance constraints through maximization operations. Step S4: Based on the multi-scale features after feature fusion, perform two-stage pose estimation. Specifically: filter the candidate pose set by matching probability threshold and iteratively optimize the six-degree-of-freedom pose parameters by combining component priority; send the pose estimation results corresponding to the six-degree-of-freedom pose parameters to the robotic arm control unit to achieve alignment between the end effector of the bomb carrier and the bomb rack.

[0031] The probabilistic multi-scale relative attitude estimation method for the end effector of a missile-carrying vehicle provided by this invention, through dynamic noise suppression, component-level multi-scale feature modeling, and real-time optimization strategies, can overcome the problems of mechanical vibration, illumination interference, and multi-scale geometric adaptation in industrial environments, achieving millimeter-level accuracy and high reliability alignment. The main technical solutions of this method include: 1) Probabilistic Awareness Dynamic Attention Mechanism. This mechanism includes feature uncertainty quantification and dynamic attention weight allocation. Feature uncertainty quantification involves probabilistically modeling point cloud features using a Gaussian Mixture Model (GMM), decomposing the feature distribution into multiple Gaussian components, and calculating the posterior probability to quantify the confidence of each feature. Dynamic attention weight allocation involves dynamically adjusting the attention weights based on feature confidence to suppress noise responses from low-confidence features and enhance geometric consistency matching in high-confidence regions.

[0032] 2) Density-adaptive multi-scale feature fusion. This fusion process includes local density threshold adjustment and hierarchical feature complementarity strategy. Local density threshold adjustment involves dynamically adjusting the matching threshold based on the local density of the point cloud, relaxing constraints in sparse regions to reduce false filtering, and strengthening geometric consistency in dense regions to improve accuracy. The hierarchical feature complementarity strategy combines the distance constraints of high, medium, and low-level features, and achieves progressive fusion through maximization operations, preserving fine-grained geometric details while enhancing global topological alignment.

[0033] 3) Integrating a two-stage optimization framework and a joint training mechanism for feature reweighting and end-to-end optimization. The two-stage optimization framework involves: in the first stage, filtering potential matching regions using a relaxed threshold; and in the second stage, refining the matching by combining hierarchical feature weights, prioritizing the use of high-dimensional semantic features to drive global alignment. The joint training mechanism simultaneously optimizes feature extraction, probabilistic modeling, and transformation matrix estimation, improving the network's overall geometric consistency modeling capability.

[0034] 4) A real-time feedback and pose optimization strategy integrating a two-stage pose correction and end-to-end real-time processing mechanism. The two-stage pose correction involves: a first stage of coarse matching, rapidly generating a candidate pose set based on a relaxed threshold (matching probability > 0.3), taking ≤ 10ms; and a second stage of fine optimization, iteratively optimizing the six-DOF parameters using component priority and performing hypothesis scoring through a parallelized RANSAC framework. The end-to-end real-time system integrates a lightweight sparse convolutional network (<1M parameters), with a single-frame processing time ≤ 30ms, and supports real-time data stream processing (frame rate ≥ 35Hz) from 4K resolution 3D cameras (such as Intel RealSense L515).

[0035] Based on the above methods and processes, millimeter-level, highly reliable alignment between the end effector of the ammunition loading vehicle and the ammunition loading rack can be achieved. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0036] In step S1, point cloud data of the end effector and the bomb rack of the bomb loading vehicle are acquired, and coordinate system transformation and noise suppression processing are performed.

[0037] Two lidar cameras are installed at the end of the robotic arm of the bomb-loading vehicle, with a camera spacing of 500mm, forming a binocular stereo vision acquisition device. The technical parameters of each camera are: measurement range 0.25-9 meters, depth accuracy ±1mm at a distance of 1 meter, frame rate 35Hz, and resolution 1024×768 pixels. Camera calibration is performed using a checkerboard calibration board for both intrinsic and extrinsic parameters. Intrinsic parameter calibration obtains the camera's focal length, principal point coordinates, and distortion coefficients, while extrinsic parameter calibration determines the relative positional relationship between the two cameras. The calibration accuracy requires a reprojection error of less than 0.5 pixels. After calibration, a transformation matrix is ​​established from the camera coordinate system to the robotic arm's base coordinate system.

[0038] In practice, the raw point cloud data needs to undergo preprocessing steps such as noise filtering, downsampling, normal vector estimation, and coordinate system transformation. As an optional implementation, noise filtering uses a statistical outlier filter to remove points that deviate from the local mean by more than two standard deviations; downsampling uses a voxel grid with a voxel size of 2 mm; normal vector estimation calculates the normal vector of each point based on nearest neighbor search with a search radius of 5 mm; finally, the point cloud is transformed from the camera coordinate system to the robotic arm's base coordinate system.

[0039] In step S2, a probabilistic multi-scale consistency model is constructed based on the noise-suppressed point cloud data. The point cloud features are probabilistically modeled using a Gaussian mixture model, the confidence level of each feature is calculated, and attention weights are dynamically allocated based on the obtained confidence levels.

[0040] like Figure 1 The diagram shows the structural flowchart of the PMSC-Net network. The PMSC-Net network includes an encoder-decoder structure, a Gaussian Mixture Model (GMM)-based information exchange network, and an adaptive multi-scale voting mechanism. The core design of the PMSC-Net network lies in fusing probabilistic modeling and multi-scale feature learning, achieving end-to-end mapping from raw point clouds to high-precision pose estimation through the encoder-decoder framework; furthermore, feature extraction of the Gaussian Mixture Model is achieved through the encoder.

[0041] The encoder employs a hierarchical feature extraction strategy, achieving progressive feature learning from geometric details to semantic understanding through layer-by-layer abstraction. Specifically, it includes an input layer, a shallow feature extraction layer, a mid-level feature extraction layer, and a deep feature extraction layer. 1) The input layer is responsible for initializing the point cloud data, fusing and encoding the spatial coordinates of the original point cloud with the calculated normal vector information. The normal vector calculation employs a principal component analysis (PCA) method, determining the local surface normal vector direction by analyzing the distribution characteristics of points in the neighborhood of each point. A normal vector consistency check is also introduced to ensure the accuracy of the direction. In the feature encoding process, in addition to basic geometric information, geometric descriptors such as local curvature and surface roughness are incorporated to enrich the geometric semantic content of the feature representation.

[0042] 2) Shallow Feature Extraction Layer: This layer focuses on capturing the local geometric structure features of the point cloud, including basic geometric elements such as edges, corners, and planes. It employs a multi-scale receptive field design, extracting local features at different scales in parallel using convolutional kernels of varying sizes. The multi-scale information is then integrated through feature concatenation or weighted fusion. This layer addresses the vanishing gradient problem in deep networks by introducing residual connections, enabling the network to learn incremental representations of features and improving feature learning efficiency. Simultaneously, a bottleneck structure is introduced within the residual block: first, 1×1×1 convolutions are used to reduce dimensionality and computation; then, 3×3×3 convolutions are used to extract features; and finally, 1×1×1 convolutions are used to restore dimensionality.

[0043] 3) The mid-level feature extraction layer captures spatial relationships between components and medium-scale geometric patterns. This layer introduces deformable convolution and spatial attention mechanisms. The deformable convolution mechanism allows the convolution kernel to adaptively adjust the sampling position based on input features, better adapting to irregular geometries such as missile racks and hooks. The spatial attention mechanism learns the importance weights of each position on the feature map, highlighting the feature responses of key geometric regions and suppressing interference from irrelevant background information. The calculation of attention weights considers not only spatial location information but also the correlation between feature channels. Joint optimization of channel attention and spatial attention enables more accurate feature selection.

[0044] 4) The deep feature extraction layer is responsible for learning high-level semantic features, including abstract information such as component categories, overall shape, and relative positional relationships. This layer employs a dense connection strategy, concatenating the feature maps from all previous layers as input to ensure sufficient flow of gradients and information. Global context modeling is achieved through a nonlocal neural network, calculating the correlation between any two points on the feature map to capture long-distance spatial dependencies. This global modeling capability is crucial for understanding the overall registration relationship between the missile rack and the missile hook.

[0045] In step S3, based on the weighted features obtained from the probabilistic multi-scale consistency model, density-adaptive multi-scale feature fusion is performed: the matching constraint strength is dynamically adjusted according to the local density distribution of the point cloud; hierarchical features at high, medium, and low scales are fused, and distance constraints are integrated through a maximization operation. Specifically, this can be achieved through the following methods: Step S3-1: Multi-scale feature fusion.

[0046] The fusion of hierarchical features at different scales is performed through a feature pyramid network. The execution logic of the feature pyramid network includes: a bottom-up path for extracting high, medium, and low-level feature maps; a top-down path for upsampling to pass high-level semantic features; and lateral connections for fusing features from the encoder and decoder at the same scale. Specifically, the design of the feature pyramid structure draws on successful experiences in object detection in computer vision, improving the model's ability to detect objects of different sizes by constructing multi-scale feature representations. The feature pyramid consists of three components: a bottom-up path, a top-down path, and lateral connections. The bottom-up path corresponds to the feature extraction process of the encoder, generating feature maps at different scales; the top-down path passes high-level semantic features to lower levels through upsampling; and the lateral connections fuse features from the encoder and decoder at the same scale.

[0047] The output of the feature pyramid network is reconstructed by the decoder, which, based on the principle of symmetry, gradually restores the spatial resolution of the feature map through transposed convolutions. Each decoding layer of the decoder establishes a skip connection with its corresponding encoding layer, fusing high-resolution detail features with low-resolution semantic features. The skip connection fusion methods include feature concatenation, element-wise addition, and gated fusion, with the optimal fusion method selected through adaptive weight learning. During feature fusion, a feature calibration module is introduced to recalibrate the fused features by learning the importance weights between channels, thereby improving the discriminative power of the feature representation.

[0048] Furthermore, the decoder integrates a multi-task learning framework to simultaneously predict the matching probability, feature descriptors, and geometric attributes of the point cloud. The matching probability prediction uses a sigmoid activation function to output the probability that each point belongs to a correct match; the feature descriptors are represented as L2-normalized high-dimensional vectors for subsequent feature matching; and the geometric attribute prediction includes surface normals, curvature, etc., to provide a basis for geometric consistency verification. The introduction of multi-task learning not only improves the network's feature learning ability but also enhances the interpretability of the prediction results.

[0049] Furthermore, sparse convolution is implemented based on a spatial hash data structure. By dividing the 3D space into a regular grid, convolution calculations are performed only in the grid cells containing point cloud data, significantly reducing memory usage and computational complexity. The core advantage of sparse convolution lies in its ability to handle point cloud data of arbitrary scale and density, avoiding the information loss problem in traditional voxelization methods. In the convolution operation, submanifold sparse convolution technology is employed to ensure that the sparsity of the output features remains consistent with the input, further improving computational efficiency.

[0050] Network regularization strategies include a combination of techniques such as batch normalization, dropout, and weight decay. Batch normalization not only accelerates network convergence but also improves the model's robustness to changes in input distribution. Adaptive dropout dynamically adjusts the dropout probability based on the training phase, using a higher dropout probability in the early stages of training to prevent overfitting and lowering the dropout probability in the later stages to maintain the model's expressive power. Weight decay constrains the complexity of model parameters through L2 regularization, preventing the model from overfitting the training data.

[0051] Gaussian Mixture Models (GMMs), as a core component of probabilistic modeling, are responsible for decomposing complex high-dimensional feature distributions into linear combinations of multiple Gaussian components, thereby achieving accurate quantification of feature uncertainty. In this embodiment, the fundamental goal of GMM probabilistic modeling is to identify high-confidence feature regions by learning probability distribution patterns in the feature space, and to provide probabilistic prior information for the subsequent dynamic attention mechanism. The core advantage of this method lies in its ability to adaptively handle noise and outliers in point cloud data, and to effectively suppress the interference of low-quality features through a probability weight allocation mechanism. Specifically: 3-1-1) Design of Intelligent GMM Model Initialization Strategy The initialization strategy of the GMM model has a significant impact on the final performance. The choice of the number of mixture components needs to strike a balance between model complexity and fitting accuracy. Too few components cannot adequately describe the complex feature distribution, while too many components may lead to overfitting and a sharp increase in computational complexity. Therefore, during initialization, an improved K-means++ clustering algorithm is first applied in the high-dimensional feature space. This algorithm ensures the diversity and representativeness of the initial cluster centers through an intelligent seed selection strategy. A feature weighting mechanism is introduced during clustering, assigning greater weight to features with higher geometric importance, making the initialization focus more on key geometric structures.

[0052] 3-1-2) Data-driven covariance matrix and mixture weight initialization Based on the clustering results in step 3-1-1), the initialization of the covariance matrix adopts a data-driven approach, determining the initial covariance matrix by analyzing the covariance structure of features within each cluster. To prevent the covariance matrix from degenerating into a singular matrix, a regularization mechanism is introduced, adding small regularization terms to the diagonal of the covariance matrix. The initialization of the mixed weights considers not only the cluster size but also the cluster compactness and separation information, ensuring that the model can better reflect the true feature distribution characteristics.

[0053] 3-1-3) Implementation of a numerically stable expectation-maximization algorithm After obtaining stable initial parameters, the Expectation-Maximization (EM) algorithm employs several improvement strategies to enhance convergence stability and computational efficiency. In the E-step posterior probability calculation, logarithmic space computation is used to avoid numerical underflow, and the log-sum-exp technique ensures numerical stability. An adaptive smoothing mechanism is also introduced in the posterior probability calculation to smooth extremely small probability values ​​and prevent certain components from being completely ignored. In the M-step parameter update, a momentum optimization strategy is employed, combining historical gradient information to accelerate the convergence process.

[0054] 3-1-4) Stability Guarantee Mechanism for High-Dimensional Space Covariance Matrix For the parameter update process in step 3-1-3), the numerical stability issue in the high-dimensional feature space is specifically considered. Cholesky decomposition is used to parameterize the covariance matrix, ensuring it always maintains positive definiteness. An adaptive regularization mechanism is also introduced, dynamically adjusting the regularization strength based on the condition number of the current covariance matrix, ensuring numerical stability while maintaining the model's expressive power. An entropy regularization term is added during the update of the mixed weights to prevent excessive concentration of weights in any component and promote diversity in the mixed components.

[0055] 3-1-5) Multi-criteria convergence judgment and global optimal solution search Based on the stable parameter updates in steps 3-1-3) and 3-1-4), the convergence of the EM algorithm is determined using multiple criteria, including the magnitude of parameter changes, the increment of the log-likelihood function, and the stability of the posterior probability distribution. The algorithm stops iterating when all criteria simultaneously satisfy the convergence condition. To address the local optimum problem, a multiple random initialization strategy is employed, running multiple independent EM processes, and selecting the result with the largest log-likelihood function as the final model.

[0056] 3-1-6) Multi-factor fusion feature confidence calculation method Based on the converged GMM model in steps 3-1-5), the calculation of feature confidence not only relies on the maximum a posteriori probability but also incorporates the entropy information of the probability distribution and the separation between components. Feature points with high confidence typically exhibit high probability on a certain Gaussian component, while this component also shows good separation from other components. The confidence calculation formula comprehensively considers multiple factors such as the sharpness of the probability peak, the concentration of the distribution, and neighborhood consistency, providing a more robust confidence assessment.

[0057] 3-1-7) Context-aware dynamic attention weight calculation Using the confidence information calculated in step 3-1-6), the dynamic attention mechanism is designed based on a confidence-driven feature reweighting strategy. It achieves adaptive feature selection by learning a non-linear mapping function from confidence to attention weights. The calculation of attention weights not only considers the confidence of the current feature point but also incorporates contextual information from neighboring features, capturing spatial correlations between features through a graph convolutional network. A gating mechanism is employed in the weight calculation process, allowing the model to dynamically adjust the attention allocation strategy based on the characteristics of the input features.

[0058] 3-1-8) Adaptive Multi-Strategy Feature Weighted Fusion Based on the attention weights in step 3-1-7), an adaptive selection mechanism for various fusion strategies is introduced during the feature weighted fusion process. In addition to basic element-wise multiplication, this includes additive fusion, gated fusion, and residual fusion, among others. The model learns the selection weights for each fusion method and chooses the most suitable fusion strategy for different types of features. The fusion process also considers the scale consistency of features, ensuring scale compatibility between features from different sources through adaptive normalization.

[0059] 3-1-9) Design of anisotropic density adaptive mechanism Building upon weighted feature fusion, the core design of multi-scale feature fusion lies in fully leveraging the complementarity of features at different levels, enhancing feature representation through information interaction between levels. The density adaptation mechanism considers not only the spatial density distribution of the point cloud but also incorporates the intensity distribution information of the feature responses. Local density calculation employs a multi-scale neighborhood analysis method, obtaining more robust density estimates through neighborhood statistics of different radii. The anisotropic characteristics of the point cloud are also considered during density calculation, replacing fixed spherical neighborhoods with adaptive ellipsoidal neighborhoods to better adapt to slender or flat geometric structures.

[0060] 3-1-10) Statistically Driven Adaptive Threshold Function Optimization Based on the density analysis results in step 3-1-9), the adaptive threshold function is designed using statistical analysis of density distribution. It achieves automatic parameter adjustment by learning the optimal mapping relationship between density and threshold. The threshold function employs a piecewise linear approximation, using different adjustment strategies in different density intervals to ensure good matching performance under various density distributions. The threshold adjustment process also incorporates a historical information feedback mechanism, dynamically adjusting the threshold parameters based on the quality of historical matching results to achieve adaptive optimization.

[0061] 3-1-11) Attention-guided hierarchical feature pyramid fusion Based on the adaptive threshold adjustment in step 3-1-10), the hierarchical feature fusion strategy effectively integrates multi-scale information by constructing a feature pyramid. High-level features focus on capturing global geometric semantic information, including macroscopic features such as overall shape and relative positional relationships; mid-level features are responsible for capturing the connectivity relationships of local geometric structures, such as edges, corners, and planar connections; low-level features retain the finest-grained geometric details, including local features such as surface texture and minor geometric changes. The feature fusion process employs an attention-weighted hierarchical fusion strategy, achieving adaptive fusion by learning the importance weights of features at different levels. The calculation of fusion weights considers multiple dimensions such as semantic consistency, geometric compatibility, and task relevance of features. A cross-scale feature interaction mechanism is also introduced during the fusion process, allowing features at different levels to directly exchange information, enhancing the richness of feature representations.

[0062] 3-1-12) Information Theory-Guided Optimization of Temporal Consistency Features After completing the hierarchical feature fusion in step 3-1-11), the complementarity analysis of the hierarchical features is achieved through feature correlation and information gain evaluation. The model automatically identifies redundant and unique information in features at different levels, maximizing the information content of the fused features through an information theory-guided feature selection mechanism. The fusion process also considers the temporal consistency of features, ensuring the stability of the feature fusion results between consecutive frames through a temporal smoothing mechanism, thereby improving the robustness of the system. The final output multi-scale fused features not only maintain the unique advantages of each level of features but also achieve the organic integration of information at different scales, providing high-quality feature representations for subsequent feature matching.

[0063] Step S3-2: Model training and optimization process.

[0064] The model training process was designed following best practices in deep learning, taking into account the unique properties of point cloud data and the specific requirements of the pose estimation task, to construct a complete training optimization workflow. The entire training process is divided into four main stages: data construction, network training, parameter optimization, and performance evaluation. Each stage contains multiple carefully designed sub-modules to ensure that the model can learn robust feature representations and accurate pose prediction capabilities. Specifically: 3-2-1) Strategies for Training Data Acquisition and Semantic Labeling The construction of training data is fundamental to the success of the model and needs to cover various scenarios and conditions that the missile loading vehicle may encounter during actual operation. The data acquisition strategy adopts a stratified sampling method to ensure the uniform distribution of training data in the attitude space. Specifically, this includes: data acquisition under different relative positional relationships, covering various situations such as normal alignment, slight offset, and large offset; data acquisition under different environmental conditions, including lighting changes such as normal illumination, strong light interference, and shadow occlusion; and data acquisition under different noise levels to simulate actual interference factors such as mechanical vibration and electronic noise.

[0065] Based on data collection, the semantic annotation of key component areas adopts a hierarchical annotation strategy, decomposing the missile rack and hook into multiple functional sub-regions. Each sub-region is assigned different weights according to its importance in the loading process, with key parts such as the lifting ring and hook contact surface being given the highest weight, while auxiliary parts such as connecting components and supporting structures have relatively lower weights. The annotation process also includes detailed recording of geometric attributes, such as surface roughness, material properties, and deformability, providing a basis for subsequent physical constraint modeling.

[0066] 3-2-2) Comprehensive Data Augmentation Strategy Oriented to Actual Working Conditions Based on the labeled dataset constructed in step 3-2-1), the data augmentation strategy was designed with full consideration of the characteristics of the shell loading vehicle's operating environment and the special properties of point cloud data. First, geometric transformation augmentation was implemented, including basic transformations such as random rotation, translation, and scaling. The selection of transformation parameters was based on statistical analysis of deviations during actual operations: rotation transformation simulated robotic arm positioning errors and workpiece clamping errors; translation transformation simulated vision system calibration errors and the influence of environmental vibrations; and scaling transformation simulated distance measurement errors and the influence of lens distortion.

[0067] Building upon geometric enhancement, a multi-level noise enhancement strategy is further implemented to simulate the effects of various real-world interference factors: Gaussian noise simulates random measurement errors of the sensor, salt-and-pepper noise simulates measurement anomalies caused by strong reflective or absorbing surfaces, and structured noise simulates the long-term effects of systematic measurement deviations such as temperature drift and mechanical loosening. The noise parameters are set based on statistical analysis of a large amount of actual measurement data to ensure the authenticity and representativeness of the enhanced data.

[0068] Simultaneously, illumination variation enhancement is implemented by altering the reflection intensity values ​​of the point cloud to simulate sensor response changes under different lighting conditions, including overall brightness adjustment, localized uneven illumination, strong light reflection, and shadow occlusion. Finally, point cloud density variation enhancement is achieved using random downsampling and upsampling techniques to simulate point cloud density changes under different distances, angles, and occlusion conditions, ensuring the model's adaptability to various input conditions.

[0069] 3-2-3) Construction of the joint loss function for pose regression and feature matching Based on the augmented dataset from step 3-2-2), the loss function is designed using a multi-task learning framework. By jointly optimizing multiple related tasks, the generalization ability and prediction accuracy of the model are improved. The main task is pose regression, and the auxiliary tasks include feature matching, confidence prediction, and geometric consistency verification. Knowledge transfer is achieved among the tasks by sharing feature representations.

[0070] The pose regression loss employs a composite distance metric, comprehensively considering the effects of translation and rotation errors. Translation errors are measured using Euclidean distance, while rotation errors are represented using quaternions and geodesic distance, avoiding the singularity problem inherent in Euler angle representation. The loss function also incorporates the robustness of Huber loss, penalizing large errors relatively less and improving the model's tolerance to outliers.

[0071] The feature matching loss is designed based on a contrastive learning framework, which learns discriminative feature representations by narrowing the distance between positive sample pairs and widening the distance between negative sample pairs. The construction of positive sample pairs considers not only spatial correspondence but also various constraints such as geometric similarity and semantic consistency. Negative sample mining employs a hard sample mining strategy, focusing on difficult negative samples that are prone to mismatch, thereby improving the model's discriminative ability.

[0072] 3-2-4) Confidence Supervision and Geometric Constraint Loss Design Building upon the basic loss in step 3-2-3), a confidence-supervised loss is further introduced, implemented through the cross-entropy loss function. This supervises the network to learn and predict the matching confidence of each feature point. Confidence labels are generated based on geometric verification and statistical analysis; matching points with good geometric consistency are marked as high-confidence, while matching points with geometric conflicts are marked as low-confidence. This confidence-supervised approach not only improves matching accuracy but also provides quality guidance for the subsequent attention mechanism.

[0073] Simultaneously, a geometric consistency loss is designed, achieved by measuring the alignment error between transformed point clouds, including multiple geometric indicators such as point-to-point distance, point-to-surface distance, and curvature similarity. The introduction of the geometric consistency loss ensures that the predicted transformation matrix is ​​not only statistically optimal but also geometrically reasonable. Robust estimation techniques are employed during the loss calculation process to reduce the impact of outliers on the overall error assessment.

[0074] Finally, a smoothness constraint loss is introduced, implemented by regularizing the feature descriptors of adjacent feature points, and a spatial regularization term is constructed based on the neighborhood structure of the point cloud. This loss function encourages spatially adjacent points to have similar feature representations, improving the spatial consistency of features. The strength of the smoothness constraint is adjusted through adaptive weights, reducing the constraint strength in boundary regions where features change drastically and strengthening the constraint in smooth regions.

[0075] 3-2-5) Multi-stage progressive training process design Based on the complete loss function constructed in steps 3-2-4), the training strategy employs a progressive learning approach, gradually improving the model's performance and robustness through multi-stage training. The entire training process is divided into three progressive stages: The first stage is the basic feature learning stage, which focuses on training the network's feature extraction capabilities. It uses relatively simple data and a large learning rate to quickly converge to a reasonable initial state. This stage primarily optimizes the feature extraction network while temporarily fixing the parameters of other modules to ensure the basic quality of the feature representation.

[0076] The second stage is the joint optimization stage. Building upon the stability of the feature network from the first stage, all network modules are trained simultaneously, using complete training data and a suitable learning rate for fine-tuning. This stage activates all loss function terms, achieving collaborative work between modules through joint optimization.

[0077] The third stage is the robustness enhancement stage. Based on the basic convergence of the model in the first two stages, more data augmentation and noise interference are introduced, and a smaller learning rate is used to achieve the final performance improvement, focusing on improving the model's generalization ability in complex environments.

[0078] 3-2-6) Adaptive Optimization Algorithms and Regularization Strategies In conjunction with the progressive training in step 3-2-5), the batch processing strategy considers the balance between GPU memory limitations and training efficiency. A dynamic batch size adjustment mechanism is employed, adaptively adjusting the batch size based on the size and complexity of the input point cloud. Smaller batches are used for dense point clouds, and larger batches are used for sparse point clouds, ensuring maximum GPU utilization. Simultaneously, gradient accumulation is introduced to achieve the effect of large-batch training by accumulating gradients from multiple small batches.

[0079] The learning rate scheduling strategy employs a combined approach, integrating the advantages of various scheduling techniques: a linear warm-up strategy is used in the early stages of training to gradually increase the learning rate and avoid instability in the initial phase; a cosine annealing strategy is used in the middle stages of training to periodically adjust the learning rate and help the model escape local optima; and an exponential decay strategy is used in the later stages of training to finely adjust the model parameters. The learning rate adjustment is also linked to the performance on the validation set, automatically reducing the learning rate when performance stagnates.

[0080] The application of regularization techniques involves the organic combination of various methods such as weight decay, label smoothing, and fusion training. Weight decay controls model complexity and prevents overfitting through L2 regularization; label smoothing improves the model's generalization ability by softening the label distribution; fusion training increases data diversity by randomly combining different samples to generate new training samples. The choice of regularization strength is determined through cross-validation to find the optimal balance between training accuracy and validation accuracy.

[0081] In step S4, a two-stage pose estimation is performed based on the multi-scale features after feature fusion.

[0082] The pose estimation algorithm converts the feature information extracted by the neural network into accurate six-degree-of-freedom transformation parameters. The algorithm design employs a hierarchical optimization strategy, implementing a complete process from rapid hypothesis generation to high-precision parameter estimation through two stages: coarse matching and fine optimization. The entire pose estimation process must not only guarantee accuracy but also meet real-time constraints; therefore, the algorithm design fully considers computational complexity and the possibility of parallelization. Specifically, it can be implemented using the following methods: Step S4-1: First stage coarse matching.

[0083] The core task of the first-stage coarse matching process is to quickly identify the most promising candidate solutions from a large number of possible pose hypotheses. The selection of feature correspondences employs a multi-layered filtering strategy. First, preliminary screening is performed based on feature similarity, using multiple metrics such as cosine distance, Euclidean distance, and Hamming distance to calculate the similarity between features. An adaptive weighting mechanism is introduced during the similarity calculation process, dynamically adjusting the weights of different distance metrics according to the dimension and type of the features. Second, a secondary screening is performed based on geometric constraints, utilizing the local geometric properties of the point cloud, such as curvature, normal vector direction, and neighborhood structure, to verify the rationality of the correspondences.

[0084] 4-1-1) Initial screening of feature similarity based on multi-metric fusion The selection of feature correspondence points employs a multi-layered filtering strategy. First, preliminary screening is performed based on feature similarity, using multiple metrics such as cosine distance, Euclidean distance, and Hamming distance to calculate the similarity between features. An adaptive weighting mechanism is introduced during the similarity calculation process, dynamically adjusting the weights of different distance metrics according to the feature's dimension and type. Specifically, for high-dimensional continuous feature vectors, cosine distance better captures directional similarity; for low-dimensional geometric features, Euclidean distance provides an intuitive spatial distance metric.

[0085] 4-1-2) Quadratic verification of correspondence driven by geometric constraints Based on the initial screening in step 4-1-1), a secondary screening is further performed based on geometric constraints. The rationality of the correspondence is verified by utilizing the local geometric properties of the point cloud, such as curvature, normal vector direction, and neighborhood structure. Geometric verification employs a multi-scale neighborhood analysis method to examine the consistency of geometric features within neighborhoods of different sizes.

[0086] Curvature consistency verification is achieved by comparing the principal curvature and Gaussian curvature of corresponding points. Correspondences with curvature differences exceeding a preset threshold are marked as unreliable. Normal vector direction verification is achieved by calculating the angle between the normal vectors of corresponding points, allowing for a certain range of angular deviations to account for noise. Neighborhood structure verification is achieved by comparing the geometric topological relationships around the corresponding point, including multiple structural features such as the relative positions, distance distributions, and angular relationships of neighboring points.

[0087] A robustness evaluation mechanism is introduced during geometric validation. Validation criteria are appropriately relaxed for geometrically unstable regions such as boundary points and points with high curvature, while validation criteria are strictly enforced for geometrically stable regions such as flat areas and interior points. The final output is a high-quality set of geometrically validated corresponding points, providing reliable input for subsequent RANSAC hypothesis generation.

[0088] 4-1-3) Quality-weighted intelligent sampling strategy Based on the set of corresponding points verified in step 4-1-2), the RANSAC algorithm employs several optimization strategies to improve the quality and efficiency of hypothesis generation. Traditional RANSAC's random sampling may generate a large number of low-quality hypotheses. The improved strategy uses weighted sampling based on feature quality, prioritizing feature points with high confidence and good geometric stability for hypothesis generation.

[0089] The weight calculation of weighted sampling takes into account multiple factors: feature confidence weight is based on the confidence score of the neural network output; geometric stability weight is based on the degree of change of local geometric features, with flat areas and areas with clear edges receiving higher weights; spatial distribution weight ensures the dispersion of sampling points in space and avoids numerical instability caused by clustered sampling.

[0090] 4-1-4) Calculation of Numerically Stable Transformation Matrix and Scoring of Multi-Criterion Assumptions Based on the intelligent sampling results in step 4-1-3), the transformation matrix is ​​calculated using a closed-form solution method based on Single-Valued Decomposition (SVD). This method can directly calculate the optimal rotation and translation parameters given corresponding point pairs. To handle the influence of noise and outliers, the input data is preprocessed before SVD calculation, including steps such as decentralization, standardization, and outlier detection. A stable numerical algorithm is used during SVD decomposition, and orthogonality constraints on the rotation matrix are ensured through orthogonalization processing.

[0091] A numerical stability check mechanism is introduced during the transformation matrix calculation to monitor numerical indicators such as the condition number and determinant value. Regularization techniques are used to correct numerical instability. The rotation matrix is ​​calculated using the extreme decomposition method to ensure that the output matrix strictly satisfies the constraints of the rotation group. The translation vector calculation considers scale invariance, and normalization is used to improve numerical accuracy.

[0092] After obtaining a stable transformation matrix, the hypothesis scoring mechanism employs a multi-criteria decision-making method, comprehensively considering multiple evaluation metrics such as the number of inliers, geometric consistency, and feature quality. The number of inliers reflects the degree of support for the hypothesis, calculated by counting the number of corresponding points within a given error threshold. Geometric consistency is evaluated by measuring the alignment quality of the transformed point cloud, including multiple geometric metrics such as point-to-point distance, point-to-surface distance, and the angle between normal vectors. Feature quality scoring is based on the confidence information output by the neural network; high-confidence features contribute more to the hypothesis score.

[0093] The final hypothesis scoring adopts a weighted combination method, and the weights of different evaluation indicators are dynamically adjusted according to the characteristics of the task and the data characteristics to ensure that the scoring results reflect both the reliability of the hypothesis and the accuracy requirements of practical applications.

[0094] Step S4-2, Second stage of refined matching.

[0095] The second-stage optimization process employs a nonlinear least squares optimization method, transforming the pose estimation problem into a parameter estimation problem. The Levenberg-Marquardt algorithm is chosen based on its stability and convergence advantages in handling nonlinear optimization problems. The algorithm implementation utilizes an adaptive damping parameter adjustment strategy, dynamically adjusting the damping coefficient according to the changing trend of the objective function to find the optimal balance between convergence speed and stability.

[0096] 4-2-1) Robust Objective Function Design for Multi-Information Source Fusion Based on the initial pose estimation from the first-stage coarse matching, the objective function design considers not only reprojection errors but also incorporates multiple information sources such as geometric constraints, physical constraints, and prior knowledge. Robust loss functions, such as Huber loss or Tukey loss, are used to mitigate the impact of outliers on the optimization process. Specifically, Huber loss employs a quadratic function form to ensure smoothness within a small error range and a linear function form to reduce the influence of outliers within a large error range; Tukey loss further suppresses the interference of extreme outliers through a dual-weighting mechanism.

[0097] Geometric constraints, including rigid body transformation constraints and boundary constraints, ensure the physical rationality of the optimization results. Rigid body transformation constraints are achieved through the orthogonality of the rotation matrix and the condition that the determinant equals 1; boundary constraints limit the variation of pose parameters within a reasonable physical range. Prior knowledge constraints, based on the mechanical structure and motion characteristics of the missile carrier, limit the search space of pose parameters, including various practical engineering constraints such as joint angle constraints, workspace constraints, and collision avoidance constraints.

[0098] The introduction of component priority weights allows the optimization process to focus on the alignment accuracy of critical components. The weight allocation is based on the importance of the component in the mounting process and the strictness of the geometric constraints: critical contact parts such as lifting rings have the highest weight because the alignment accuracy of these parts directly affects the success rate and safety of the mounting; the weight allocation also considers the geometric features of the components, with parts with rich geometric features having higher weights and parts with sparse geometric features having relatively lower weights.

[0099] 4-2-2) Stable numerical optimization implementation driven by automatic differentiation Based on the objective function constructed in step 4-2-1), the Jacobian matrix is ​​calculated using automatic differentiation, avoiding the complexity and potential errors of manual derivation. Automatic differentiation not only improves computational accuracy but also simplifies algorithm maintenance and expansion. Forward-mode automatic differentiation is suitable for cases with low parameter dimensions, while backward-mode automatic differentiation is suitable for cases with low objective function dimensions. The algorithm automatically selects the optimal differentiation mode to improve computational efficiency.

[0100] The numerical stability of the Jacobian matrix is ​​ensured through condition number monitoring and regularization techniques. When the condition number is too large, a regularization term is automatically introduced to stabilize the calculation process. The regularization parameter is selected using the L-curve method to find the optimal balance between the data fitting term and the regularization term. The matrix inversion process employs singular value decomposition (SVD) to avoid numerical instability by truncating small singular values.

[0101] Multiple convergence criteria are introduced during iterative optimization, including parameter change, objective function change, and gradient norm. This multi-criteria approach more accurately identifies the convergence state, avoiding premature stopping or invalid iterations: the parameter change criterion monitors the relative changes in parameters between adjacent iterations; the objective function change criterion monitors the degree of improvement in the optimization objective; and the gradient norm criterion monitors the convergence of the optimization direction. The convergence threshold is set based on a balance between accuracy requirements and computational resources, and the optimal threshold parameter is determined experimentally.

[0102] 4-2-3) Hierarchical Geometric Consistency Verification Mechanism Based on the converged pose parameters optimized in step 4-2-2), geometric consistency verification, as a crucial step in assessing the quality of pose estimation, employs a multi-level verification strategy to ensure the reliability of the results. The verification process consists of three progressive levels: The first level of point-level verification involves calculating the distance error for each corresponding point pair after transformation and statistically analyzing the error distribution characteristics. Point-level verification employs a nearest neighbor search algorithm, accelerating the search process through efficient data structures such as KD-trees. Nearest neighbor search not only calculates distance information but also verifies the consistency of normal vectors, ensuring the geometric rationality of the correspondence. Robust statistical methods are used in the error statistics process, employing statistics such as the median and quartiles instead of the mean to reduce the impact of outliers.

[0103] The second level of validation involves fitting the point cloud to local planes and calculating the alignment error between these planes. The RANSAC algorithm is used to improve robustness in plane fitting, and alignment quality is evaluated by comparing plane normals and inter-plane distances. The calculation of the normal vector angle difference considers the directional ambiguity of the normal vectors, avoiding misjudgments caused by 180-degree flips by selecting a smaller angle. Statistical analysis of the angle difference not only calculates the average value but also analyzes the variance and skewness of the distribution, comprehensively evaluating the consistency of the normal vectors.

[0104] The third layer, volumetric verification, calculates the overlap and shape similarity of the overall point cloud. Volumetric verification is achieved by calculating metrics such as the overlap volume ratio and shape descriptor similarity of the point clouds. Curvature similarity is calculated using the principal curvature matching method, which assesses the similarity of local geometric structures by comparing the magnitude and direction of the principal curvatures.

[0105] 4-2-4) Adaptive Quality Assessment and Threshold Optimization Strategy Based on the multi-level validation results of steps 4-2-3), the quality assessment threshold is set based on statistical analysis of a large amount of experimental data. The optimal discrimination threshold is determined through ROC curve analysis. The threshold setting process employs cross-validation, dividing the dataset into training and test sets. The optimal threshold is learned on the training set, and the generalization performance is validated on the test set. The threshold setting also considers the accuracy requirements of different application scenarios, providing multiple accuracy modes for users to choose from: high-precision mode is suitable for critical payload tasks, standard-precision mode is suitable for routine operations, and fast mode is suitable for scenarios with high real-time requirements.

[0106] The adaptive threshold adjustment mechanism dynamically adjusts the threshold parameters based on real-time performance feedback, achieving a dynamic balance between accuracy and speed. The mechanism employs a sliding window statistical method, adjusting the threshold parameters based on the statistical characteristics of the most recent matching results. When the matching success rate decreases, the threshold is automatically widened to improve recall; when the false match rate increases, the threshold is automatically tightened to improve precision.

[0107] The evaluation process outputs comprehensive evaluation information, including matching confidence scores, geometric consistency indices, and expected accuracy ranges, providing a basis for subsequent pose control decisions. High-confidence matching results are directly used for control command generation, medium-confidence results trigger a re-matching process, and low-confidence results trigger a manual intervention mechanism.

[0108] 4-2-5) Design of Pose Error Closed-Loop Control System Based on the quality assessment results of step 4-2-4), the pose error feedback control system adopts a closed-loop control strategy to achieve precise alignment by real-time monitoring of pose estimation errors and generating corresponding control commands. The control system adopts a hierarchical control architecture, with the upper layer responsible for path planning and trajectory generation, and the lower layer responsible for real-time pose adjustment and error compensation.

[0109] The decomposition of the transformation matrix employs a numerically stable algorithm to avoid the influence of singularities and numerical errors. The transformation from rotation matrix to Euler angles considers the gimbal lock problem, using quaternions as an intermediate representation to ensure the continuity and uniqueness of the transformation. Quaternion representation not only avoids singularity issues but also provides more intuitive interpolation and filtering operations. The handling of translation components considers coordinate system transformation and scale consistency issues.

[0110] The inverse kinematics solution employs a numerical optimization method, using the pseudo-inverse of the Jacobian matrix to map from Cartesian space to joint space. Constraints such as joint limiting and singularity avoidance are considered during the solution process to ensure reachability and safety: joint limiting is handled using a penalty function method, adding an extra penalty term when the joint angle approaches its limit; singularity avoidance is achieved through operational index monitoring, automatically adjusting the motion direction when approaching a singular configuration. In the case of multiple solutions, the minimum motion criterion is used to select the optimal solution, reducing mechanical wear and energy consumption.

[0111] 4-2-6) Reinforcement Learning-Driven Adaptive Parameter Optimization Based on the implementation effect of the control system in steps 4-2-5), the adaptive adjustment strategy achieves intelligent parameter optimization by establishing a mapping relationship between error and parameter adjustment. The mapping relationship is learned using reinforcement learning, gradually optimizing the adjustment strategy through interaction with the environment. The state space of the reinforcement learning agent includes multi-dimensional information such as current pose error, historical adjustment effects, and environmental change characteristics; the action space includes various parameter adjustment options; and the reward function design comprehensively considers multiple indicators such as accuracy improvement, time consumption, and stability.

[0112] Parameter tuning not only considers current performance but also predicts performance changes after tuning, avoiding parameter oscillations and overtuning. The predictive model is implemented using a neural network, trained on historical data to obtain a nonlinear mapping relationship between parameter tuning and performance changes. The tuning strategy also incorporates a balance mechanism between exploration and utilization, exploring optimal parameter configurations while ensuring current performance.

[0113] like Figure 2 The bimodal probability density distribution plot shown illustrates the fitting results of the Gaussian Mixture Model (GMM) to the feature data. The red curve represents the probability density function fitted by the GMM, exhibiting a clear bimodal characteristic. The left peak is located at approximately -2.5 eigenvalues, with a peak probability density of approximately 0.26, while the right peak is located at approximately 1.5 eigenvalues, with a peak probability density of approximately 0.22. A low probability density region of approximately 0.08 is formed near the eigenvalue of 0 between the two peaks, constituting a typical valley structure. Figure 2 The blue dots in the diagram mark two high-confidence regions, located near the two probability density peaks. These regions indicate that features have a high probability of occurrence at these locations, and the model's predictions for these regions have high confidence; they are likely to correspond to key geometric features or stable matching points in the point cloud data. The bimodal distribution indicates the existence of two distinct clustering regions in the feature space; this separation is beneficial for feature classification and matching processes.

[0114] From a technical perspective, the high probability density in the peak region provides a solid theoretical foundation for confidence calculation, while the low probability density in the valley region can effectively filter noise and outliers. The low probability region between the two peaks can also serve as a boundary reference for classification decisions. In practical applications of pose estimation, this probability distribution characteristic supports various optimization strategies: feature points located in high confidence regions can be prioritized for matching, corresponding weights can be assigned to different features based on probability density, and low-quality features can be identified and suppressed through probability distribution to improve overall matching accuracy.

[0115] This bimodal distribution provides an ideal probabilistic basis for feature confidence evaluation based on GMM, effectively supporting the implementation of dynamic attention mechanisms and feature weighted fusion strategies. This enables the model to adaptively identify and utilize high-quality feature information while suppressing noise and outlier interference, thereby improving the robustness and accuracy of the entire pose estimation system.

[0116] based on Figure 3 The analysis of the multi-scale feature fusion process shown demonstrates a meticulously designed four-level feature fusion system. Specifically, Figure 3 The geometric detail encoding layer in (a) exhibits a dense pattern of orange-yellow spots. These high-density dot activations are specifically responsible for capturing the finest-grained geometric details, including microstructural features such as structural gaps and pattern textures, demonstrating the ability to accurately model local details. Figure 3 The component structure modeling layer in (b) shows magenta-red activation regions. These activation patterns are relatively clustered and have clear structural boundaries, specifically handling medium-scale semantic components such as window frames and roof connections. Compared to the dispersed patterns in the detail layer, the activation in this layer is more concentrated and structured, reflecting effective identification of medium-level semantic components. Figure 3 The global topological feature layer in (c) exhibits a blue-green gradient distribution, covering the widest range and representing the encoding of global structural changes by high-dimensional features. This layer is mainly responsible for processing macroscopic information such as building outlines and scene topology, reflecting the system's deep understanding of overall spatial relationships and large-scale geometric structures. Figure 3 The cross-scale feature fusion layer in (d) exhibits a complex pattern of multi-color fusion, organically combining the feature representations of the first three layers. Through a cross-level attention mechanism, this layer achieves effective integration of multi-scale patterns from fine-grained geometric details to global topological structures. The entire hierarchical probabilistic fusion mechanism constructs a multi-level evidence system through three heterogeneous distance constraints. The HM layer distance captures the topological features of fine-grained local structures, the HL layer distance maintains the rigid alignment of building outlines through large-scale geometric consistency constraints, and the ML layer distance suppresses outlier interference through mid-level feature matching, forming a pyramid-shaped complementary structure that enhances geometric consistency and suppresses noise propagation.

[0117] Example 2 This embodiment discloses a probabilistic multi-scale system for estimating the relative attitude of a missile-carrying vehicle at the end of its flight path.

[0118] A probabilistic multi-scale system for estimating the terminal relative attitude of a missile-carrying vehicle includes: The point cloud data acquisition module is configured to acquire point cloud data of the end effector and the bomb rack of the bomb loading vehicle, and perform coordinate system transformation and noise suppression processing. The probabilistic multi-scale consistency modeling module is configured to: construct a probabilistic multi-scale consistency model based on noise-suppressed point cloud data; The feature fusion module is configured to perform density-adaptive multi-scale feature fusion based on the weighted features obtained from the probabilistic multi-scale consistency model. Specifically, it dynamically adjusts the matching constraint strength according to the local density distribution of the point cloud; it fuses hierarchical features at high, medium, and low scales and integrates distance constraints through a maximization operation. The pose estimation module is configured to perform two-stage pose estimation based on the multi-scale features after feature fusion. Specifically, it filters the candidate pose set by matching probability thresholds and iteratively optimizes the six-degree-of-freedom pose parameters by combining component priority. The pose estimation results corresponding to the six-degree-of-freedom pose parameters are sent to the robotic arm control unit to achieve alignment between the end effector of the bomb carrier and the bomb rack. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0119] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in the probabilistic multi-scale-based method for estimating the terminal relative attitude of a missile-carrying vehicle as described in Embodiment 1 of this disclosure.

[0120] Example 4 The purpose of this embodiment is to provide an electronic device.

[0121] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the probabilistic multi-scale-based method for estimating the terminal relative attitude of a missile-carrying vehicle as described in Embodiment 1 of this disclosure.

[0122] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0123] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0124] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for estimating the relative pose of a terminal end of a hanging vehicle based on probabilistic multiscale, characterized in that, The method comprises the following steps: Obtain the point cloud data of the end effector of the bomb hanging vehicle and the bomb rack, and perform coordinate system conversion and noise suppression processing; Based on the noise-suppressed point cloud data, a probabilistic multi-scale consistency model is constructed; Based on the weighted features obtained from the probabilistic multi-scale consistency model, density-adaptive multi-scale feature fusion is performed, specifically: dynamically adjusting the matching constraint strength according to the local density distribution of the point cloud; integrating the hierarchical features at high, medium and low scales by maximizing operation to integrate the distance constraint; Based on the multi-scale features after feature fusion, two-stage pose estimation is performed, specifically: filtering the candidate pose set through a matching probability threshold, and iteratively optimizing the six-degree-of-freedom pose parameters combined with component priority; the pose estimation result corresponding to the six-degree-of-freedom pose parameters is sent to the mechanical arm control unit to realize the alignment of the end effector of the bomb hanging vehicle and the bomb rack.

2. The probabilistic multi-scale based relative pose estimation method of a carriage end of a hanging bomb according to claim 1, characterized in that, The noise suppression processing includes: using a statistical outlier filter to remove noise points deviating from the local mean by a predetermined standard deviation threshold.

3. The probabilistic multi-scale based relative pose estimation method of a carriage end of a hanging bomb according to claim 1, characterized in that, The probabilistic multi-scale consistency model is constructed by modeling the point cloud features with a Gaussian mixture model, calculating the confidence of each feature, and dynamically assigning attention weights based on the obtained confidence.

4. The probabilistic multi-scale based relative pose estimation method of a carriage end of a hanging bomb according to claim 3, characterized in that, The feature extraction of the Gaussian mixture model is realized by an encoder, which adopts a hierarchical feature extraction strategy, specifically including an input layer, a shallow feature extraction layer, a middle feature extraction layer and a deep feature extraction layer.

5. The probabilistic multi-scale based relative pose estimation method of a carriage end of a hanging bomb according to claim 1, characterized in that, The hierarchical feature fusion at different scales is performed by a feature pyramid network, and the execution logic of the feature pyramid network includes: a bottom-up path for extracting high, medium and low hierarchical feature maps, a top-down path for transmitting high-level semantic features through upsampling, and a lateral connection path for fusing features of the same scale encoder and decoder.

6. The probabilistic multi-scale based relative pose estimation method of a cannon on a truck end according to claim 5, characterized in that, The output of the feature pyramid network is reconstructed by a decoder, and the decoder is based on the symmetry principle, and each decoding layer is connected with the corresponding encoding layer by a jump connection; wherein the fusion mode of the jump connection includes three strategies of feature splicing, element addition and gate fusion, and the optimal fusion mode is selected through adaptive weight learning.

7. The probabilistic multi-scale based relative pose estimation method of a carriage at the end of a bomb run of claim 1, wherein, The candidate pose set is filtered through a matching probability threshold as the first stage of two-stage pose estimation, which is realized by sparse convolution under a parallel RANSAC framework, that is, based on the coarse matching geometric constraint condition, a plurality of candidate relative pose hypotheses are generated; the second stage of two-stage pose estimation is to iteratively optimize the six-degree-of-freedom pose parameters combined with the component priority, which is realized by using a nonlinear least squares optimization method.

8. A probabilistic multi-scale based relative attitude estimation system for a hanging vehicle end, characterized in that, The method comprises the following steps: The point cloud data acquisition module is configured to obtain the point cloud data of the end effector of the bomb hanging vehicle and the bomb rack, and perform coordinate system conversion and noise suppression processing; The probabilistic multi-scale consistency modeling module is configured to construct a probabilistic multi-scale consistency model based on the noise-suppressed point cloud data; The feature fusion module is configured to perform density-adaptive multi-scale feature fusion based on the weighted features obtained from the probabilistic multi-scale consistency model, specifically: dynamically adjusting the matching constraint strength according to the local density distribution of the point cloud; integrating the hierarchical features at high, medium and low scales by maximizing operation to integrate the distance constraint; The pose estimation module is configured to: perform two-stage pose estimation according to the fused multi-scale features, specifically: screening a candidate pose set through a matching probability threshold, and iteratively optimizing six-degree-of-freedom pose parameters in combination with component priorities; and transmitting the pose estimation result corresponding to the six-degree-of-freedom pose parameters to a mechanical arm control unit to realize alignment of the end effector of the bomb loading vehicle and the bomb rack.

9. A computer-readable storage medium having stored thereon a program, characterized in that, The program, when executed by a processor, implements the steps in the probability multi-scale-based end effector relative pose estimation method of the bomb loading vehicle according to any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and capable of running on the processor, characterized by The processor, when executing the program, implements the steps in the probability multi-scale-based end effector relative pose estimation method of the bomb loading vehicle according to any one of claims 1-7.

Images