A method and system for 3D point cloud classification

Abstract

This invention discloses a method and system for 3D point cloud classification, belonging to the field of 3D point cloud processing and intelligent sensing technology. The method first preprocesses the input point cloud and, combining local geometric relationships, model sensitivity responses, and anomaly triggering features, quantifies the risk of each point in the point cloud to obtain a point-level risk distribution. Then, based on the point-level risk distribution, the point cloud regions are graded, and differentiated purification methods are applied to different risk regions. Finally, the purified point cloud is input into a classification model to obtain the target classification result. The proposed erosion strategy generation and threshold dynamic adjustment mechanism can adaptively determine the risk classification threshold based on sample characteristics and risk distribution status, generating corresponding purification intensity and processing parameters. Simultaneously, the proposed feedback learning and adaptive strategy update mechanism can dynamically modify the risk quantification rules and processing strategies based on the classification results and purification effects, forming a closed-loop optimization. This invention can reduce the impact of anomalies, disturbances, and triggering points on the classification process while preserving the structural information of the target entity, improving the stability, adaptability, and practical application reliability of 3D point cloud classification, and is suitable for applications such as intelligent robots, autonomous driving, and industrial inspection.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud processing and deep learning robustness technology, and particularly to a method and system for 3D point cloud classification. Background Technology

[0002] 3D point clouds, as an important form of spatial data representation, are widely used in scenarios such as autonomous driving, robot navigation, industrial inspection, and 3D reconstruction. Deep learning-based point cloud classification methods can automatically extract discriminative features from point clouds. However, during actual acquisition and transmission, point cloud data is easily affected by various complex noises and outliers, such as sensor measurement errors, outliers caused by environmental occlusion and reflection, and geometric distortions caused by sparse or uneven sampling density. These factors can lead to unstable model predictions and a significant decrease in accuracy.

[0003] Furthermore, with the deployment of point cloud models in security-sensitive scenarios, adversarial attacks and backdoor attacks targeting point clouds are gradually increasing. Research shows that attackers can construct adversarial examples by adding, deleting, or perturbing a small number of points, or implant specific triggering patterns into training / test data, causing the model to output incorrect results under certain conditions, thereby threatening system security. Because point cloud data is disordered, sparse, and structurally complex, outliers often exhibit locality and concealment. Traditional filtering strategies relying on fixed thresholds or simple geometric rules are insufficient to effectively remove risky points while preserving the details of normal point clouds.

[0004] Existing robust enhancement and protection techniques mainly include outlier removal, point cloud resampling, and adversarial training. However, some methods are highly dependent on the type and intensity of noise, resulting in limited generalization ability; others require additional training or introduce complex models, leading to increased computational overhead; and still others struggle to reliably identify high-risk points that truly affect model decisions when faced with multiple perturbations or changing attack strategies, easily exhibiting either under- or over-filtering issues, thus impacting classification performance and practical usability.

[0005] Therefore, there is an urgent need for a technical solution that can quantitatively assess and adaptively filter key risk points in point clouds under complex conditions such as various noises, adversarial disturbances, and backdoor triggering, so as to reduce the risk of model decision-making while preserving the geometric and semantic information of point clouds as much as possible, and improve the robustness and security of 3D point cloud classification. Summary of the Invention

[0006] The purpose of this invention is to achieve quantitative identification and adaptive corrosion purification adjustment of high-risk points in three-dimensional point clouds without significantly damaging the geometric and semantic information of point clouds, and to propose a method and system for three-dimensional point cloud classification.

[0007] This invention designs a point-level risk metric function to assess and dynamically classify the vulnerability of each point in a point cloud. By combining point cloud data characteristics and corrosion response feedback, it adaptively selects the corrosion type, severity, and purification intensity. It performs adaptive corrosion / removal on high-risk points, mild corrosion suppression on medium-risk areas, and maintains the original state of low-risk areas as much as possible. This enables dynamic adjustment of the corrosion strategy and purification ratio, reducing decision-making risk and improving classification robustness and security.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a method and system for three-dimensional point cloud classification, comprising the following steps:

[0009] Step 1: Calculation of comprehensive risk assessment: The control end first constructs multiple types of corrosion samples (including one or more of noise, geometric transformation, structural modification, sampling, and deformation distortion) on the input point cloud based on a preset corrosion set, and analyzes the differences in geometric response and model output of the point cloud before and after corrosion to preliminarily assess the robustness risk level of the samples; then, it calculates the comprehensive risk index by combining the point-level risk metric function to provide a basis for subsequent strategy selection.

[0010] Step 2, Point-level Risk Quantification and Vulnerability Analysis: The processing end performs point-level risk calculation on the standardized point cloud. The point-level risk calculation includes at least one or more of risk assessment based on local geometric neighborhood and risk assessment based on gradient attribution / sensitivity analysis. The point-level risk scores are then normalized to obtain the point-level vulnerability distribution and ranking results.

[0011] Step 3: Corrosion Strategy Generation and Dynamic Threshold Adjustment: The control end performs fusion analysis on point-level risk information and point cloud data feature information. The data feature information includes one or more of density distribution, noise level, geometric complexity, symmetry, and connectivity. Based on the fusion analysis results, the purification threshold and target corrosion quantity are dynamically determined, and an adaptive corrosion configuration is generated. The adaptive corrosion configuration is used to mark corrosion type combinations, severity, and purification intensity parameters. Finally, the adaptive corrosion configuration is sent to the processing end.

[0012] Step 4: Adaptive Corrosion and Hierarchical Purification Execution: After receiving the adaptive corrosion configuration, the processing end performs hierarchical processing on the point cloud according to the configuration. Specifically, the point cloud is divided into high-risk areas, medium-risk areas, and low-risk areas. Strong corrosion / removal or a higher proportion of point removal is performed on high-risk areas, mild corrosion suppression is performed on medium-risk areas, and the original point distribution is maintained or only necessary shape alignment is performed on low-risk areas. The processed point cloud is then filled in with points and its distribution is aligned to obtain a purified point cloud that meets the preset point count and statistical consistency requirements.

[0013] Step 5, Robust Reasoning and Effect Evaluation: Input the cleaned point cloud into the classification model to obtain the classification output; at the same time, calculate one or more of the following: changes in risk indicators, geometric change indicators, and diversity indicators before and after cleanup, to evaluate the effectiveness of the current corrosion and cleanup strategy.

[0014] Step 6, Feedback Learning and Adaptive Strategy Update: The control unit constructs a feedback signal based on the evaluation results obtained in Step 5, updates the strategy weights and parameter selection rules of the corrosion method, and adjusts the corrosion type combination, severity, and purification threshold of subsequent samples accordingly to achieve closed-loop optimization of the adaptive corrosion mechanism.

[0015] This invention also proposes a method and system for three-dimensional point cloud classification. The system applies the three-dimensional point cloud classification method and system described above. The system includes: a point cloud preprocessing module, a point-level risk quantification module, an adaptive corrosion selection and parameter adjustment module, a hierarchical corrosion purification execution module, a robust classification and effect evaluation module, and a feedback learning and strategy update module.

[0016] The modules deployed on the control end (server / central node) include: adaptive erosion selection and parameter adjustment module, robust classification and effect evaluation module, and feedback learning and policy update module;

[0017] The modules deployed on the processing end (terminal / edge device / inference node) include: point cloud preprocessing module, point-level risk quantification module, and hierarchical corrosion purification execution module; in an integrated deployment scenario, all of the above modules can be deployed on the same computing device.

[0018] The point cloud preprocessing module is used to: unify the format of the input point cloud, normalize the coordinates, and align the number of points to obtain a standardized point cloud input.

[0019] The point-level risk quantification module is used to: calculate point-level risk scores based on local geometric relationships and gradient attribution / sensitivity analysis, and perform normalization and sorting to obtain the set of vulnerable points and the overall risk index;

[0020] The adaptive corrosion selection and parameter adjustment module is used to: combine the feature analysis results of point cloud data with historical feedback weights to adaptively determine parameters such as corrosion type combination, severity, and purification threshold / proportion, and generate adaptive corrosion configuration;

[0021] The graded corrosion purification execution module is used to: perform graded processing on the point cloud according to the adaptive corrosion configuration, divide the point cloud into high-risk areas, medium-risk areas and low-risk areas, perform strong corrosion / removal or higher proportion of point removal on high-risk areas, perform mild corrosion suppression on medium-risk areas, keep the low-risk areas as they are, and perform point number completion and distribution alignment to output a purified point cloud.

[0022] The robust classification and effectiveness evaluation module is used to classify and reason about the cleaned point cloud, and calculate one or more of the following: changes in risk indicators, geometric change indicators, and diversity indicators before and after cleanup, in order to evaluate the corrosion cleanup effect.

[0023] The feedback learning and strategy update module is used to: generate feedback signals based on the effect evaluation results, update the erosion method selection strategy weights and parameter adjustment rules, and realize closed-loop optimization of the adaptive erosion mechanism.

[0024] Preferably, in step 1, the calculation of a comprehensive risk assessment specifically includes the following steps:

[0025] Step 1.1: Corrosion Sample Construction and Corrosion Response Calculation. First, the control unit receives the input point cloud sample. Where, p i ∈R 3

[0026] This represents the 3D coordinates of the i-th point, and N represents the number of points in the point cloud. The control terminal is based on a preset erosion stage. Construct multiple types of corrosion samples from the input point cloud, where C k The k-th corrosion method is indicated, which includes one or more of the following: noise corrosion, geometric transformation corrosion, structural modification corrosion, sampling corrosion, and deformation distortion corrosion.

[0027] For the corrosion sample corresponding to the k-th corrosion mode at severity level s, it is denoted as:

[0028]

[0029] Where s∈1,2,...,S, and S represents the total number of preset corrosion severity levels.

[0030] To evaluate the difference in point cloud response before and after corrosion, the change in geometric response is defined as:

[0031]

[0032] in, Indicates corrosion sample In the middle and the original point p i The corresponding points; when corrosion causes a change in the number of points, the corresponding relationship can be determined by nearest neighbor matching.

[0033] Furthermore, the change in the model output response is defined as:

[0034]

[0035] Among them, g θ (·) represents the classification model fθ The output vector, logit vector, or intermediate discriminant feature vector of (·).

[0036] By using the aforementioned changes in geometric response and model output response, the impact of different corrosion methods on the input point cloud can be measured from both the geometric space and the model decision space, providing basic response information for subsequent comprehensive risk assessment.

[0037] Step 1.2, Point-level Risk Quantification Calculation. Based on the obtained corrosion response information, the control unit further performs risk quantification analysis on each point in the input point cloud to identify vulnerable points that have a high impact on model decisions.

[0038] Let the point features of the l-th layer of the classification model be represented as follows:

[0039]

[0040] in, Let d represent the feature vector corresponding to i points in the l-th layer, and d represent the feature dimension.

[0041] Define the model scoring function as S f (X), then the point-level risk of the i-th point in the l-th layer is represented as:

[0042]

[0043] in, This represents the strength of the influence of the feature perturbation of the i-th point in the l-th layer on the model scoring function.

[0044] To improve the stability of point-level risk assessment, the risks from multiple point feature layers are normalized and aggregated to obtain the final point-level risk. ri

[0045]

[0046] Where L′ represents the number of point feature layers participating in the aggregation, ||·|| F This represents the Frobenius norm.

[0047] Preferably, local geometric risks can also be constructed by combining local geometric neighborhood information of the point cloud. Let... Point p i Given the set of k nearest neighbors, the local geometric risk can be expressed as:

[0048]

[0049] Furthermore, the gradient attribution risk and local geometric risk are fused to obtain a comprehensive point-level risk.

[0050]

[0051] Here, Norm(·) represents the normalization operation; β1 and β2 represent the fusion weights, and satisfy β1 + β2 = 1. This yields the point-level risk set.

[0052]

[0053] Step 1.3, Tail Risk Calculation. Since outliers, anti-perturbation points, or backdoor trigger points in point clouds are usually few in number but high in risk, this invention further introduces a tail risk metric to characterize the high-risk tail in the point-level risk distribution.

[0054] Let random variable R follow a point-level risk distribution P(·), then the value risk at confidence level α is defined as:

[0055] VaR α (R)=inf{γ:P(R<=γ)>=α}

[0056] Conditional Value at Risk (VaR) is defined as exceeding VaR. α The conditional expectation of (R), i.e.:

[0057] CVaR α (R)=E(R|R>=VaR α (R))

[0058] For discrete point-level risk sets Its discrete form can be written as:

[0059]

[0060] Where I(·) represents the indicator function, Representing the discrete probability distribution, preferably, taking... CVaR α (R) can effectively characterize the impact of high-risk points in the input point cloud on the overall decision stability, thereby obtaining the tail risk index of the sample.

[0061] Step 1.4: Adaptive Corrosion Robustness Assessment and Comprehensive Risk Calculation. After obtaining the point-level risk and tail-level risk, the control system further combines the robustness performance under different corrosion modes to generate a comprehensive risk index for the current sample, providing a basis for subsequent adaptive corrosion strategy selection.

[0062] For the k-th corrosion mode, the corrosion robustness score of the sample under the corresponding corrosion conditions is defined as CRS. k(X) represents the minimum corrosion severity level that causes a change in the classification model's prediction. A larger value indicates stronger robustness of the sample to the k-th corrosion mode. It can be expressed as:

[0063]

[0064] Based on all corrosion methods, the corrosion robustness vector of the current sample can be constructed:

[0065]

[0066] Furthermore, the average corrosion susceptibility is defined as:

[0067]

[0068] Where ε is a small constant to prevent the denominator from being zero.

[0069] Finally, by integrating tail risk, corrosion vulnerability, and changes in average output response, a comprehensive risk index is obtained:

[0070]

[0071] Where λ1, λ2, and λ3 are weighting coefficients, and satisfy:

[0072] λ1+λ2+λ3=1

[0073] Therefore, the control unit can obtain a comprehensive risk assessment result applicable to the current input point cloud. This comprehensive risk assessment result simultaneously characterizes the distribution features of high-risk points within the sample and the sample's robustness under various corrosion conditions, thus providing a basis for subsequent steps such as corrosion strategy generation, dynamic threshold adjustment, and tiered purification execution.

[0074] Preferably, in step 2, a point-level risk quantification and vulnerability analysis method includes the following steps:

[0075] Step 2.1: Point cloud standardization processing. The processing unit receives the point cloud samples to be analyzed. Next, the point cloud is first standardized to reduce the impact of translation and scale differences on the subsequent risk quantification results. Let the standardized point cloud be represented as... The standardized coordinates of the i-th point can then be expressed as:

[0076] p i ′=(p i -μ) / s

[0077] Where μ represents the center position of the input point cloud, and s represents the scale normalization factor. Preferably, μ and s are expressed as:

[0078]

[0079] s = max 1≤i≤N ||p i -μ||2

[0080] Through the above standardization process, different input samples can be subjected to subsequent point-level risk analysis under a unified coordinate scale, thereby improving the consistency of vulnerability identification results.

[0081] Step 2.2, Point-level Vulnerability Calculation. After obtaining the standardized point cloud, the processing unit performs point-level risk quantification on each point in the point cloud. Let the point features of the l-th layer of the classification model be represented as... in, Let S represent the feature vector corresponding to the i-th point in the l-th layer. Define the model scoring function as S. f (X′), then the sensitivity risk of the i-th point can be calculated based on the gradient attribution information as follows:

[0082]

[0083] Where L′ represents the number of point feature layers participating in the aggregation.

[0084] Furthermore, to reflect the local structural changes in the point cloud, the processing unit combines local geometric neighborhood information to calculate the local geometric risk of the i-th point. Let... Point p i If the set of k nearest neighbors of ′ is given, then the local geometric risk can be expressed as:

[0085]

[0086] Subsequently, the sensitivity risk and local geometric risk are fused to obtain the comprehensive vulnerability score of the i-th point:

[0087] q i =η1·Norm(a i )+η2·Norm(b i )

[0088] Where Norm(·) represents the normalization operation, η1 and η2 represent the fusion weights, and satisfy:

[0089] η1+η2=1

[0090] Through the above processing, point-level risk characteristics can be characterized from both model sensitivity and local geometry, thereby obtaining more stable vulnerability assessment results.

[0091] Step 2.3: Vulnerability Ranking and Risk Area Delineation. Based on the comprehensive vulnerability score q of each point. iThe processing unit can sort the points in the point cloud and further divide them into high-risk, medium-risk, and low-risk areas. Let the high-risk threshold and medium-risk threshold be τ, respectively. h and τ m And satisfying τ h >τ m Then the set of high-risk points can be represented as:

[0092] Ω h =p i ′|q i ≥τ h

[0093] The set of medium-risk points can be represented as:

[0094] Ω m =p i ′|τ m ≤q i <τ h

[0095] The set of low-risk points can be represented as:

[0096] Ω l =p i ′|q i <τ m

[0097] Therefore, the processing end can obtain the point-level vulnerability distribution, vulnerability ranking results, and risk area division results of the input point cloud, thus providing a basis for the generation of corrosion strategies, dynamic adjustment of thresholds, and hierarchical purification execution in subsequent steps.

[0098] Preferably, in step 3, a corrosion strategy generation and threshold dynamic adjustment method specifically includes the following steps:

[0099] Step 3.1: Point Cloud Feature Information Extraction and Statistical Representation. After completing the point-level risk quantification and vulnerability analysis in Step 2, the control unit has obtained the comprehensive vulnerability score distribution, vulnerability ranking results, and the division information of high-risk, medium-risk, and low-risk areas of the input point cloud. To further generate an adaptive corrosion configuration that matches the current sample, the control unit processes the standardized point cloud... Extract the overall feature information.

[0100] Preferably, the control terminal can calculate the average local spacing, local disturbance intensity, and risk dispersion of the point cloud, respectively. Let point p... i The set of k nearest neighbors of ′ is The average local spacing can then be expressed as:

[0101]

[0102] The intensity of a local disturbance can be expressed as:

[0103]

[0104] Furthermore, let the average of the overall vulnerability scores for each point be:

[0105]

[0106] The dispersion of point-level risk can then be expressed as:

[0107]

[0108] Based on the above features, the overall feature vector of the input point cloud can be constructed:

[0109]

[0110] The feature vector z(X) is used to characterize the density distribution characteristics, local disturbance level, and risk distribution status of the current sample, thereby providing a basis for subsequent corrosion type selection and parameter adjustment.

[0111] Step 3.2: Adaptive Erosion Strategy Generation. After obtaining the overall feature information of the point cloud, the control unit further combines the comprehensive risk index obtained in Step 1 and the risk area division results obtained in Step 2 to adaptively score various erosion methods.

[0112] Let the high-risk area be Ω h The proportion of high-risk points can then be expressed as:

[0113] φ h (X)=|Ω h | / N

[0114] Among them, |Ω h | indicates the number of points in the set of high-risk points.

[0115] For the k-th corrosion method, its corresponding strategy score is defined as:

[0116]

[0117] Among them, w k γ represents the feature weight vector corresponding to the k-th corrosion method. k and ξ k ρ(X) represents the influence coefficients of the comprehensive risk index and the proportion of high-risk points, respectively, and represents the comprehensive risk index obtained in step 1.

[0118] To obtain the relative priority of different corrosion methods, the strategy score is normalized to obtain the selection weight of the k-th corrosion method:

[0119]

[0120] Furthermore, the target severity parameter corresponding to the k-th corrosion method is calculated based on the selection weights:

[0121]

[0122] Where S represents the total number of preset corrosion severity levels, This represents the target severity parameter corresponding to the k-th corrosion method. The larger the value, the stronger the effect of the corrosion method on the current sample.

[0123] Preferably, the control terminal can select a preset threshold ε. c Select the target corrosion set from all corrosion methods:

[0124] M(X)={C k π k (X)>=ε c , k = 1, 2, ..., K

[0125] Therefore, the control unit can adaptively determine the preferred combination of corrosion types and their corresponding severity parameters based on the point-level risk distribution and overall characteristic information of the current sample.

[0126] Step 3.3: Threshold Dynamic Adjustment and Configuration Generation. After completing the corrosion mode screening, the control unit further dynamically adjusts the risk threshold and purification intensity in the subsequent purification process based on the current sample's risk level and structural characteristics.

[0127] Let the preset high-risk benchmark threshold and medium-risk benchmark threshold be respectively and The dynamic high-risk threshold can then be expressed as:

[0128]

[0129] The dynamic risk threshold can be expressed as:

[0130]

[0131] Where α1, α2, β1, and β2 are adjustment coefficients, and satisfy the following conditions:

[0132] Furthermore, to quantify the overall treatment intensity in the subsequent purification process, the purification intensity parameter is defined as:

[0133]

[0134] Wherein, ω1, ω2 and ω3 are the fusion weights.

[0135] Finally, the control unit generates the adaptive corrosion configuration for the current sample:

[0136]

[0137] Wherein, Γ(X) represents the adaptive corrosion configuration result for the current input point cloud, and the configuration result includes at least a set of target corrosion types, a dynamic risk threshold, and a purification intensity parameter. The control terminal sends the adaptive corrosion configuration to the processing terminal for use in subsequent adaptive corrosion and graded purification execution steps.

[0138] Preferably, in step 4, an adaptive corrosion and hierarchical purification method includes the following steps:

[0139] Step 4.1: Dynamic Risk Area Extraction and Processing Object Determination. After receiving the adaptive corrosion configuration Γ(X) from the control terminal in Step 3, the processing terminal combines it with the point-level vulnerability score q of the current sample. i and dynamic risk thresholds and The points in the input point cloud are re-divided into risk areas to determine the objects to be treated in subsequent graded purification.

[0140] Let high-risk areas, medium-risk areas, and low-risk areas be represented as follows: and Then we have:

[0141]

[0142] Furthermore, to ensure that the target severity parameter generated in step 3 can be actually used for the purification execution of the current sample, the processing end discretizes the target severity parameter corresponding to the k-th corrosion method to obtain the actual execution severity:

[0143]

[0144] in, This indicates the actual severity level of the k-th corrosion method on the current sample.

[0145] Through the above processing, the processing end can obtain the dynamic risk area division results for the current sample and the actual execution severity parameters, providing a basis for subsequent differentiated purification execution.

[0146] Step 4.2: Differentiated treatment for high-risk areas and medium-to-low-risk areas. For high-risk areas, the treatment unit applies the purification intensity parameter λ from the adaptive corrosion configuration. p (X) Determine the high-purification ratio. Let the treatment intensity of the high-risk area be δ. h (X) can then be expressed as:

[0147] δ h (X)=min{1,δ0+κ h λ p (X)}

[0148] Where δ0 represents the basic high-risk treatment ratio, κ h This indicates the high-risk purification adjustment coefficient.

[0149] Furthermore, the treatment unit performs strong corrosion inhibition or high-proportion spot removal on high-risk areas, resulting in high-risk purification outcomes.

[0150]

[0151] in, This indicates a high-risk area requiring intensive cleanup operations, used based on the target corrosion type set M(X) and the actual severity of the corrosion. and processing intensity δ h (X) Perform removal, suppression or strong correction on high-risk points.

[0152] For medium-risk areas, mild corrosion inhibition and localized structural modification are applied at the treatment end. Let the treatment intensity for the medium-risk area be δ. m (X), then we have:

[0153] δ m (X)=κ m λ p (X)

[0154] Among them, K m This represents the risk mitigation adjustment coefficient.

[0155] For any point p in the medium-risk area i The corrected point coordinates can be expressed as:

[0156]

[0157] This yields the point set of the medium-risk region after mild suppression and local structural correction:

[0158]

[0159] For low-risk areas, the processing end maintains the original point distribution, or only performs necessary shape alignment processing, thereby obtaining the processing results for low-risk areas:

[0160]

[0161] in, This indicates a conservative alignment operation for low-risk areas, used to preserve the main structural information of these areas.

[0162] Through the above processing, the processing end can achieve differentiated purification for different risk areas: high-risk areas are purified in a focused manner, medium-risk areas are slightly modified, and low-risk areas maintain the original structural stability.

[0163] Step 4.3: Point Completion and Distribution Alignment. After completing the differentiated processing for high-risk, medium-risk, and low-risk areas, the processing unit merges the processing results for each area to obtain an intermediate cleaned point cloud:

[0164] X c =X h UX m ∪X l

[0165] Considering that the removal of high-risk points and local corrections may lead to changes in the number of points or uneven point distribution, the processing end further performs point count completion and distribution alignment processing on the intermediate cleaned point cloud. Let the preset target number of points be N0, then the final cleaned point cloud can be represented as:

[0166]

[0167] in, This indicates point completion and distribution alignment operations, used to adjust the intermediate cleaned point cloud into an output point cloud that meets the preset point count and statistical consistency requirements.

[0168] Therefore, the processing end can obtain a clean point cloud X that meets the preset requirements for the number of points, structural consistency, and risk suppression. clean This provides an input basis for subsequent robust reasoning and effectiveness evaluation.

[0169] Preferably, in step 5, a robust reasoning and effectiveness evaluation method specifically includes the following steps:

[0170] Step 5.1: Robust inference execution of the cleaned point cloud. After completing the adaptive erosion and hierarchical cleanup execution in Step 4, the processing end obtains a cleaned point cloud X that meets the preset requirements for the number of points and distribution consistency. clean To evaluate the improvement effect of the current purification strategy on the model's discrimination results, the processing unit inputs the purified point cloud into the classification model for robust inference.

[0171] Let the classification model be f θ (·), the corresponding output vector or logit vector is represented as g. θ If (·), then the classification prediction result of the purified point cloud can be expressed as:

[0172]

[0173] Furthermore, the prediction confidence level corresponding to the purified point cloud can be expressed as:

[0174] m(X clean )=maxg θ (X clean )

[0175] in, X represents the predicted category for cleaning up the point cloud. clean This represents the maximum response intensity of the purified point cloud in the classification model.

[0176] Through the above processing, the inference results and response intensity information of the purified point cloud under the current classification model can be obtained, providing a basis for subsequent effect evaluation.

[0177] Step 5.2, Risk Suppression Effectiveness Evaluation. To evaluate the effectiveness of the current purification strategy in suppressing the distribution of high-risk points, the processing end re-performs point-level risk quantification on the purified point cloud, obtaining the purified point-level risk set. Let the purified point cloud X... clean The comprehensive point-level risk of each point is represented as follows: The purified point-level risk set can then be represented as:

[0178]

[0179] Where N0 represents the target number of points in the purified point cloud.

[0180] Based on the original point-level risk set R and the purified point-level risk set R clean The tail risk reduction rate can be defined as:

[0181]

[0182] Where, Δ risk (X) represents the decrease in tail risk after purification relative to the original input sample. The larger the value, the more obvious the effect of the current purification strategy on suppressing the impact of high-risk points.

[0183] Furthermore, to evaluate the degree to which the purification strategy improves the overall risk index, the improvement in overall risk can be defined as:

[0184] Δ ρ (X)=ρ(X)-ρ(X) clean )

[0185] Where ρ(X) represents the comprehensive risk index of the original input point cloud in step 1, ρ(X) clean This represents the comprehensive risk index for cleaning up point clouds.

[0186] Therefore, the processing end can evaluate the risk suppression effect of the current purification strategy from two aspects: tail risk suppression and overall risk improvement.

[0187] Step 5.3, Geometric Preservation and Overall Effect Evaluation. Based on the risk suppression effect evaluation, to further evaluate the ability of the cleaned point cloud to preserve the original geometric structure, the processing end calculates the cleaned point cloud X. clean The geometric change metric between the point cloud and the standardized input point cloud X′. Preferably, the Chamfer distance is used as the geometric change metric, which can be expressed as:

[0188]

[0189] Among them, D cd (X′,X clean The value indicates the degree of geometric difference between the point cloud before and after purification. The smaller the value, the better the purified point cloud retains its original structural features.

[0190] Furthermore, to measure the diversity of point distribution in the purified point cloud, a diversity index is defined as:

[0191]

[0192] in, Let Ψ(X) represent the i-th point in the purified point cloud. clean () indicates the overall distribution expansion of the purified point cloud.

[0193] Finally, by combining the tail risk reduction rate, predicted response strength, geometric change index, and diversity index, a comprehensive evaluation score for the current strategy can be constructed:

[0194] E(X) = v1Norm(Δ risk (X))+v2Norm(m(X clean ))-V3Norm(D cd (X′,X clean ))+v4Norm(Ψ(X clean ))

[0195] Where v1, v2, v3, and v4 are the fusion weights, and satisfy the following:

[0196] v1 + v2 + v3 + V4 = 1

[0197] Therefore, the system can comprehensively evaluate the effectiveness of the current adaptive corrosion and purification strategy from multiple dimensions such as risk suppression, model response, geometry preservation and distribution diversity, and provide a basis for subsequent feedback learning and strategy updates.

[0198] Preferably, in step 6, a feedback learning and adaptive policy update method specifically includes the following steps:

[0199] Step 6.1: Construction of Evaluation Result Feedback Signal. After completing the robust inference and effect evaluation in Step 5, the system has obtained a comprehensive evaluation result of the current adaptive corrosion and hierarchical cleanup strategy in terms of risk suppression, model response improvement, geometric structure preservation, and distribution consistency. In order to enable the evaluation result to further influence the corrosion strategy generation and cleanup execution process of subsequent samples, the control end first uses the comprehensive evaluation score E(X) and tail risk reduction rate Δ obtained in Step 5 as the basis for the evaluation. risk (X), Overall risk improvement Δ ρ (X) and geometric change index D cd (X′,X clean Construct a feedback signal.

[0200] Let the feedback signal of the current sample be F(X), then it can be expressed as:

[0201] F(X)=χ1E(X)+χ2Δ risk (X)+χ3Δ ρ (X)-χ4D cd (X′,X clean )

[0202] Where χ1, χ2, χ3, and χ4 are the feedback fusion weights, and satisfy the following:

[0203] χ1+χ2+χ3+χ4=1

[0204] Thus, the control end can obtain a unified feedback signal that simultaneously reflects the risk suppression effect, the model response improvement effect, and the geometry preservation capability.

[0205] Step 6.2: Corrosion Method Weight Update. After receiving the feedback signal, the control unit further updates the strategy weights of each corrosion method to improve the ability to prioritize effective corrosion methods.

[0206] Let the policy weight of the k-th erosion method at the t-th round of update be... Under the influence of the feedback signal, the updated policy weights can be expressed as:

[0207]

[0208] Where λ represents the update step size, π k (X) represents the selection weight of the k-th corrosion method in step 3.

[0209] To obtain the normalized erosion method priority, the updated selection probability can be further defined as follows:

[0210]

[0211] in, This represents the probability of choosing the k-th corrosion method in the next round of strategy generation.

[0212] Through the above update mechanism, the control end can dynamically adjust the relative priority of each corrosion method according to the current assessment results, so that the selection of corrosion methods for subsequent samples is more in line with the risk distribution characteristics and purification requirements.

[0213] Step 6.3: Severity Parameter and Threshold Configuration Update. Based on updating the corrosion mode weights, the control system further adaptively adjusts the corrosion severity parameters and risk threshold configurations to enable subsequent remediation strategies to more precisely match the risk level of the input samples.

[0214] Let the target severity parameter of the k-th corrosion method at the t-th update be... The updated target severity parameter can then be expressed as:

[0215]

[0216] Where, μ s This indicates the severity adjustment step size.

[0217] Furthermore, let the dynamic high-risk threshold and dynamic medium-risk threshold generated in step 3 be respectively... and Its update form can then be expressed as:

[0218]

[0219] Where, μ h and μ m Let represent the adjustment step sizes for the high-risk threshold and the medium-risk threshold, respectively, and ensure that the updated values ​​still satisfy the following:

[0220]

[0221] Through the above processing, the control end can make linkage corrections to the corrosion severity and threshold configuration based on the evaluation results of the current sample, thereby improving the targeting of subsequent adaptive corrosion strategies.

[0222] Step 6.4, Adaptive Strategy Closed-Loop Update and Output. After updating the erosion method weights, severity parameters, and risk threshold configurations, the control unit further generates an update strategy configuration for subsequent input samples.

[0223] Let the updated set of policy parameters be:

[0224]

[0225] Based on the updated set of policy parameters, the adaptive policy output for the next round of sample processing can be constructed:

[0226]

[0227] in, This represents a policy update mapping function used to map the updated weight, severity, and threshold parameters to an executable adaptive erosion configuration.

[0228] Therefore, the control end can feed back the evaluation results in step 5 to the strategy generation and purification execution process involved in steps 3 and 4, forming a closed-loop optimization mechanism for subsequent input samples, thereby continuously improving the system's ability to suppress high-risk points, anomalies, anti-disturbance points or backdoor trigger points and its classification robustness.

[0229] Compared with the prior art, the present invention has the following beneficial effects:

[0230] (1) By using the point-level risk quantification method, the importance and vulnerability of each point in the point cloud are finely characterized, and the high-risk points are highlighted by combining the tail risk index, which can effectively improve the ability to identify anomalies, disturbances and triggers.

[0231] (2) Through the adaptive corrosion strategy generation mechanism, the corrosion type and severity parameters can be dynamically selected based on the risk distribution, local geometric structure and overall statistical characteristics of the sample, thereby improving the adaptability between the purification strategy and the sample characteristics.

[0232] (3) By implementing a graded purification mechanism targeting high-risk, medium-risk, and low-risk areas, the impact of high-risk points can be reduced while minimizing damage to normal structural areas, thereby improving the quality and classification stability of the purified point cloud.

[0233] (4) By constructing an evaluation feedback-driven closed-loop update mechanism, the subsequent strategies can be dynamically adjusted based on the robust inference results, risk improvement amount and geometry preservation effect, thereby enhancing the system's robustness and generalization ability in complex disturbance environments. Attached Figure Description

[0234] Figure 1 This is a flowchart illustrating a method and system for three-dimensional point cloud classification provided in an embodiment of the present invention.

[0235] Figure 2 This is a schematic diagram of a method and system for three-dimensional point cloud classification provided in an embodiment of the present invention.

[0236] Figure 3 The experimental comparison diagram shows a method and system for three-dimensional point cloud classification provided in this embodiment of the invention with traditional methods.

[0237] Figure 4This invention provides a method and system architecture diagram for three-dimensional point cloud classification. Detailed Implementation

[0238] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0239] It should be noted that, to avoid obscuring important aspects of the invention with unnecessary details, the accompanying drawings only show structural modules and processing flows closely related to the invention, while omitting other details that are not directly related to the invention. Those skilled in the art should understand that the module division in the drawings can be equivalently adjusted according to specific application scenarios without departing from the core ideas of the invention.

[0240] Example 1, as Figure 1 The diagram shown is a flowchart of a method and system for three-dimensional point cloud classification proposed in this invention. To facilitate understanding of the technical solutions of the embodiments of this application, a typical execution flow of an embodiment of this application is first given:

[0241] Step 1, Comprehensive Risk Assessment: After receiving the input point cloud data, the control terminal constructs multiple types of corrosion samples based on a preset corrosion set, and jointly analyzes the differences in geometric response between the original point cloud and the point cloud after corrosion, as well as the changes in model output, to obtain the overall risk characterization of the samples under different perturbation conditions. On this basis, the comprehensive risk index of the input samples is calculated by combining the tail risk measurement results, providing a basis for subsequent point-level risk quantification and corrosion strategy generation.

[0242] Step 2, Point-level Risk Quantification and Vulnerability Analysis: Based on the comprehensive risk information issued by the control terminal, the processing end performs point-level risk calculation on each point in the standardized point cloud. The point-level risk calculation includes risk assessment based on local neighborhood geometry, sensitivity analysis based on gradient attribution, or a combination of both. Subsequently, the point-level risk scores are normalized to obtain the point-level vulnerability distribution and vulnerability ranking results.

[0243] Step 3: Corrosion Strategy Generation and Dynamic Threshold Adjustment: The control unit performs fusion analysis on point-level risk information and point cloud data features, including but not limited to point cloud density distribution, noise level, local geometric complexity, symmetry, and connectivity. Based on the fusion analysis results, the control unit dynamically determines the purification threshold and target corrosion quantity corresponding to high-risk, medium-risk, and low-risk areas, generates an adaptive corrosion configuration that matches the current input sample, and sends the adaptive corrosion configuration to the processing unit for execution.

[0244] Step 4: Adaptive Corrosion and Hierarchical Purification Execution: After receiving the adaptive corrosion configuration, the processing end performs hierarchical processing on the point cloud; among them, strong corrosion suppression, key removal, or a high proportion of point removal are performed on high-risk areas; mild disturbance suppression, local structure correction, or conservative resampling are performed on medium-risk areas; and the original distribution is maintained or only necessary shape alignment is performed on low-risk areas; after completing the regionalization processing, the point cloud is filled in and the distribution is aligned to obtain a purified point cloud that meets the preset point count requirements and statistical consistency.

[0245] Step 5, Robust Inference and Effect Evaluation: Input the cleaned point cloud into the classification model or detection model for robust inference to obtain the corresponding output results; at the same time, evaluate the changes in risk indicators, local geometric changes and point cloud distribution diversity before and after cleanup to determine the effectiveness of the current corrosion and cleanup strategy, and generate corresponding evaluation results to feed back to the control end.

[0246] Step 6, Feedback Learning and Adaptive Strategy Update: The control terminal constructs a feedback signal based on the evaluation results output in Step 5, and updates the weight allocation rules, threshold adjustment rules, and corrosion parameter selection mechanism of the corrosion strategy. This enables subsequent input samples to automatically adjust the corrosion type combination, corrosion intensity, and purification threshold based on historical processing effects, thereby achieving closed-loop adaptive optimization for robust point cloud processing tasks.

[0247] Figure 2 This invention demonstrates one possible implementation: The robust point cloud processing system based on adaptive corrosion and point-level risk quantification can be broadly divided into two parts: a control unit and a processing unit. A closed-loop processing flow is formed around the input of the point cloud data to be processed, risk assessment, strategy generation, hierarchical purification, robust inference, and feedback updates. Specifically, the control unit is mainly responsible for comprehensive risk assessment, corrosion strategy generation, and feedback update control; the processing unit is mainly responsible for point-level risk quantification, hierarchical purification execution, and robust inference assessment. The control unit constructs multiple types of corrosion samples based on the input point cloud samples and assesses their comprehensive risk. Combining the point-level risk distribution and robustness assessment results returned by the processing unit, it dynamically generates adaptive corrosion configuration parameters and updates subsequent processing strategies. The processing unit performs hierarchical purification processing on the point cloud based on the adaptive corrosion configuration parameters and outputs classification decision results, thereby achieving robust optimization for point cloud classification tasks.

[0248] This embodiment mainly considers the robust recognition problem of point cloud samples under complex perturbation or adversarial interference scenarios. The specific steps include:

[0249] Step 1, Comprehensive Risk Assessment: The control terminal receives the point cloud dataset to be processed, constructs multiple types of corrosion samples from the input point cloud samples, and obtains the comprehensive risk index and feature summary of the samples through geometric response difference analysis and output change analysis, providing a basis for the subsequent generation of corrosion strategies.

[0250] Step 2, Point-level Risk Quantification: Based on the preliminary risk analysis results provided by the control end, the processing end performs point-level risk quantification on the point cloud samples, assessing the vulnerability of each point from two aspects: local geometric structure and gradient attribution sensitivity, and obtaining the point-level risk distribution and vulnerability ranking results through normalization processing.

[0251] Step 3: Corrosion Strategy Generation: After receiving the comprehensive risk index, sample feature information, and point-level risk information, the control terminal performs fusion analysis on the risk information and data features, adaptively determines the purification threshold and corrosion intensity, generates corresponding adaptive corrosion configuration parameters, and sends these configuration parameters to the processing terminal.

[0252] Step 4: Graded purification execution: The processing end performs graded purification processing on the point cloud sample according to the adaptive corrosion configuration parameters; among them, strong corrosion suppression or point removal is performed on high-risk areas, suppression adjustment is performed on medium-risk areas, and the basic structure of low-risk areas remains unchanged. Furthermore, point number completion and distribution alignment are completed to obtain the purified point cloud sample.

[0253] Step 5, Robust Inference Evaluation: The processing end inputs the purified point cloud samples into the classification model for robust inference, and comprehensively evaluates the risk changes, geometric preservation, and purification effect of the samples before and after purification. Finally, it outputs the classification decision results and corresponding evaluation feedback information.

[0254] Step 6, Feedback Update Control: The control unit receives the robustness evaluation results and feedback signals returned in Step 5, updates the erosion strategy weights and parameter rules to form a new feedback update control strategy, and further applies the updated strategy to the comprehensive risk assessment and erosion strategy generation process of subsequent point cloud samples to achieve closed-loop optimization.

[0255] Through the above steps, this invention can organically combine point-level risk quantification results with an adaptive erosion mechanism, effectively suppressing the adverse effects of high-risk and vulnerable regions on model inference results while ensuring the main geometric structure information of the point cloud, thereby improving the robustness and stability of point cloud classification tasks.

[0256] To verify the effectiveness of the proposed robust point cloud defense method based on point-level risk quantization and adaptive erosion, this invention experimentally validates the defense framework under various complex interference scenarios, including random outliers, adversarial point noise, and backdoor-triggered attacks. Furthermore, the backdoor detection capability of this invention is evaluated under four typical backdoor attack conditions: PCBA, PointBA-I, PointBA-O, and IRBA. The experimental results are shown in Tables 1 and 2.

[0257] Table 1. Representative experimental results of the present invention in the point cloud robust classification task.

[0258] ModelNet40 PointNet Random outliers Global random outliers 88.0 ModelNet40 PointNet Random outliers Local random outliers 70.0 ModelNet40 PointNet Counteracting added noise Hausdorff Distance skill points 84.0 ModelNet40 PointNet Counteracting added noise Chamfer Distance skill points 85.3 ModelNet40 PointNet Backdoor trigger noise Poison Tag Backdoor Trigger 85.7 ModelNet40 PointNet Backdoor trigger noise Clean label backdoor trigger 78.3 ShapeNetPart PointNet Random outliers Global random outliers 95.7 ShapeNetPart PointNet Random outliers Local random outliers 85.5 ShapeNetPart PointNet Counteracting added noise Hausdorff Distance skill points 96.3 ShapeNetPart PointNet Counteracting added noise Chamfer Distance skill points 94.2 ShapeNetPart PointNet Backdoor trigger noise Poison Tag Backdoor Trigger 88.9 ShapeNetPart PointNet Backdoor trigger noise Clean label backdoor trigger 75.4

[0259] Table 2. Backdoor detection results of the present invention under four common backdoor attack scenarios.

[0260] ModelNet40 0.816 / 0.772 0.854 / 0.881 0.741 / 0.804 0.708 / 0.782 0.780 / 0.810 ShapeNetPart 0.851 / 0.821 0.883 / 0.907 0.772 / 0.836 0.754 / 0.818 0.815 / 0.846

[0261] Secondly, to further verify the effectiveness of the point cloud robust defense method described in this invention, this embodiment conducted a comparative experiment with traditional robust methods such as ROR, SOR, and PCN. The results are as follows: Figure 3 As shown. Figure 3 The paper presents a comparison of the classification accuracy of various methods on the ModelNet40 and ShapeNetPart datasets. The test scenarios include random global outliers, adversarial noise, and backdoor triggering noise. As shown in the figure, the present invention exhibits superior classification performance under all the aforementioned interference scenarios, especially maintaining high accuracy under backdoor triggering noise conditions. This demonstrates that the present invention can effectively reduce the interference of outliers and malicious triggering points on point cloud classification models, exhibiting strong robustness and practical value.

[0262] This embodiment also provides a method and system for three-dimensional point cloud classification, such as Figure 4 As shown.

[0263] The system employs the robust point cloud defense method described above, and includes: a point cloud preprocessing module, a point-level risk quantification module, an adaptive corrosion selection and parameter adjustment module, a hierarchical corrosion purification execution module, a robust classification and effect evaluation module, a feedback learning and strategy update module, a model output analysis and comprehensive risk update module, and a purified point cloud generation module.

[0264] The modules deployed on the control end include: an adaptive corrosion selection and parameter adjustment module, a robust classification and effect evaluation module, a feedback learning and strategy update module, and a model output analysis and comprehensive risk update module.

[0265] The modules deployed on the processing end include: point cloud preprocessing module, point-level risk quantification module, hierarchical corrosion and purification execution module, and purification point cloud generation module.

[0266] The point cloud preprocessing module is used to unify the format, normalize the coordinates, and align the points of the input point cloud to obtain standardized point cloud data.

[0267] The point-level risk quantification module is used to score the risk of each point in the point cloud based on local geometric relationships, gradient attribution, or sensitivity analysis results, and to complete normalization and sorting to obtain point-level risk distribution information.

[0268] The adaptive corrosion selection and parameter adjustment module is used to adaptively determine parameters such as corrosion type, corrosion intensity and purification threshold by combining point cloud data features, risk quantification results and historical feedback information.

[0269] The graded corrosion and purification execution module is used to perform graded processing on the point cloud according to the parameter adjustment results, to perform strong corrosion or removal on high-risk areas, to perform suppression processing on medium-risk areas, and to maintain the original structure or make slight adjustments to low-risk areas.

[0270] The purification point cloud generation module is used to complete the number of points and align the distribution of the point cloud after graded corrosion purification, and output a purification point cloud with a more stable structure.

[0271] The robust classification and effectiveness evaluation module is used to perform robust classification inference on the cleaned point cloud and to comprehensively evaluate the cleanup results by combining risk changes, geometric preservation and backdoor trigger suppression effects.

[0272] The model output analysis and comprehensive risk update module is used to analyze the classification output results and update the comprehensive risk information in combination with the current purification effect;

[0273] The feedback learning and strategy update module is used to iteratively update the corrosion strategy weights and parameter adjustment rules based on the effect evaluation results and comprehensive risk update results, thereby achieving closed-loop optimization of the system.

[0274] Example 2: This example considers a 3D point cloud classification application scenario suitable for intelligent sorting production lines of automotive parts. A manufacturing company deploys structured light cameras and robotic arms at the end of the production line to automatically identify and grasp parts such as housings, brackets, connectors, and valve bodies. Due to interference from metal reflections, dust adhesion, partial occlusion, and abnormal trigger points in the industrial environment, the collected point cloud data is prone to outliers and local abnormal point clusters, leading to misclassification of structurally similar parts by the classification model. Combining the point-level risk quantification and adaptive corrosion purification method proposed in this invention, the above point cloud classification task can be achieved according to the following steps:

[0275] Step 1: Point Cloud Preprocessing. The processing unit first performs centering, scale normalization, and point alignment on the input point cloud to obtain a standardized point cloud input:

[0276] P = p i i=1 N P′=p i ′i=1 N

[0277] Its standardization process satisfies:

[0278]

[0279] Step 2: Point-level risk quantification. The processing unit performs local geometric analysis, gradient sensitivity analysis, and anomaly trigger response analysis on the standardized point cloud to calculate the point-level risk score for each point.

[0280] r i =αd i +βg i +γb i

[0281] The risk scores were then normalized.

[0282]

[0283] In the process of identifying valve body components, the risk values ​​corresponding to the edges of mounting holes, flange connection areas, and local reflective areas are usually higher.

[0284] Step 3: The control unit adaptively determines the purification threshold based on the risk distribution and divides the point cloud into high-risk, medium-risk, and low-risk areas.

[0285] τ h =μ r +λ1σ r , τ m =μ r +λ2σ r

[0286] ρ h =p i ′|q i ≥τ h

[0287] ρ m =p i ′|T m ≤q i < τ h

[0288] ρ l =p i ′|q i <τ m

[0289] Step 4: Perform graded corrosion purification at the processing end. High-risk areas undergo strong corrosion or removal, medium-risk areas undergo suppression treatment, and low-risk areas retain their main structure, resulting in a purified point cloud.

[0290]

[0291] Then, point completion and distribution alignment are performed to obtain the final purified point cloud:

[0292]

[0293] Step 5: Input the purified point cloud into the classification model for robust classification and obtain the classification result:

[0294]

[0295] In the scenario of identifying bracket-type components, after purification, key structural features such as corners, connecting grooves and edge contours can be better preserved, thereby reducing the probability of misclassification with similar connectors.

[0296] Step 6: The control unit updates the strategy parameters based on the classification results and purification effect, forming a closed-loop optimization process. The update format is as follows:

[0297]

[0298] The overall loss can be expressed as:

[0299] L=ω1·Lcls+ω2·Rres+ω3·Dgeo

[0300] Through the above steps, this embodiment can effectively suppress the interference of random outliers, local abnormal point clusters and backdoor trigger points on point cloud classification results in complex industrial environments, and improve the robustness and recognition accuracy of the classification model while maintaining the main geometric structure.

[0301] Example 3: This example considers using the method of the present invention to construct a robust 3D point cloud classification system for embodied intelligent robots. In embodied intelligent application scenarios such as home services, warehouse picking, and intelligent companionship, robots typically rely on LiDAR, depth cameras, or binocular vision devices to perceive their surroundings and perform point cloud classification and recognition of targets such as tables, chairs, boxes, cups, door handles, tools, and obstacles to support tasks such as grasping, obstacle avoidance, manipulation, and interaction. However, in actual operation, the environment often presents problems such as changes in lighting, interference from transparent or reflective objects, partial occlusion, background clutter, and maliciously triggered markers, which can easily lead to abnormal points, local noise accumulation, or induced triggering areas in the input point cloud, thereby affecting the robot's environmental understanding ability and task execution stability.

[0302] Step 1: The robot first preprocesses the raw point cloud collected by the sensors, including coordinate unification, scale normalization, and point alignment, to obtain a standardized point cloud input. Through this processing, target point clouds collected at different positions, poses, and distances are mapped to a unified representation space, providing a foundation for subsequent risk analysis and robust classification.

[0303] Step 2: The robot performs point-level risk quantification analysis on the standardized point cloud, comprehensively considering local geometric anomalies, responses in classification-sensitive areas, and anomaly triggering characteristics to assess the risk level of each point in the point cloud. When home service robots identify small targets such as cups, remote controls, and medicine boxes, areas with edge defects, locally highly reflective areas, and manually added anomaly markers are usually identified as high-risk areas; when warehouse robots identify bins, packaging boxes, and pallets, anomalies near occlusion boundaries and stacking contact surfaces also typically have a high risk.

[0304] Step 3: The control unit adaptively generates a tiered purification strategy based on the uploaded risk distribution results. For high-risk areas, the system selects stronger erosion, removal, or suppression methods; for medium-risk areas, a milder purification process is used; and for low-risk areas, the original geometric structure is preserved as much as possible. This process does not rely on a fixed threshold but dynamically adjusts the purification intensity based on the current sample characteristics, historical classification results, and risk distribution, enabling the system to adapt to changes in point cloud quality in different home, warehouse, and human-computer interaction environments.

[0305] Step 4: The robot performs graded erosion purification on the point cloud according to the purification strategy issued by the control terminal. If necessary, it completes the point cloud by adding points and aligning their distribution to generate a more stable target point cloud representation. For critical areas commonly encountered by embodied robots when performing grasping tasks, such as cup handles, door handles, and drawer edges, the system suppresses interference from abnormal points while preserving as many main structural features as possible that are closely related to grasping and operation, so as to avoid the purification process adversely affecting subsequent task planning.

[0306] Step 5: Input the purified point cloud into the classification model for robust classification to obtain the target category recognition result. For example, in a home organization scenario, the robot can reliably distinguish objects such as cups, lunch boxes, medicine bottles, and controllers even with cluttered desktops and partial occlusions; in a warehouse operation scenario, the robot can maintain a relatively stable object category judgment ability even with stacked boxes, reflective plastic sealing, and partial missing parts, thereby improving the success rate of grasping and handling tasks.

[0307] Step 6: The control unit updates the risk quantification strategy and purification parameters based on the classification results, purification effect, and feedback from subsequent task execution, forming a closed-loop optimization mechanism. When the system continuously detects a specific type of abnormal interference in a certain scenario, it can automatically increase the risk sensitivity of the relevant area and enhance the purification intensity of subsequent samples; when the target structure is relatively small or the task is highly dependent on geometric details, the purification intensity can be appropriately reduced to achieve a balance between robustness and structure preservation.

[0308] The above steps can effectively improve the robustness of embodied intelligent robots in 3D point cloud classification in complex real-world environments, reduce the impact of outliers, local noise, and induced triggering areas on perception results, and thus enhance the stability, safety, and environmental adaptability of robots in grasping, obstacle avoidance, manipulation, and interaction tasks.

[0309] This invention addresses the challenges of outlier interference, adversarial noise, and the difficulty in uniformly defending against backdoor-triggered attacks in practical applications of 3D point cloud classification. It proposes a method and system for 3D point cloud classification. This method constructs a point-level risk assessment mechanism to comprehensively quantify the local geometric anomalies, gradient sensitivity, and trigger response characteristics of each point in the input point cloud. Based on this, it classifies the point cloud regions according to risk, thereby adaptively selecting erosion, suppression, and purification strategies of different intensities. This mechanism can prioritize weakening the interference of high-risk outliers on classification decisions while preserving the main geometric structure in low-risk regions as much as possible, theoretically balancing anomaly suppression and structure preservation capabilities. Experimental results show that this invention effectively improves the robustness and stability of the point cloud classification model under various complex interference scenarios, including random outliers, adversarial noise, and backdoor-triggered noise. It demonstrates superior classification performance compared to traditional robust processing methods on multiple datasets and typical models, exhibiting good adaptability and practical value.

[0310] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and such modifications or substitutions should all fall within the protection scope of the present invention.

Claims

1. A method and system for classifying three-dimensional point clouds, characterized in that: The method is used to suppress outliers, counteract disturbances and backdoor triggers that interfere with the classification results during point cloud classification. It includes a point cloud preprocessing branch, a point-level risk quantification branch, an adaptive corrosion strategy generation and threshold dynamic adjustment branch, a hierarchical corrosion purification branch, a robust classification and effect evaluation branch, and a feedback learning and adaptive strategy update branch. The point cloud preprocessing branch is used to perform format unification, coordinate normalization and point number alignment processing on the input point cloud to obtain a standardized point cloud; the point-level risk quantification branch is used to combine the local geometric relationship of the point cloud, the model's sensitive response and anomaly triggering characteristics to perform risk assessment on each point in the point cloud to obtain the point-level risk distribution. The adaptive corrosion strategy generation and threshold dynamic adjustment branch is used to adaptively determine the thresholds for dividing high-risk, medium-risk, and low-risk areas based on point-level risk distribution and current sample characteristics, and generate corresponding corrosion intensity and purification parameters. The hierarchical corrosion purification branch is used to perform strong corrosion, suppression, smoothing, or light adjustment processing on different risk areas according to the corrosion intensity and purification parameters, so as to weaken the influence of outliers and maintain the main geometric structure. The robust classification and effect evaluation branch is used to input the purified point cloud into the classification model for robust inference, and evaluate the purification effect based on the classification results, risk changes, and structure preservation. The feedback learning and adaptive strategy update branch is used to update the point-level risk quantification rules, erosion strategies and related parameters based on the evaluation results, so as to form a closed-loop optimization mechanism.

2. The method and system for three-dimensional point cloud classification according to claim 1, characterized in that: Includes the following steps: Step 1: Perform format unification, coordinate normalization, and point number alignment on the input point cloud to obtain standardized point cloud data; Step 2: Based on the local geometric relationships, model-sensitive responses, and anomaly triggering characteristics of the standardized point cloud, the risk of each point in the point cloud is quantified to obtain the point-level risk distribution; Step 3: Based on the point-level risk distribution and the feature information of the current sample, divide the point cloud region into high-risk, medium-risk, and low-risk regions, and generate the corresponding corrosion intensity, purification threshold, and processing parameters. Step 4: Based on the processing parameters, perform strong corrosion, removal, or strong suppression treatment on high-risk areas, perform suppression, smoothing, or weak corrosion treatment on medium-risk areas, and maintain the main structure or perform minor adjustments on low-risk areas. Step 5: Complete the number of points and align their distribution in the point cloud after the graded corrosion and purification process to generate a purified point cloud; Step 6: Input the purified point cloud into the point cloud classification model for classification reasoning, and evaluate the purification effect based on the classification results, risk changes, and structure preservation. Step 7: Based on the purification effect evaluation results, update the point-level risk quantification rules, corrosion strategies and related parameters to form a feedback learning and adaptive strategy update mechanism.

3. The method and system for three-dimensional point cloud classification according to claim 2, wherein step 3 includes the following steps: Step 3.1: Extract the overall feature information of the standardized point cloud. Let the standardized point cloud be... Point p i The set of k nearest neighbors of ' is The average local spacing is then expressed as: The intensity of the local disturbance is expressed as: Let the overall risk score for each point be q. i The average value is: The dispersion of point-level risk is then expressed as: Construct the overall feature vector of the input point cloud based on the above features: Step 3.2: Generate an adaptive corrosion strategy based on the overall feature vector and point-level risk distribution, assuming the high-risk region is Ω. h The percentage of high-risk points is expressed as follows: f h (X)=|Ω h | / N For the k-th corrosion method, its strategy score is defined as: The strategy score is normalized to obtain the selection weight of the k-th corrosion method: The target severity parameter corresponding to the kth corrosion method is calculated based on the selected weights: Based on the preset selection threshold ε c Select the target corrosion set from all corrosion modes: M(X)={C k :p k (X)>=e c ,k=1,2,…,K} Step 3.3: Based on the current risk level and structural characteristics of the sample, dynamically adjust the risk threshold and purification intensity during the purification process. Let the preset high-risk baseline threshold and medium-risk baseline threshold be respectively... and The dynamic high-risk threshold is then expressed as: The dynamic risk threshold is expressed as: The purification intensity parameter is defined as follows: Finally, the adaptive erosion configuration for the current sample is generated: The adaptive corrosion configuration includes at least a set of target corrosion types, a dynamic risk threshold, and a purification intensity parameter, and is used for subsequent graded corrosion purification execution.

4. The method and system for three-dimensional point cloud classification according to claim 2, wherein step 7 includes the following steps: Step 7.1: Construct a feedback signal based on the purification effect evaluation results in Step 6; let the comprehensive evaluation score of the current sample be E(X), and the tail risk reduction rate be Δ. risk (X), the overall risk improvement is Δ ρ (X), the geometric change index between the standardized point cloud and the cleaned point cloud is D. cd (X′,X clean If ), then the feedback signal is represented as: F(X)=χ1E(X)+χ2Δ risk (X)+χ3Δ ρ (X)-χ4D cd (X′,X clean ) Where χ1, χ2, χ3, and χ4 are the feedback fusion weights, and satisfy the following: χ1+χ2+χ3+χ4=1 Step 7.2: Update the strategy weights of the corrosion methods based on the feedback signal; let the strategy weight of the k-th corrosion method in the t-th round of update be... The selection weight of the k-th corrosion method in step 3 is π. k (X), then the updated policy weights are expressed as: Where λ represents the update step size; further, the updated policy weights are normalized to obtain the selection probability of the k-th erosion method in the next round of policy generation: Step 7.3: Based on updating the strategy weights of the corrosion methods, the corrosion severity parameter and risk threshold are updated in a linked manner; let the target severity parameter of the k-th corrosion method in the t-th round of update be... The updated target severity parameter is then expressed as: Where, μ s This indicates the severity adjustment step size; further, let the dynamic high-risk threshold and dynamic medium-risk threshold generated in step 3 be respectively... and The update form is then expressed as: Where, μ h and μ m Let represent the adjustment step sizes for the high-risk threshold and the medium-risk threshold, respectively, and ensure that the updated values ​​still satisfy the following: Step 7.4: Based on the updated erosion mode selection probability, target severity parameter, and risk threshold, generate the set of adaptive strategy parameters for the next round of sample processing. Based on this, the adaptive strategy output for the next round of sample processing is constructed: in, This represents the policy update mapping function, which maps the updated policy parameters to an executable adaptive erosion configuration.

5. The method and system for three-dimensional point cloud classification according to claim 2, wherein step 2 includes the following steps: Step 2.1, Point Cloud Standardization Processing. The processing unit receives the point cloud samples to be analyzed. Next, the point cloud is first standardized to reduce the impact of translation and scale differences on the subsequent risk quantification results. Let the standardized point cloud be represented as... The standardized coordinates of the i-th point can then be expressed as: p i ′=(p i -μ) / s Where μ represents the center position of the input point cloud, and s represents the scale normalization factor. Preferably, μ and s are expressed as: s=max 1≤i≤N ||p i -m||2 Through the above standardization process, different input samples can be subjected to subsequent point-level risk analysis under a unified coordinate scale, thereby improving the consistency of vulnerability identification results. Step 2.2, Point-level Vulnerability Calculation. After obtaining the standardized point cloud, the processing unit performs point-level risk quantification on each point in the point cloud. Let the point features of the l-th layer of the classification model be represented as... in, Let S represent the feature vector corresponding to the i-th point in the l-th layer. Define the model scoring function as S. f (X′), then the sensitivity risk of the i-th point can be calculated based on the gradient attribution information as follows: Where L′ represents the number of point feature layers participating in the aggregation. Furthermore, to reflect the local structural changes in the point cloud, the processing unit combines local geometric neighborhood information to calculate the local geometric risk of the i-th point. Let... Point p i If the set of k nearest neighbors of ′ is given, then the local geometric risk can be expressed as: Subsequently, the sensitivity risk and local geometric risk are fused to obtain the comprehensive vulnerability score of the i-th point: q i =η1·Norm(a i )+η2·Norm(b i ) Where Norm(·) represents the normalization operation, η1 and η2 represent the fusion weights, and satisfy: η1+η2=1 Through the above processing, point-level risk characteristics can be characterized from two dimensions: model sensitivity and local geometry, thereby obtaining more stable vulnerability assessment results. Step 2.3, Vulnerability Ranking and Risk Area Delineation. Based on the comprehensive vulnerability score q of each point. i The processing unit can sort the points in the point cloud and further divide them into high-risk, medium-risk, and low-risk areas. Let the high-risk threshold and medium-risk threshold be τ, respectively. h and τ m And satisfying τ h >τ m Then the set of high-risk points can be represented as: Ω h =p i ’|q i ≥τ h The set of medium-risk points can be represented as: Oh m =p i ′|T m ≤q i <t h The set of low-risk points can be represented as: Oh l =p i ′|q i <t m This yields the point-level risk distribution, risk ranking, and risk region division results of the input point cloud, which can be used for subsequent erosion strategy generation and dynamic threshold adjustment.

6. A method and system for classifying three-dimensional point clouds, wherein, The system applies the three-dimensional point cloud point-level risk quantification and adaptive corrosion purification method described above. The system includes: a point cloud preprocessing module, a point-level risk quantification module, an adaptive corrosion selection and parameter adjustment module, a hierarchical corrosion purification execution module, a purified point cloud generation module, a robust classification and effect evaluation module, and a feedback learning and strategy update module. The modules deployed on the control end include: an adaptive corrosion selection and parameter adjustment module, a robust classification and effect evaluation module, and a feedback learning and strategy update module; The modules deployed on the processing end include: point cloud preprocessing module, point-level risk quantification module, hierarchical corrosion and purification execution module, and purification point cloud generation module; The point cloud preprocessing module is used to perform format unification, coordinate normalization, and point number alignment on the input point cloud. The point-level risk quantification module is used to: assess the risk of each point in the point cloud based on local geometric relationships, model sensitive response, and anomaly triggering characteristics; The adaptive corrosion selection and parameter adjustment module is used to: generate a graded purification strategy based on point-level risk distribution and sample characteristics, and dynamically determine the corrosion intensity and purification threshold. The graded corrosion and purification execution module is used to: perform strong corrosion, removal, or strong inhibition treatment on high-risk areas; perform inhibition or smoothing treatment on medium-risk areas; and maintain the main structure or perform minor adjustments on low-risk areas. The purified point cloud generation module is used to: perform point number completion and distribution alignment on the purified point cloud to generate a purified point cloud; The robust classification and effectiveness evaluation module is used to: perform classification reasoning on the cleaned point cloud, and evaluate the classification results, risk changes and structure retention before and after cleanup; The feedback learning and strategy update module is used to update the point-level risk quantification rules, adaptive corrosion strategies, and related parameters based on the purification effect evaluation results.

Images