Automatic labeling method for automatic driving scene data based on incremental learning

Abstract

The invention discloses an automatic labeling method for automatic driving scene data based on incremental learning. The method comprises the following steps: S1, acquiring multi-source data and preprocessing the multi-source data; s2, performing feature mapping by adopting topological residual projection, constructing an automatic labeling model, and generating an entity labeling result; s3, analyzing an unlabeled region, extracting local residual features and updating an entity labeling result; s4, carrying out feature decomposition by adopting discrete wave domain adaptation, extracting a multi-scale feature identifier, and executing high-frequency feature reconstruction; s5, matching multi-scale feature identifiers, screening low-confidence regions, and optimizing uncertain region annotations; s6, performing incremental feature mapping in combination with the newly added data, and updating the automatic labeling model; and S7, performing continuous iterative updating by using an incremental learning method, and outputting an adaptive updating result of the entity annotation. According to the method, topological residual projection, discrete wave domain adaptation and incremental learning are combined, the labeling precision and adaptability are improved, manpower is reduced, and efficient and accurate labeling is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic data annotation, and particularly to an automatic annotation method for autonomous driving scenario data based on incremental learning. Background Art

[0002] Autonomous driving technology has achieved rapid development in recent years. The accuracy of the perception system is crucial for the safety and reliability of autonomous driving. The autonomous driving perception system mainly relies on various sensors, such as lidar, cameras, and millimeter-wave radars, to obtain environmental information, and improves the recognition ability of key targets such as roads, vehicles, pedestrians, and traffic signs through data annotation and training models. High-quality annotation data is a key factor in training an autonomous driving perception model. Therefore, the efficient and accurate annotation of autonomous driving scenario data has become one of the core issues of concern in the industry.

[0003] Existing autonomous driving data annotation methods mainly rely on manual annotation or semi-automatic annotation methods. Manual annotation requires a large amount of human input, has a low annotation efficiency, and there are also subjective errors and consistency problems. The semi-automatic annotation method usually relies on a pre-trained model for preliminary annotation, and then manual review and correction are carried out. Although this method improves the efficiency to a certain extent, it still cannot effectively adapt to the changes in new scenarios and new environments. Especially in the case of complex road structures, variable weather conditions, and dynamic changes in traffic flow, the generalization ability of existing automatic annotation models is weak, and there are prone to annotation errors or uncertain regions.

[0004] In addition, traditional automatic annotation methods usually adopt a static learning method, that is, a fixed annotation model is obtained through one-time training and then applied to different data. However, due to the highly dynamic nature of autonomous driving scenarios, the continuous changes in new scenarios, new targets, new road structures, and new traffic rules make it difficult for the fixed model to adapt to all environments. Once the model performance deteriorates, a large amount of data needs to be collected again for retraining, which not only consumes a large amount of time and computing resources, but also increases the cost of data storage and management. Therefore, the adaptability of traditional automatic annotation methods is poor and it is difficult to meet the requirements of autonomous driving systems for real-time update and efficient annotation.

[0005] To address this issue, some studies have introduced the concept of incremental learning, which enables the annotation system to continuously optimize the model as new data is added, without the need to retrain the entire dataset. The introduction of incremental learning methods helps reduce computational overhead and improve the model's adaptability to new scenarios. However, the application of existing incremental learning methods in autonomous driving scenarios still faces the following problems: First, traditional incremental learning methods mainly focus on updating model parameters and are less adaptable to changes in data features, resulting in the incremental learning of new data may affect the performance of the existing model; Second, existing methods are difficult to efficiently manage unlabeled regions and cannot reasonably optimize low-confidence data, making the annotation results still have a certain degree of uncertainty; Finally, existing incremental learning methods usually rely on deep neural networks, which are vulnerable to the problem of catastrophic forgetting, that is, the model will forget important information in the old data when learning new data, resulting in a decline in annotation accuracy.

[0006] To improve the annotation efficiency and accuracy of autonomous driving scenario data, some studies have attempted to combine multiple optimization strategies, such as projection mapping based on topological relationships, feature calibration based on wave domain transformation, etc. However, when dealing with large-scale autonomous driving data, existing topological mapping methods are easily affected by data distribution drift, resulting in a decline in the generalization ability of the annotation model. In addition, the application of wave domain transformation in autonomous driving data is still immature. Existing methods mainly rely on frequency domain filtering and feature decomposition, but lack an adaptive optimization strategy for low-confidence regions, resulting in low accuracy in the annotation boundary regions.

[0007] In addition, when dealing with the fusion of multi-source sensor data, existing autonomous driving data annotation methods often adopt direct splicing or simple alignment methods, ignoring the differences in spatio-temporal dimensions of different sensor data. This data fusion method is prone to cause inconsistencies in annotation information. Especially in the process of detecting and tracking dynamic targets, annotation errors may accumulate continuously, affecting the reliability of the final annotation results. Therefore, how to provide an automatic annotation method for autonomous driving scenario data based on incremental learning, which can adaptively update in different autonomous driving scenarios, improve the annotation efficiency and accuracy, reduce manual intervention, and optimize the annotation quality of low-confidence regions, is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose an automatic annotation method for autonomous driving scenario data based on incremental learning. The present invention optimizes the feature mapping ability of the annotation model through a topological residual projection method, decomposes and calibrates multi-scale features using a discrete wave domain adaptation algorithm, and dynamically updates the annotation model through an incremental learning mechanism to ensure that the system can adapt to changes in different driving scenarios. The present invention has the advantages of strong adaptability, high annotation efficiency, strong ability to optimize low-confidence regions, and low consumption of computing resources, improves the annotation quality of autonomous driving scenario data, reduces the need for manual intervention, and provides more accurate training data for the autonomous driving perception system.

[0009] An automatic annotation method for autonomous driving scenario data based on incremental learning according to an embodiment of the present invention includes the following steps:

[0010] S1. Obtain multi-source data in the autonomous driving scenario and preprocess the multi-source data;

[0011] S2. Perform feature mapping on the preprocessed multi-source data using topological residual projection, construct an automatic annotation model, and generate entity annotation results for static and dynamic targets within the scenario;

[0012] S3. Analyze the topological structure of the unannotated region, extract local residual features, adjust the topological residual projection mapping strategy, and update the entity annotation results;

[0013] S4. Decompose the features of the entity annotation results using a discrete wave domain adaptation algorithm, extract multi-scale feature identifiers, and perform high-frequency feature reconstruction to improve the annotation accuracy of the boundary region within the scenario;

[0014] S5. Perform feature matching on the reconstructed multi-scale feature identifiers, screen out low-confidence annotation regions, and dynamically optimize the annotation of uncertain regions using an adaptive filtering method;

[0015] S6. Combine the optimized region annotation results with the newly added multi-source data, perform incremental feature mapping on the newly added multi-source data using topological residual projection, and update the automatic annotation model;

[0016] S7. Based on the automatic annotation model, use an incremental learning method to perform continuous iterative updates and output an adaptive update result of the entity annotation.

[0017] Optionally, the multi-source data includes lidar point cloud data, camera image data, and millimeter wave radar data, and the preprocessing includes time synchronization, noise filtering, and spatial transformation.

[0018] Optionally, the S2 specifically includes:

[0019] S21. Construct a feature space matrix based on the preprocessed multi-source data, and the feature space matrix is set as where m is the number of data samples and n is the data dimension;

[0020] S22. Perform feature mapping on the feature space matrix F using topological residual projection to generate a topological relationship matrix where k represents the feature dimension after dimensionality reduction, and the mapping weight is adjusted through residual calculation;

[0021] S23. Perform classification mapping on the topological relationship matrix T, and determine the attribution of static and dynamic targets in the scene according to the target category labels C = {c1, c2,..., c l} to obtain the entity annotation result;

[0022] S24. Calculate the annotation residual matrix R = F - TW, where W is the projection mapping matrix, and through the optimization objective function:

[0023]

[0024] where argmin is the independent variable value when the function takes the minimum value, the projection mapping matrix W is adjusted to optimize the annotation accuracy, λ is the regularization coefficient, and ||·||1 is the 1-norm;

[0025] S25. Generate an automated annotation model based on the optimized topological relationship matrix T and the adjusted projection mapping matrix W, and store the entity annotation result.

[0026] Optionally, the specific steps of S3 include:

[0027] S31. Construct a topological residual perturbation matrix based on the annotation residual to characterize the local structure offset of the unannotated area:

[0028]

[0029] where, P T is the topological residual perturbation matrix, is the annotation residual matrix, is the topological relationship matrix, is the projection mapping matrix, α and β are balance coefficients, is the Laplacian operator of the annotation residual matrix R for local gradient adjustment, m is the number of data samples, n is the data dimension, and k is the feature dimension after dimensionality reduction;

[0030] S32. Extract local feature components using the topological residual perturbation matrix and construct a local residual feature matrix:

[0031]

[0032] Among them, R L is the local residual feature matrix, is the i-th row data of the topological residual perturbation matrix P T , T i is the i-th row data of the topological relationship matrix T, ||T i || is the norm of the topological vector T i ; is the transpose of the gradient of the j-th column of the topological relationship matrix T, ||T j || 2 is the squared norm of the topological vector T j , γ is the trade-off parameter;

[0033] S33. Dynamically update the projection mapping matrix W using the local residual feature matrix R L to obtain the updated projection mapping matrix:

[0034]

[0035] Among them, W ′ is the updated projection mapping matrix, η is the adjustment step size, is the transpose of the j-th column of the topological relationship matrix T, ||T j || 2 is the squared norm of the topological vector T j , δ is the regularization parameter, W i is the i-th row data of the projection mapping matrix W, ||W i || is the norm of the projection mapping vector W i ;

[0036] S34. Based on the adjusted projection mapping updated matrix W ′ , optimize the topological residual projection mapping strategy and update the entity annotation result.

[0037] Optionally, the S4 specifically includes:

[0038] S41. Perform discrete wavelet transform on the annotation residual matrix R to obtain the wave domain feature matrix B:

[0039]

[0040] Among them, B i,j is the i,j-th component of the wave domain feature matrix, R p,q is the element at the p,q position in the annotation residual matrix, ψ i,j (p,q) is the wavelet basis function at scale i and direction j, μ is the boundary enhancement coefficient, is the gradient calculation result of the annotation residual matrix, ε and ∈ are the number of rows and columns of the input matrix respectively;

[0041] S42. Perform discrete quantization mapping based on the wave domain feature matrix B, extract multi-scale feature identifiers, and obtain the adaptive feature calibration matrix D:

[0042]

[0043] where D u,v is the u,v-th element of the adaptive feature calibration matrix, B s,t is the s,t-th component of the wave domain feature matrix, min(B) and max(B) respectively represent the minimum and maximum values of matrix B, Q is the quantization level, and Φ s,t (u,v) is the feature mapping function at scale ι and direction κ;

[0044] S43. Perform high-frequency feature compensation based on the adaptive feature calibration matrix D, optimize the topological residual projection mapping strategy, and obtain the projection mapping quadratic update matrix:

[0045]

[0046] where W″ is the projection mapping quadratic update matrix, W ′ is the projection mapping update matrix, D r,t is the r,t-th element in the adaptive feature calibration matrix, φ r,t is the feature compensation basis function, λ is the regularization coefficient, is the Laplace transform of the feature calibration matrix D, Γ p,q is the high-frequency feature compensation weight, Π and Ω are respectively the feature dimension ranges, and P and Q are the optimization window sizes;

[0047] S44. The projection mapping quadratic update matrix W″, as the optimization result of the topological residual projection mapping strategy, is used to update the topological relationship matrix T and adjust the entity annotation result, further improving the annotation accuracy of the boundary region within the scene.

[0048] Optionally, the S5 specifically includes:

[0049] S51. Construct the feature matching degree matrix M, and calculate the similarity between the adaptive feature calibration matrix D and the projection mapping quadratic update matrix W″:

[0050]

[0051] where M i,j is the i,j-th element of the matching degree matrix, D p,q is the p,q-th element of the adaptive feature calibration matrix, W″ p,q is the p,q-th element of the projection mapping quadratic update matrix, ||D p,q || and ||W p,q || are respectively the norms of the vectors, Θi,j (p, q) is a feature matching weighting function, m is the number of data samples, and n is the data dimension;

[0052] S52. Set a confidence threshold τ, screen the low-confidence regions, and calculate the confidence adjustment matrix C:

[0053]

[0054] where C x,y is the x, y element of the confidence adjustment matrix, δ(M r,s < τ) is an indicator function that takes a value of 1 when M r,s is less than the confidence threshold τ and 0 otherwise, Θ r,s (x, y) is the compensation weight for the low-confidence region, and Π and Ω are the feature dimension ranges respectively;

[0055] S53. Use the confidence adjustment matrix C to perform adaptive filtering on the annotation residual matrix R to obtain the optimized confidence enhancement matrix G:

[0056]

[0057] where G a,b is the a, b position element of the optimized confidence enhancement matrix, C r,t is the r, t element of the confidence adjustment matrix, R r,t is the r, t element of the annotation residual matrix, is the Gaussian weighted filter kernel, σ is the filter scale parameter, λ is the regularization coefficient, is the Laplace transform of the confidence adjustment matrix, Ι and K are the filter window sizes;

[0058] S54. Use the optimized confidence enhancement matrix G as the input to update the automatic annotation model and improve the annotation accuracy of the low-confidence regions.

[0059] Optionally, the specific steps of S6 include:

[0060] S61. Obtain the newly added multi-source data, perform preprocessing, combine it with the confidence enhancement matrix G to construct an incremental feature sample set, normalize the newly added data through a sample balancing strategy, match the feature distribution of the existing annotation samples, and calculate the feature similarity score;

[0061] S62. Calculate the feature distribution deviation of the incremental feature sample set, identify the feature drift regions, set a threshold to screen the effective incremental samples, use the statistical distribution comparison method to calculate the KL divergence between the feature distribution of the newly added samples and the existing data, and eliminate the high-deviation samples to ensure the stability of the incremental training data;

[0062] S63. Calculate the topological residuals for the filtered incremental feature sample set, map the features of the new samples to the existing topological relation matrix T, and generate the incremental topological matrix T. ′ , and adjust the topological structure through an adaptive regularization strategy to reduce the distribution drift while maintaining the topological consistency of the new samples.

[0063] S64. Based on the incremental topological matrix T ′ , calculate the incremental learning loss function, optimize the automatic annotation model using regularization constraints, and obtain the projection mapping three - time update matrix:

[0064]

[0065] where W″′ is the projection mapping three - time update matrix, W″ is the projection mapping two - time update matrix, W is the projection mapping matrix, T′ i is the i - th sample feature vector of the incremental topological matrix, Y i is the annotation vector corresponding to this sample, λ is the regularization coefficient, m is the number of data samples, argmin is the independent variable value when the function takes the minimum value, and adjust the projection mapping matrix W to optimize the annotation accuracy.

[0066] S65. Apply the projection mapping three - time update matrix W″′ to the newly added multi - source data, perform automatic annotation prediction, and store the annotation results. At the same time, for the low - confidence prediction regions, trigger the dynamic adjustment mechanism to improve the annotation accuracy.

[0067] The beneficial effects of the present invention are as follows:

[0068] First of all, the present invention performs feature mapping on the autonomous driving scenario data through the topological residual projection method. Compared with the traditional static annotation method, this method can effectively extract the topological relationship of the data, reduce the impact of data distribution drift on the annotation accuracy, and improve the adaptability of the annotation system in different scenarios. By dynamically adjusting the topological mapping strategy, the present invention can accurately capture the local features of the unannotated regions, improve the annotation ability for complex traffic environments, and avoid the annotation failure problem caused by the fixed mapping rules in the prior art.

[0069] Secondly, the present invention uses a discrete wave domain adaptation algorithm to perform feature decomposition and high-frequency reconstruction on the labeled data, optimizing the labeling results in low-confidence regions and enhancing the robustness of the labeling system. When dealing with autonomous driving scenario data, the prior art is easily affected by factors such as lighting, weather, and occlusion, resulting in blurred labeling boundaries or mislabeling. The present invention extracts feature information at different scales through wave domain feature decomposition and adopts an adaptive calibration strategy to compensate and optimize the high-frequency feature regions, significantly improving the accuracy of the labeling system in the boundary regions. In addition, combined with the adaptive filtering method, the present invention can automatically screen low-confidence labeling regions and perform dynamic optimization, thereby reducing labeling errors and improving the overall data labeling quality.

[0070] Finally, the present invention continuously optimizes the labeling process through an incremental learning mechanism, avoiding the high computational cost problem of the traditional method that requires frequent retraining of the entire dataset. The introduction of incremental learning enables the labeling system to be updated in real time according to newly collected data without affecting the labeling accuracy of the existing data, solving the problem of the traditional method forgetting old knowledge when learning new scenarios. The incremental feature mapping method adopted by the present invention enables new data to be quickly adapted based on the existing basis, reducing the occupation of computing resources and improving the real-time performance of automatic labeling at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0072] Figure 1 is a flowchart of an automatic labeling method for autonomous driving scenario data based on incremental learning proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0074] Refer to Figure 1 , an automatic labeling method for autonomous driving scenario data based on incremental learning, including the following steps:

[0075] S1. Obtain multi-source data in the autonomous driving scenario and preprocess the multi-source data;

[0076] S2. Use topological residual projection to perform feature mapping on the preprocessed multi-source data, construct an automatic labeling model, and generate entity labeling results for static and dynamic targets in the scenario;

[0077] S3. Perform topological structure analysis on the unlabeled regions, extract local residual features, adjust the topological residual projection mapping strategy, and update the entity annotation results;

[0078] S4. Use the discrete wave domain adaptation algorithm to perform feature decomposition on the entity annotation results, extract multi-scale feature identifiers, and perform high-frequency feature reconstruction to improve the annotation accuracy of the boundary regions within the scene;

[0079] S5. Perform feature matching on the reconstructed multi-scale feature identifiers, filter out the low-confidence annotation regions, and use the adaptive filtering method to dynamically optimize the annotation of the uncertain regions;

[0080] S6. Combine the optimized region annotation results with the newly added multi-source data, perform incremental feature mapping on the newly added multi-source data using topological residual projection, and update the automated annotation model;

[0081] S7. Based on the automated annotation model, use the incremental learning method to perform continuous iterative updates and output the adaptive update results of the entity annotation.

[0082] By combining the topological residual projection method and the discrete wave domain adaptation algorithm, the present invention realizes the efficient automatic annotation of autonomous driving scene data, and uses the incremental learning mechanism to optimize the annotation process, enabling the system to adapt to the changing driving environment. Compared with the traditional annotation method, the present invention significantly improves the annotation efficiency, reduces the need for manual intervention, and performs excellently in optimizing low-confidence regions, making the annotation results more accurate and stable.

[0083] In this embodiment, the multi-source data includes lidar point cloud data, camera image data, and millimeter wave radar data, and the preprocessing includes time synchronization, noise filtering, and spatial transformation.

[0084] By acquiring lidar point cloud data, camera image data, and millimeter wave radar data, and performing time synchronization, noise filtering, and spatial transformation, the present invention ensures the spatio-temporal consistency of the data and improves the adaptability of the annotation system in complex environments. Time synchronization eliminates the data delay between multiple sensors, making data fusion more accurate. Noise filtering reduces false detections and false targets, improving data reliability. Spatial transformation ensures the unity of the data coordinates of different sensors, providing a stable data foundation for subsequent feature mapping and annotation.

[0085] In this embodiment, the specific steps of S2 are as follows:

[0086] S21. Construct a feature space matrix based on the preprocessed multi-source data, and the feature space matrix is set as where m is the number of data samples and n is the data dimension;

[0087] S22. Use topological residual projection to perform feature mapping on the feature space matrix F to generate a topological relationship matrix where k represents the feature dimension after dimensionality reduction, and the mapping weights are adjusted through residual calculation;

[0088] S23. Perform classification mapping on the topological relationship matrix T, and make attribution judgments on static and dynamic targets in the scene according to the target category labels C = {c1, c2,..., c l} to obtain the entity annotation result;

[0089] S24. Calculate the annotation residual matrix R = F - TW, where W is the projection mapping matrix, through the optimization objective function:

[0090]

[0091] where argmin is the independent variable value when the function takes the minimum value, adjust the projection mapping matrix W to optimize the annotation accuracy, λ is the regularization coefficient, and ||·||1 is the 1-norm;

[0092] S25. Based on the optimized topological relationship matrix T and the adjusted projection mapping matrix W, generate an automated annotation model and store the entity annotation result.

[0093] The present invention uses the topological residual projection method to perform feature mapping on the preprocessed multi-source data, enabling automated annotation to extract target features based on topological information without being affected by data distribution drift. This method effectively reduces the feature offset of static and dynamic targets in different sensor data, makes entity annotation more stable, and improves annotation accuracy. Compared with the traditional annotation method based on fixed feature matching, the present invention can better adapt to the changes in the road environment and enhance the annotation ability for complex scenes.

[0094] In this embodiment, the specific steps of S3 are as follows:

[0095] S31. Based on the annotation residuals, construct a topological residual perturbation matrix to characterize the local structure offset of the unannotated area:

[0096]

[0097] where P T is the topological residual perturbation matrix, is the annotation residual matrix, is the topological relationship matrix, is the projection mapping matrix, α and β are balance coefficients, is the Laplacian operator of the annotation residual matrix R for local gradient adjustment, m is the number of data samples, n is the data dimension, and k is the feature dimension after dimensionality reduction;

[0098] S32. Extract local feature components using the topological residual perturbation matrix and construct a local residual feature matrix:

[0099]

[0100] Among them, R L is the local residual feature matrix, is the i-th row data of the topological residual perturbation matrix P T and T i is the i-th row data of the topological relationship matrix T. ||T i || is the norm of the topological vector T i . is the transpose of the gradient of the j-th column of the topological relationship matrix T, and ||T j || 2 is the squared norm of the topological vector T j . γ is a trade-off parameter;

[0101] S33. Dynamically update the projection mapping matrix W using the local residual feature matrix R L to obtain an updated projection mapping matrix:

[0102]

[0103] Among them, W ′ is the updated projection mapping matrix, η is the adjustment step size, is the transpose of the j-th column of the topological relationship matrix T, and ||T j || 2 is the squared norm of the topological vector T j . δ is a regularization parameter, and W i is the i-th row data of the projection mapping matrix W. ||W i || is the norm of the projection mapping vector W i .

[0104] S34. Based on the adjusted updated projection mapping matrix W ′ , optimize the topological residual projection mapping strategy and update the entity annotation results.

[0105] In the present invention, topological structure analysis and local residual feature extraction are adopted in the unlabeled area to effectively identify the low-confidence areas in the data, and by dynamically adjusting the topological residual projection strategy, the annotation system can be adaptively optimized for different environmental changes. This method significantly reduces the mislabeling rate in complex scenarios. Especially in the case of drastic lighting changes and complex scene structures, it can automatically adjust the annotation strategy to improve the overall annotation consistency.

[0106] In this embodiment, the specific content of S4 includes:

[0107] S41. Perform a discrete wavelet transform on the labeled residual matrix R to obtain a wave domain feature matrix B:

[0108]

[0109] where B i,j is the i,j-th component of the wave domain feature matrix, R p,q is the element at the p,q position in the labeled residual matrix, ψ i,j (p,q) is the wavelet basis function at scale i and direction j, μ is the boundary enhancement coefficient, is the gradient calculation result of the labeled residual matrix, ε and ∈ are the number of rows and columns of the input matrix respectively;

[0110] S42. Based on the wave domain feature matrix B, perform discrete quantization mapping, extract multi-scale feature identifiers, and obtain an adaptive feature calibration matrix D:

[0111]

[0112] where D u,v is the u,v-th element of the adaptive feature calibration matrix, B s,t is the s,t-th component of the wave domain feature matrix, min(B) and max(B) represent the minimum and maximum values of matrix B respectively, Q is the quantization level, Φ s,t (u,v) is the feature mapping function at scale ι and direction κ;

[0113] S43. Based on the adaptive feature calibration matrix D, perform high-frequency feature compensation, optimize the topological residual projection mapping strategy, and obtain a projection mapping secondary update matrix:

[0114]

[0115] where W″ is the projection mapping secondary update matrix, W ′ is the projection mapping update matrix, D r,t is the r,t-th element in the adaptive feature calibration matrix, φ r,t is the feature compensation basis function, λ is the regularization coefficient, is the Laplace transform of the feature calibration matrix D, Γ p,q is the high-frequency feature compensation weight, Π and Ω are the feature dimension ranges respectively, P and Q are the optimization window sizes;

[0116] S44. The projection mapping secondary update matrix W″, as the optimization result of the topological residual projection mapping strategy, is used to update the topological relationship matrix T and adjust the entity annotation result, further improving the annotation accuracy of the boundary area within the scene.

[0117] The present invention uses a discrete wave domain adaptation algorithm to perform feature decomposition on the entity annotation results, extract multi-scale feature identifiers, and perform high-frequency feature reconstruction to optimize the annotation accuracy of the boundary region. This method addresses the problems of fuzzy annotation boundaries and information loss in low-confidence regions, uses multi-scale wave domain transformation to enhance target features, and combines high-frequency feature compensation to reduce the uncertainty of the target boundary. Compared with traditional single-scale feature matching, the present invention can more finely optimize the target contour and improve the accuracy of key targets in the scene, especially performing better in complex environments such as at night and in rainy days.

[0118] In this embodiment, step S5 specifically includes:

[0119] S51. Construct a feature matching degree matrix M, and calculate the similarity between the adaptive feature calibration matrix D and the projection mapping quadratic update matrix W″:

[0120]

[0121] where M i,j is the i,j-th element of the matching degree matrix, D p,q is the p,q-th element of the adaptive feature calibration matrix, W″ p,q is the p,q-th element of the projection mapping quadratic update matrix, ||D p,q || and ||W″ p,q || are the norms of the vectors respectively, Θ i,j (p,q) is the feature matching weighting function, m is the number of data samples, and n is the data dimension;

[0122] S52. Set a confidence threshold τ, screen the low-confidence regions, and calculate the confidence adjustment matrix C:

[0123]

[0124] where C x,y is the x,y-th element of the confidence adjustment matrix, δ(M r,s <τ) is the indicator function, which takes the value of 1 when M r,s is less than the confidence threshold τ, and 0 otherwise, Θ r,s (x,y) is the low-confidence region compensation weight, and Π and Ω are the feature dimension ranges respectively;

[0125] S53. Use the confidence adjustment matrix C to perform adaptive filtering on the annotation residual matrix R to obtain the optimized confidence enhancement matrix G:

[0126]

[0127] where G a,b is the a,b-position element of the optimized confidence enhancement matrix, Cr,t is the r, t-th element of the confidence adjustment matrix, R r,t is the r, t-th element of the annotation residual matrix, is the Gaussian weighted filter kernel, σ is the filter scale parameter, and λ is the regularization coefficient, is the Laplace transform of the confidence adjustment matrix, Ι and K are the filter window sizes;

[0128] S54. Use the optimized confidence enhancement matrix G as the input, update the automatic annotation model, and improve the annotation accuracy of low-confidence regions.

[0129] The present invention performs feature matching on the reconstructed multi-scale feature identifiers, screens low-confidence annotation regions, and combines an adaptive filtering method to dynamically optimize the annotation of uncertain regions. The present invention can automatically detect regions with low annotation confidence and perform dynamic optimization through feature matching, making the annotation results more robust. By using the adaptive filtering method, the optimized annotation system can continuously adjust the annotation strategy in a dynamic environment, improve the annotation accuracy of low-confidence regions, reduce the dependence on manual correction compared with static annotation methods, and improve the data processing efficiency.

[0130] In this embodiment, the specific steps of S6 are as follows:

[0131] S61. Obtain new multi-source data, perform preprocessing, combine the confidence enhancement matrix G, construct an incremental feature sample set, normalize the new data through a sample balancing strategy, match the feature distribution of existing annotation samples, and calculate the feature similarity score;

[0132] S62. Calculate the feature distribution deviation of the incremental feature sample set, identify the feature drift region, set a threshold to screen effective incremental samples, use the statistical distribution comparison method to calculate the KL divergence between the feature distribution of the new samples and the existing data, and eliminate high-deviation samples to ensure the stability of the incremental training data;

[0133] S63. Calculate the topological residual of the filtered incremental feature sample set, map the features of the new samples to the existing topological relationship matrix T, and generate an incremental topological matrix T ′ , and adjust the topological structure through an adaptive regularization strategy to reduce the distribution drift while maintaining the topological consistency of the new samples;

[0134] S64. Based on the incremental topological matrix T ′ calculate the incremental learning loss function, optimize the automatic annotation model using regularization constraints, and obtain the projection mapping three-time update matrix:

[0135]

[0136] Among them, W″′ is the projection mapping triple update matrix, W″ is the projection mapping double update matrix, W is the projection mapping matrix, T′ i is the i-th sample feature vector of the incremental topology matrix, Y i is the annotation vector corresponding to the sample, λ is the regularization coefficient, m is the number of data samples, argmin is the independent variable value when the function takes the minimum value, and the projection mapping matrix W is adjusted to optimize the annotation accuracy;

[0137] S65. Apply the projection mapping triple update matrix W″′ to the newly added multi-source data, perform automatic annotation prediction, and store the annotation results. At the same time, for the low-confidence prediction area, trigger the dynamic adjustment mechanism to improve the annotation accuracy.

[0138] The present invention adopts the topological residual projection method to perform incremental feature mapping on the newly added multi-source data and update the automatic annotation, enabling the annotation system to have the incremental learning ability and adapt to new scenarios without re-training. This method can dynamically adjust the topological mapping structure during the annotation process, enabling the system to quickly adapt and automatically annotate when facing new road environments and traffic targets without a large amount of manual intervention. Compared with the traditional static annotation method, the present invention significantly reduces the computational overhead, improves the real-time performance and adaptability of the system, and makes it more suitable for the requirements of autonomous driving data annotation.

[0139] Embodiment 1:

[0140] To verify the feasibility of the present invention in implementation, the present invention is applied to a certain intelligent driving test scenario, which includes urban roads, highways, and complex intersections, covering different weather conditions such as sunny, cloudy, and rainy days. The test area is located in a certain autonomous driving test base, covering an area of 50 square kilometers, equipped with a variety of sensors and monitoring devices to collect high-precision autonomous driving scenario data. The test environment includes complex traffic elements such as different types of vehicles, pedestrians, traffic signals, and road signs to ensure the diversity and complexity of the test data.

[0141] In this scenario, the present invention automatically annotates the data collected by autonomous driving through a topological residual projection method and a discrete wave domain adaptation algorithm, and adopts an incremental learning mechanism to dynamically update the annotation results. During the testing process, the intelligent driving vehicle is equipped with lidar, cameras, and millimeter-wave radars, travels on the road, and collects environmental data in real time. After the data is transmitted to the annotation system, time synchronization, noise filtering, and spatial transformation are first performed to ensure data consistency. Then, the topological residual projection method is used to map the features of the data, extract the topological relationships of key targets such as roads, vehicles, and pedestrians, and generate preliminary annotation results. For low-confidence regions, the discrete wave domain adaptation algorithm is used for multi-scale feature analysis to optimize the annotation accuracy of the boundary regions. Finally, the annotated data enters the incremental learning process, enabling the system to continuously optimize according to newly collected data and improve its adaptability to new scenarios and new targets.

[0142] During the testing process, in the same autonomous driving environment, the traditional manual annotation, semi-automatic annotation, and the automatic annotation method of the present invention are compared to evaluate the annotation efficiency, accuracy, and the optimization of low-confidence regions. A total of about 500,000 frames of data are collected in the experiment, including 200,000 frames of urban road scenarios, 150,000 frames of highway scenarios, 100,000 frames of complex intersection scenarios, and 50,000 frames of special weather scenarios. The traditional manual annotation method requires an average of 5 minutes per frame for annotation, with relatively low annotation efficiency, and in complex traffic environments, the subjective errors of manual annotation are relatively large. The semi-automatic annotation method combines a pre-trained model for preliminary annotation, but in new scenarios, the annotation accuracy drops significantly and requires a large amount of manual correction.

[0143] Table 1 Comparison experimental data table of automatic annotation methods

[0144]

[0145] The automatic annotation method of the present invention can complete the annotation in only 0.8 seconds per frame on average under the same amount of data, with the annotation efficiency increased by about 375 times, and at the same time, the annotation confidence reaches 97.5%, which is 4.2% higher than that of the semi-automatic annotation method. In addition, in terms of the optimization effect of low-confidence regions, the present invention performs better than the traditional method. During the testing process, in scenarios with drastic changes in light, the vehicle boundary blur degree is relatively high for the traditional annotation method, and the mislabeling rate reaches 8.7%; in scenarios with dense pedestrians at complex intersections, the mislabeling rate is as high as 12.3%. After adopting the automatic annotation method of the present invention, the mislabeling rates are reduced to 2.4% and 3.1% respectively. Especially in rainy and night environments, the discrete wave domain adaptation algorithm of the present invention can effectively identify low-confidence regions and perform adaptive optimization, reducing the annotation error by more than 60%.

[0146] Through comparative tests, the present invention performs excellently in terms of annotation efficiency, annotation accuracy, and optimization of low-confidence regions, can effectively reduce the need for manual intervention, improve the quality of data annotation, and provide higher-quality training data for the autonomous driving perception system.

[0147] The test results show that the present invention has significant advantages in large-scale autonomous driving data annotation, can not only improve the annotation efficiency, but also reduce the annotation error, and still maintain high-precision annotation in complex environments, providing reliable data support for the optimization of the autonomous driving perception system.

[0148] As mentioned above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An automatic annotation method for autonomous driving scenario data based on incremental learning, characterized in that, It includes the following steps: S1. Obtain multi-source data in the autonomous driving scenario and preprocess the multi-source data; S2. Use topological residual projection to perform feature mapping on the preprocessed multi-source data, construct an automated annotation model, and generate entity annotation results for static and dynamic targets within the scenario; S3. Conduct topological structure analysis on the unannotated area, extract local residual features, adjust the topological residual projection mapping strategy, and update the entity annotation results; S4. Use the discrete wave domain adaptation algorithm to perform feature decomposition on the entity annotation results, extract multi-scale feature identifiers, and perform high-frequency feature reconstruction to improve the annotation accuracy of the boundary area within the scenario; S5. Perform feature matching on the reconstructed multi-scale feature identifiers, screen out low-confidence annotation areas, and use an adaptive filtering method to dynamically optimize the annotation of uncertain areas; S6. Combine the optimized regional annotation results with the newly added multi-source data, use topological residual projection to perform incremental feature mapping on the newly added multi-source data, and update the automated annotation model; S7. Based on the automated annotation model, use the incremental learning method to perform continuous iterative updates and output the adaptive update results of entity annotation.

2. The automatic annotation method for autonomous driving scenario data based on incremental learning according to claim 1, wherein The multi-source data includes lidar point cloud data, camera image data, and millimeter-wave radar data, and the preprocessing includes time synchronization, noise filtering, and spatial transformation.

3. The automatic annotation method for autonomous driving scenario data based on incremental learning according to claim 1, wherein, The specific content of S2 includes: S21. Construct a feature space matrix based on the preprocessed multi-source data, and the feature space matrix is set as where m is the number of data samples and n is the data dimension; S22. Perform feature mapping on the feature space matrix F using topological residual projection to generate a topological relationship matrix where k represents the feature dimension after dimensionality reduction, and the mapping weights are adjusted through residual calculation; S23. Perform a classification mapping on the topological relationship matrix T, and determine the attribution of static and dynamic targets in the scene according to the target category labels C = {c1, c2,..., c l}, to obtain the entity annotation result; S24. Calculate the annotation residual matrix R = F - TW, where W is the projection mapping matrix, and optimize the projection mapping matrix W through the optimization objective function: where argmin is the independent variable value when the function takes the minimum value, adjust the projection mapping matrix W to optimize the annotation accuracy, λ is the regularization coefficient, and ||·||1 is the 1-norm; S25. Based on the optimized topological relationship matrix T and the adjusted projection mapping matrix W, generate an automated annotation model and store the entity annotation results.

4. The automatic annotation method for autonomous driving scenario data based on incremental learning according to claim 1, wherein The specific content of S3 includes: S31. Based on the annotation residuals, construct a topological residual perturbation matrix to characterize the local structure offset of the unannotated area: Among them, P T is the topological residual perturbation matrix, is the annotation residual matrix, is the topological relationship matrix, is the projection mapping matrix, α and β are balance coefficients, is the Laplacian operator of the annotation residual matrix R for local gradient adjustment, m is the number of data samples, n is the data dimension, and k is the feature dimension after dimensionality reduction; S32. Use the topological residual perturbation matrix to extract local feature components and construct a local residual feature matrix: Among them, R L is the local residual feature matrix, is the i-th row data of the topological residual perturbation matrix P T and T i is the i-th row data of the topological relationship matrix T. ||T i || is the norm of the topological vector T i , is the transpose of the gradient of the j-th column of the topological relationship matrix T. ||T j || 2 is the squared norm of the topological vector T j , and γ is the trade-off parameter; S33. Utilize the local residual feature matrix R L Dynamically update the projection mapping matrix W to obtain the updated projection mapping matrix: Among them, W ′ is the projection mapping update matrix, η is the adjustment step size, is the transpose of the j-th column of the topological relationship matrix T, ||T j || 2 is the squared norm of the topological vector T j δ is the regularization parameter, W i is the i-th row data of the projection mapping matrix W, ||W i || is the norm of the projection mapping vector W i ; S34. Update matrix W of the projection mapping based on the adjusted projection mapping, optimize the topological residual projection mapping strategy, and update the entity annotation result. ′ , optimize the topological residual projection mapping strategy, and update the entity annotation result.

5. A method for automatically annotating autonomous driving scenario data based on incremental learning according to claim 1, characterized in that, The specific content of S4 includes: S41. Perform discrete wavelet transform on the annotation residual matrix R to obtain the wave domain feature matrix B: Among them, B i,j is the i,j-th component of the wave domain feature matrix, R p,q is the element at the p,q position in the annotation residual matrix, ψ i,j (p,q) is the wavelet basis function at scale i and direction j, μ is the boundary enhancement coefficient, is the gradient calculation result of the annotation residual matrix, ε and ∈ are the number of rows and columns of the input matrix respectively; S42. Based on the wave domain feature matrix B, perform discrete quantization mapping, extract multi-scale feature identifiers, and obtain the adaptive feature calibration matrix D: where D u,v is the u,v-th element of the adaptive feature calibration matrix, B s,t is the s,t-th component of the wave domain feature matrix, min(B) and max(B) respectively represent the minimum and maximum values of matrix B, Q is the quantization level, and Φ s,t (u, v) is the feature mapping function at scale ι and direction κ; S43. Based on the adaptive feature calibration matrix D, perform high-frequency feature compensation, optimize the topological residual projection mapping strategy, and obtain the projection mapping secondary update matrix: Among them, W ″ is the projection mapping quadratic update matrix, and W ′ is the projection mapping update matrix. D r,t is the r, t-th element in the adaptive feature calibration matrix, φ r,t is the feature compensation basis function, λ is the regularization coefficient, is the Laplace transform of the feature calibration matrix D, Γ p,q is the high-frequency feature compensation weight, Π and Ω are the feature dimension ranges respectively, and P and Q are the optimization window sizes; S44. Projection mapping secondary update matrix W ″ As the optimization result of the topological residual projection mapping strategy, it is used to update the topological relation matrix T and adjust the entity annotation result, further improving the annotation accuracy of the boundary region within the scene.

6. The automatic annotation method for autonomous driving scenario data based on incremental learning according to claim 1, wherein The specific content of S5 includes: S51. Construct a feature matching degree matrix M, and calculate the similarity between the adaptive feature calibration matrix D and the projection mapping quadratic update matrix W ″ between: Among them, M i,j is the i,j-th element of the matching degree matrix, D p,q is the p,q-th element of the adaptive feature calibration matrix, W″ p,q is the p,q-th element of the projection mapping quadratic update matrix, ||D p,q || and ||W″ p,q || are the norms of the vectors respectively, Θ i,j (p,q) is the feature matching weighting function, m is the number of data samples, and n is the data dimension; S52. Set the confidence threshold τ, screen out low-confidence areas, and calculate the confidence adjustment matrix C: where C x,y is the x,y-th element of the confidence adjustment matrix, δ(M r,s < τ) is the indicator function, which takes the value of 1 when M r,s is less than the confidence threshold τ, and 0 otherwise, Θ r,s (x,y) is the compensation weight for the low-confidence region, and Π and Ω are the ranges of the feature dimensions respectively; S53. Use the confidence adjustment matrix C to perform adaptive filtering on the annotation residual matrix R to obtain the optimized confidence enhancement matrix G: Among them, G a,b is the element at the a, b position of the optimized confidence enhancement matrix, C r,t is the r, t-th element of the confidence adjustment matrix, R r,t is the r, t-th element of the annotation residual matrix, is the Gaussian weighted filter kernel, σ is the filter scale parameter, and λ is the regularization coefficient, is the Laplace transform of the confidence adjustment matrix, Ι and K are the filter window sizes; S54. Use the optimized confidence enhancement matrix G as the input, update the automated annotation model, and improve the annotation accuracy of low-confidence areas.

7. A method for automatically annotating autonomous driving scenario data based on incremental learning according to claim 1, characterized in that, The specific content of S6 includes: S61. Obtain newly added multi-source data, perform preprocessing, combine with the confidence enhancement matrix G to construct an incremental feature sample set, normalize the newly added data through a sample balancing strategy, match the feature distribution of existing labeled samples, and calculate the feature similarity score; S62. Calculate the feature distribution deviation for the incremental feature sample set, identify the feature drift region, set a threshold to filter valid incremental samples, use a statistical distribution comparison method to calculate the KL divergence between the feature distribution of the newly added samples and the existing data, and eliminate high-deviation samples to ensure the stability of the incremental training data; S63. Calculate the topological residuals for the filtered incremental feature sample set, map the features of the new samples to the existing topological relationship matrix T, and generate an incremental topological matrix T ′ , and adjust the topological structure through an adaptive regularization strategy to reduce the distribution drift while maintaining the topological consistency of the new samples; S64. Based on the incremental topology matrix T ′ Calculate the incremental learning loss function, optimize the automatic annotation model using regularization constraints, and obtain the projection mapping three-update matrix: Among them, W″′ is the projection mapping three - time update matrix, and W ″ is the projection mapping two - time update matrix, W is the projection mapping matrix, T′ i is the i - th sample feature vector of the incremental topology matrix, Y i is the annotation vector corresponding to the sample, λ is the regularization coefficient, m is the number of data samples, argmin is the independent variable value when the function takes the minimum value, and the projection mapping matrix W is adjusted to optimize the annotation accuracy; S65. Apply the projection mapping three-time updated matrix W″′ to the newly added multi-source data, perform automatic annotation prediction, and store the annotation results. At the same time, trigger a dynamic adjustment mechanism for low-confidence prediction regions to improve the annotation accuracy.

Images