A data annotation method and system for autonomous driving
By using adaptive discrete wave domain-topology joint optimization and topology residual projection matrix update, the problems of blurred boundaries and poor handling of low-confidence regions in autonomous driving data annotation are solved, achieving efficient and accurate data annotation iteration and model adaptability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEBEI BINSONG TECHNOLOGY CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-06-02
AI Technical Summary
Existing autonomous driving data annotation technologies are difficult to adapt to complex and dynamic autonomous driving scenarios. They suffer from poor handling of boundaries and low-confidence regions, low model iteration efficiency, and high computational overhead. Furthermore, incremental learning schemes cannot effectively balance the differences between new and old data.
An adaptive discrete wave domain-topology joint optimization method is adopted to initially label multi-source sensor data, select high-value samples for manual verification and optimization, and use the topological residual projection matrix to update the model to achieve closed-loop iteration.
It improves the accuracy and consistency of annotation results, reduces computational overhead, enables efficient and accurate model iteration, and allows for timely feedback and closed-loop iteration to adapt to new scenarios and problems.
Smart Images
Figure CN122132814A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data annotation, and in particular to a data annotation method and system for autonomous driving. Background Technology
[0002] The implementation and development of autonomous driving technology heavily rely on high-quality, large-scale labeled data. Environmental data collected by multi-source sensors needs to be accurately labeled before it can serve as the core basis for model training and iteration. However, existing autonomous driving data labeling technologies still have many key problems, making it difficult to adapt to the complexity, dynamism, and long-tail requirements of autonomous driving scenarios. These problems are mainly reflected in the following aspects:
[0003] Insufficient accuracy of automated annotation and poor handling of boundaries and low-confidence regions: Existing automated annotation models are prone to blurring and offsetting of the boundary regions of objects in the scene. At the same time, they have low confidence in the annotation of complex scenes such as lighting changes, target occlusion, and motion blur, making it difficult to accurately identify low-quality annotated regions, which greatly reduces the effectiveness of the annotation results.
[0004] Existing annotation models mostly adopt a full retraining mode for updates, which has high computational overhead and low iteration efficiency. Some incremental learning schemes cannot effectively balance the learning of new data and the retention of old knowledge, which can easily lead to the "catastrophic forgetting" problem where the model's ability to annotate historical scenes declines. Furthermore, it is difficult to quantify the difference between new data and existing knowledge topology, and it is impossible to achieve accurate micro-updates of the model. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a more accurate data annotation method and system that can be updated in a closed loop.
[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:
[0007] A data annotation method for autonomous driving, the key of which includes the following steps:
[0008] Acquire multi-source sensor data for autonomous driving to be labeled, and perform data preprocessing and modal alignment;
[0009] Preliminary entity annotation results are generated by using a pre-trained automated annotation model to perform initial annotation on the processed multi-source sensor data.
[0010] Adaptive discrete wave domain-topology joint optimization is performed on the preliminary entity annotation results;
[0011] Based on a deep active learning strategy, high-value samples are automatically evaluated and selected from the data after joint optimization, with the information gain for adapting to new scenarios as the maximum screening criterion.
[0012] The selected high-value samples are submitted to manual review and correction to obtain high-quality ground truth labels. At the same time, the low-confidence labeled areas are dynamically optimized using an adaptive filtering method to generate supplementary optimized labels.
[0013] The high-quality ground truth labels and the supplementary optimized labels are used together as training data for incremental learning to drive the updating of the mapping matrix based on topological residual projection.
[0014] The updated labeled model is deployed to the autonomous driving system for road testing, and its performance is monitored. New scene data that is not well processed by the model is collected, and the above labeling and optimization process is retried to carry out closed-loop iteration.
[0015] Preferably, the data preprocessing and modal alignment specifically include:
[0016] The 3D point cloud data is projected into a bird's-eye view and rasterized. Taking the vehicle as the origin, the 3D point cloud data is projected onto the ground plane and divided into grids. The grid is encoded as a three-channel feature of maximum height, point cloud density, and reflection intensity. A 2D bird's-eye view feature map is generated based on the grid data. Simultaneously, the camera image, millimeter-wave radar data, and 2D bird's-eye view feature map are mapped to the same coordinate system to obtain a 2D multimodal feature map.
[0017] Preferably, the adaptive discrete wave domain-topology joint optimization specifically includes:
[0018] The preliminary entity annotation results are decomposed into sub-bands containing low-frequency approximation components and high-frequency detail components by using discrete wavelet transform.
[0019] The high-frequency detail components obtained from the decomposition are enhanced and selectively reconstructed to sharpen object boundaries;
[0020] Structural constraints are applied to the wavelet enhancement region based on the topological residual perturbation matrix.
[0021] The reconstructed multi-scale features are matched and evaluated, and regions with a label confidence level lower than a set threshold are selected as the low-confidence label regions.
[0022] Preferably, the deep active learning-based strategy, using maximizing information gain for adaptation to new scenarios as the selection criterion, specifically includes:
[0023] The calculation of the information gain integrates at least one of the following uncertainty measures: the entropy H(y|x) of the model prediction, the BALD(x) value based on Bayesian active learning divergence, and the task-level uncertainty for autonomous driving planning tasks.
[0024] The task-level uncertainties include displacement errors in the prediction of the vehicle's trajectory, the risk of soft collisions between the vehicle and the trajectories of surrounding traffic participants, and the confidence entropy of the prediction of the trajectories of surrounding intelligent agents.
[0025] Preferably, the driving mechanism updates the mapping matrix based on the topological residual projection, specifically including:
[0026] The autonomous driving scenario can be understood as a topological structure in a high-dimensional feature space;
[0027] When processing new data, a wave domain-topology perturbation matrix is constructed. Quantitative analysis of the residual shift between new data features and existing knowledge topology. The formula for the joint wave domain-topology perturbation matrix is as follows:
[0028]
[0029] in To label the residual matrix, It is a topological relation matrix. The projection mapping matrix, The Laplacian operator for the residual matrix is given, and WEMap is the wavelet boundary enhancement confidence map. and This is the balance coefficient;
[0030] Based on the residual analysis, incremental learning is used to dynamically adjust the projection mapping matrix, thereby achieving incremental learning of the model.
[0031] Preferably, the step of dynamically adjusting the projection mapping matrix using incremental learning specifically involves:
[0032] The model before the update is used as the teacher model, and the model to be updated is used as the student model. During incremental training, the student model learns the scene topology distribution and the boundary feature distribution after wave domain optimization from the teacher model. A consistency constraint loss is constructed based on the wave domain-topology joint perturbation matrix to constrain the output differences between the student model and the teacher model in low confidence regions, boundary regions, and new scene regions. Only the high residual regions indicated by the joint perturbation matrix are updated with weights, keeping the original scene feature mapping unchanged.
[0033] A data annotation system for autonomous driving, used to implement the above method, characterized in that it includes:
[0034] The data acquisition and preprocessing module is used to acquire and synchronize multi-source sensor data;
[0035] The initial annotation module integrates an automated annotation model for performing preliminary entity annotation on preprocessed data;
[0036] The joint optimization module performs adaptive discrete wave domain-topology joint optimization on the preliminary entity annotation results;
[0037] The deep active learning module evaluates unlabeled data based on information gain and automatically filters high-value samples.
[0038] The human-machine collaborative review and optimization module presents the high-value samples for human review and performs algorithm optimization on the low-confidence regions.
[0039] The topological residual projection and incremental learning module analyzes the residuals between new data and existing knowledge topology based on manual review results and optimization results for low-confidence regions, and updates the annotation model accordingly.
[0040] The closed-loop iteration management module manages the deployment of updated models, road test monitoring, and the collection of new problem data to drive a new round of iterations.
[0041] Preferably, the joint optimization module specifically includes:
[0042] The wavelet transform unit is used to perform a two-dimensional discrete wavelet transform on the input labeled image and perform multi-level decomposition.
[0043] The high-frequency processing unit is used to process and reconstruct the high-frequency subband coefficients obtained from the decomposition.
[0044] Topological constraint cells are used to construct topological residuals and apply structure-preserving constraints to the wavelet enhancement region.
[0045] The confidence assessment unit is used to calculate the confidence level of the region label and output the low-confidence region identifier.
[0046] Preferably, the deep active learning module specifically includes:
[0047] The information gain calculation unit calculates the prediction entropy, Bayesian uncertainty, and task-related uncertainty for each sample in the unlabeled data pool, and integrates them into a comprehensive information gain score.
[0048] The sample sorting and filtering unit sorts all samples according to the comprehensive information gain score and selects a predetermined number of samples with the highest scores as the high-value samples for output.
[0049] The beneficial effects of adopting the above technical solution are as follows:
[0050] This invention performs multi-scale feature decomposition and high-frequency feature reconstruction on the preliminary annotation results based on adaptive discrete wave domain-topology joint optimization, effectively sharpening object boundaries and improving the clarity of annotation edges. It solves the problem of blurred boundary annotations in scenarios such as motion blur, low light, and target occlusion, and improves the accuracy and consistency of the overall annotation results.
[0051] This invention proposes a mapping matrix update scheme based on topological residual projection. It quantitatively analyzes the residual offset between new data and existing knowledge by constructing a topological residual perturbation matrix. Combined with incremental learning, it dynamically adjusts the projection mapping matrix. The projection mapping matrix is only slightly modified based on incremental data. The computational cost is much lower than full retraining, and it can effectively constrain the retention of old knowledge, thereby achieving efficient and accurate iteration of the annotation model and improving the accuracy of the annotation system.
[0052] This invention calculates information gain by integrating prediction entropy, Bayesian divergence, and multi-dimensional indicators of autonomous driving task-level uncertainty. This information gain serves both as a high-value sample selection tool and as an iteration threshold, enabling new scenario problems and annotation defects discovered after model deployment to be promptly fed back to the annotation stage, thus completing a closed-loop iteration. Attached Figure Description
[0053] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0054] Figure 1 This is a schematic diagram of a data annotation method for autonomous driving proposed in this invention. Detailed Implementation
[0055] To make the above-mentioned objectives, features, and advantages of the present invention more apparent and understandable, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0056] This invention proposes a data annotation method and system for autonomous driving, logically forming a complete closed loop from data acquisition and intelligent processing to model evolution. Its core data flow follows a path of "perception-analysis-optimization-iteration," such as... Figure 1 The specific steps are as follows:
[0057] Acquire multi-source sensor data for autonomous driving to be labeled, and perform data preprocessing and modal alignment;
[0058] Preliminary entity annotation results are generated by using a pre-trained automated annotation model to perform initial annotation on the processed multi-source sensor data.
[0059] Adaptive discrete wave domain-topology joint optimization is performed on the preliminary entity annotation results;
[0060] Based on a deep active learning strategy, high-value samples are automatically evaluated and selected from the data after wave domain adaptation processing, with the information gain of adapting to new scenarios as the maximum screening criterion.
[0061] The selected high-value samples are submitted to manual review and correction to obtain high-quality ground truth labels. At the same time, the low-confidence labeled areas are dynamically optimized using an adaptive filtering method to generate supplementary optimized labels.
[0062] High-quality ground truth labels and supplementary optimized labels are used together as training data for incremental learning to drive the updating of the mapping matrix based on topological residual projection.
[0063] The updated labeled model is deployed to the autonomous driving system for road testing, and its performance is monitored. New scene data that is not well processed by the model is collected, and the above labeling and optimization process is retried to carry out closed-loop iteration.
[0064] Multi-source data input and preprocessing: The system receives multi-source sensor data from autonomous vehicles in real time, including but not limited to LiDAR point clouds, camera images, and millimeter-wave radar. The preprocessing unit performs bird's-eye view projection and rasterization encoding on the 3D point cloud data. With the vehicle as the origin, the 3D point cloud data is projected onto the ground plane and divided into grids. The grid encoding is a three-channel feature of maximum height, point cloud density, and reflection intensity. A 2D bird's-eye view feature map is generated based on the grid data. Simultaneously, the camera images, millimeter-wave radar data, and the 2D bird's-eye view feature map are mapped to the same coordinate system to obtain a 2D multimodal feature map.
[0065] Feature extraction and preliminary annotation: The preprocessed data is fed into a feature extraction network. In practice, a fusion network based on ResNet, PointNet++, or other architectures can be used to generate high-dimensional features. Subsequently, using the initialized or updated topological residual projection module from the previous iteration, the features are mapped to a low-dimensional topological relation space, and the preliminary entity annotation results for static and dynamic targets in the scene, along with the corresponding annotation residuals, are output.
[0066] Adaptive Discrete Wave Domain-Topology Joint Optimization: Preliminary annotation results, especially those of boundary regions, are fed into the joint optimization module for multi-scale decomposition and high-frequency reconstruction to improve boundary accuracy. Simultaneously, the system integrates annotation residuals, wave domain analysis results, and uncertainty estimates from the model itself to automatically identify regions with low annotation confidence.
[0067] This module is specifically designed to handle boundary blurring and target occlusion issues in annotation, and achieves global-local collaborative optimization through feedback with the topology module.
[0068] Step 1: Multi-scale wave domain decomposition
[0069] The preliminary annotation results output by the topological residual projection module are decomposed into multiple scales using two-dimensional discrete wavelet transform.
[0070] Choose a suitable wavelet basis function. Specifically, you can choose Daubechies'db4 or the bioorthogonal CDF9 / 7.
[0071] Perform one-level or multi-level decomposition. Taking one-level decomposition as an example, the labeled region image I is decomposed into four sub-bands:
[0072] LL (Low-Frequency Approximation): Contains the main outline and background information of the image.
[0073] LH (Horizontal Detail), HL (Vertical Detail), and HH (Diagonal Detail): These collectively contain high-frequency detail information such as the edges and textures of an image, and the precise boundary information of objects is mainly contained within them.
[0074] Step 2: High-frequency feature reconstruction and boundary enhancement
[0075] The blurred boundaries of an object correspond to the attenuation of high-frequency information or noise interference. The core of this step is to enhance the useful edge signals in the high-frequency subbands while suppressing noise.
[0076] The high-frequency subband coefficients of LH, HL, and HH are analyzed. By setting an adaptive threshold, edge coefficients exceeding the threshold are retained, coefficients below the threshold are attenuated, and noise regions are filtered out.
[0077] The processed high-frequency coefficients are combined with the LL subband and subjected to inverse discrete wavelet transform to reconstruct an image of the labeled region with clearer and sharper boundaries. This process can effectively improve the inaccurate boundary problems caused by motion blur, low light, or partial occlusion.
[0078] The third step is topological constraints. In the process of wavelet multi-scale decomposition and high-frequency reconstruction of the preliminary entity annotation results, only enhancing and sharpening the high-frequency detail components can easily lead to problems such as over-sharpening of boundaries, local area breakage, and distortion of target contours. This can destroy the overall topological structure of the relative positional relationships, connectivity relationships, and hierarchical relationships between targets in the autonomous driving scenario, resulting in the annotation results not conforming to the physical constraints of the real traffic scenario in terms of spatial structure.
[0079] Therefore, this invention introduces topological constraints based on wavelet boundary enhancement, and the specific implementation process is as follows:
[0080] Based on the target space adjacency relationship, relative position and connectivity, a scene structure prior is constructed, and a topological constraint mask is generated by combining the labeled residual distribution. The wavelet enhancement intensity is dynamically adjusted according to the mask. The enhancement amplitude is reduced in stable structural regions, and the boundary is sharpened in low confidence edge regions while maintaining structural integrity, so as to ensure that only the sharpness is optimized without changing the scene topology.
[0081] Step 4: Identification and Optimization of Low-Confidence Regions
[0082] Recognition: The features of the reconstructed image in the wave domain are matched with the original predicted features, and the difference is calculated. Simultaneously, the topological residual perturbation matrix P is combined. t Regions with large local gradients. Both indicate regions with low confidence in the system's annotations.
[0083] Optimization: For these low-confidence regions, a bilateral filter combining spatial and range information is used for dynamic refinement to obtain optimized annotation results.
[0084] Information gain-based active learning and sample selection: The system's built-in deep active learning module analyzes massive unlabeled data pools or low-confidence optimization regions, and uses an information gain calculation method that integrates task uncertainty to automatically select the batch of samples that are most valuable for improving the model's adaptability to new scenarios.
[0085] This module drives the sample selection of the entire closed-loop system, with the goal of selecting those samples that best expose the blind spots in the model's cognition and improve its adaptability to new scenarios.
[0086] Step 1: Define and quantify the cognitive blind spots of the model
[0087] The system defines the cognitive blind spot from three dimensions and provides a basis for subsequent information gain calculation:
[0088] Physical perception blind spot: The unobservable grid area calculated based on the sensor's field of view and occlusion.
[0089] Data knowledge blind spots: The model's lack of understanding of rare or complex scenarios not covered by the training data.
[0090] Task reasoning blind spot: The area where the model produces high uncertainty or erroneous outputs in tasks such as trajectory prediction and interactive understanding.
[0091] Step 2: Designing multi-dimensional information gain calculation criteria
[0092] For each sample x in the unlabeled data pool, the system calculates its information gain value to quantify the expected value of labeling that sample. The main calculation methods include:
[0093] Predicting entropy :
[0094]
[0095] Entropy represents the posterior probability that the model labels sample x as class c. The higher the entropy value, the less certain the model is about the class of that sample.
[0096] Bayesian active learning of divergence :
[0097]
[0098] in This method represents the mutual information between model parameters θ and labeled results y under sample x and dataset D. It tends to select samples where different settings of model parameters produce highly divergent predictions and can effectively identify samples that help clarify model parameters.
[0099] Task-level uncertainty: For autonomous driving planning tasks, metrics such as trajectory displacement error, soft collision risk, and agent prediction entropy are calculated. These metrics directly reflect the model's cognitive shortcomings in safety-critical tasks, and specific scores are determined based on practical experience and expert opinions.
[0100] The multidimensional information gain value is quantified by combining the above three aspects and the final result is obtained by weighted summation. The weights of the three aspects are determined by which aspects are emphasized more in actual use.
[0101] Step 3: Sample Screening and Batch Construction
[0102] The system comprehensively applies the above criteria to score and rank the unlabeled data pool. A greedy strategy of batch active learning is employed, selecting a batch of samples that maximizes the overall information gain in each iteration and submitting it for manual review.
[0103] Manual key verification and correction data generation: High-value samples selected are submitted to the manual annotation platform for key verification and correction. The corrected results form high-quality ground truth labels; at the same time, for the low-confidence annotation regions identified in step 3, an adaptive filtering method is used to dynamically optimize the labeling using a bilateral filtering algorithm that combines spatial and range information. The optimized results are automatically verified by the system to form supplementary optimized labels.
[0104] Incremental learning and model evolution: High-quality ground truth labels are merged with supplementary optimized labels as training data for incremental learning, driving the update of the mapping matrix based on topological residual projection. Specifically:
[0105] Step 1: Construction and Representation of High-Dimensional Feature Space Topology
[0106] Suppose that after passing through the feature extraction network, we obtain a feature space matrix F∈R containing m samples. m×n Where n is the feature dimension. This invention treats F as a point cloud in a high-dimensional space, and to characterize its topological structure, it employs continuous cohomology for multi-scale analysis:
[0107] Based on the point cloud F, a distance threshold ε is defined.
[0108] Let ε gradually increase from 0 to construct a sequence of simple complexes. When the distance between two points is less than ε, connect them to form an edge (1-simplex); when the three edges form a triangle, they form a face (2-simplex), and so on.
[0109] During this filtering process, the "birth" and "death" parameters of topological features in each dimension are recorded. The results are represented as persistence maps or barcodes. Among them, topological features with long persistence are regarded as stable structural signals in the scene.
[0110] Step 2: Initialization and calculation of the topological residual perturbation matrix
[0111] The core of this module is learning a projection mapping matrix W∈R. n×k Map F to the dimension-reduced topological relation matrix T = F·W, T∈R m×k T should reflect the stable topological pattern in F.
[0112] After initial annotation, calculate the global annotation residual matrix R = FT·W T To quantitatively analyze the local offset between new data features and existing knowledge topology, a topological residual perturbation matrix is constructed. Its initialization formula is: Where ∇²R is the Laplacian operator of the residual matrix R, used to extract local gradient information, emphasizing drastic changes in the residuals at boundaries and in uncertain regions. WEMap is a wavelet boundary enhancement confidence map, where α, β, and This is an adjustable balance coefficient. It encodes the structural errors of the model under a specific data distribution.
[0113] Step 3: Dynamically adjust the projection mapping matrix using incremental learning.
[0114] The model before the update is used as the teacher model, and the model to be updated is used as the student model. During incremental training, the student model learns the scene topology distribution and the boundary feature distribution after wave domain optimization from the teacher model. A consistency constraint loss is constructed based on the wave domain-topology joint perturbation matrix to constrain the output differences between the student model and the teacher model in low confidence regions, boundary regions, and new scene regions. Only the high residual regions indicated by the joint perturbation matrix are updated with weights, keeping the original scene feature mapping unchanged.
[0115] The updated student model was redeployed to process the new data and initiate a new cycle of "road test monitoring - issue collection," thereby enabling the system to continuously evolve.
[0116] Based on the above, the complete closed-loop iterative process of this invention is implemented as follows:
[0117] System initialization: Use a basic dataset to train the initial feature extraction network and topological residual projection module to form the initial automated annotation model M0.
[0118] Road Test Data Retrieval and Triggering: Deploy M0 on test vehicles for road testing. The system monitors model performance in real time and automatically retrieves the complete sensor data sequence for the current and preceding / following time windows when the following conditions are triggered:
[0119] 1) The model output confidence level is lower than the threshold;
[0120] 2) There is a significant difference between the predicted trajectory and the driver's actual operation;
[0121] 3) Potentially high-risk scenarios were detected.
[0122] Automated processing and sample selection: The collected data enters the cloud processing platform. First, it undergoes automatic annotation and optimization by the joint optimization module. Subsequently, the active learning module calculates on a large number of unlabeled or low-quality labeled frames to select samples with high information gain.
[0123] Human-machine collaborative review: High-value samples selected are pushed to the manual annotation platform. Annotators use the intelligent pre-annotation and auxiliary tools provided by the system to perform efficient review and key corrections, generating high-quality labels.
[0124] Incremental learning and model update: Manually corrected data is merged with low-confidence region data automatically optimized by the system to form an incremental dataset. The projection mapping matrix is dynamically adjusted using incremental learning to obtain a new model M1. This process emphasizes the preservation of old knowledge under constraints.
[0125] Validation and Redeployment: Validate M1 in a large-scale simulation testing system to ensure its performance is no less than M0 in both challenging and routine scenarios. After successful validation, deploy M1 to the vehicle and return to the road test data collection and triggering step to begin the next iteration.
[0126] Through the aforementioned closed-loop process, the system can continuously identify problems and collect key data from real-world road tests, driving the update of the topological residual projection mapping matrix W, thereby updating the entire automated annotation model. The updated model is applied to new data annotation, and its performance is then monitored through road tests to identify new long-tail problems, triggering a new round of data collection and filtering, forming a continuously self-improving data closed loop. This achieves a spiral increase in annotation accuracy and scene adaptability, forming a self-driven, continuously evolving autonomous driving data annotation and model training system.
[0127] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A data annotation method for autonomous driving, characterized in that, Includes the following steps: Acquire multi-source sensor data for autonomous driving to be labeled, and perform data preprocessing and modal alignment; Preliminary entity annotation results are generated by using a pre-trained automated annotation model to perform initial annotation on the processed multi-source sensor data. Adaptive discrete wave domain-topology joint optimization is performed on the preliminary entity annotation results; Based on a deep active learning strategy, high-value samples are automatically evaluated and selected from the data after joint optimization, with the information gain for adapting to new scenarios as the maximum screening criterion. The selected high-value samples are submitted to manual review and correction to obtain high-quality ground truth labels. At the same time, the low-confidence labeled areas are dynamically optimized using an adaptive filtering method to generate supplementary optimized labels. The high-quality ground truth labels and the supplementary optimized labels are used together as training data for incremental learning to drive the updating of the mapping matrix based on topological residual projection. The updated labeled model is deployed to the autonomous driving system for road testing, and its performance is monitored. New scene data that is not well processed by the model is collected, and the above labeling and optimization process is retried to carry out closed-loop iteration.
2. The data annotation method for autonomous driving according to claim 1, characterized in that, The data preprocessing and modal alignment specifically include: The 3D point cloud data is projected into a bird's-eye view and rasterized. Taking the vehicle as the origin, the 3D point cloud data is projected onto the ground plane and divided into grids. The grid is encoded as a three-channel feature of maximum height, point cloud density, and reflection intensity. A 2D bird's-eye view feature map is generated based on the grid data. Simultaneously, the camera image, millimeter-wave radar data, and 2D bird's-eye view feature map are mapped to the same coordinate system to obtain a 2D multimodal feature map.
3. The data annotation method for autonomous driving according to claim 1, characterized in that, The adaptive discrete wave domain-topology joint optimization specifically includes: The preliminary entity annotation results are decomposed into sub-bands containing low-frequency approximation components and high-frequency detail components by using discrete wavelet transform. The high-frequency detail components obtained from the decomposition are enhanced and selectively reconstructed to sharpen object boundaries; Structural constraints are applied to the wavelet enhancement region based on the topological residual perturbation matrix. The reconstructed multi-scale features are matched and evaluated, and regions with a label confidence level lower than a set threshold are selected as the low-confidence label regions.
4. The data annotation method for autonomous driving according to claim 1, characterized in that, The deep active learning-based strategy, using maximizing information gain for adaptation to new scenarios as the selection criterion, specifically includes: The calculation of the information gain integrates at least one of the following uncertainty measures: the entropy H(y|x) of the model prediction, the BALD(x) value based on Bayesian active learning divergence, and the task-level uncertainty for autonomous driving planning tasks. The task-level uncertainties include displacement errors in the prediction of the vehicle's trajectory, the risk of soft collisions between the vehicle and the trajectories of surrounding traffic participants, and the confidence entropy of the prediction of the trajectories of surrounding intelligent agents.
5. A data annotation method for autonomous driving according to claim 1, characterized in that, The driver updates based on the mapping matrix of the topological residual projection, specifically including: The autonomous driving scenario can be understood as a topological structure in a high-dimensional feature space; When processing new data, a wave domain-topology perturbation matrix is constructed. The residual offset between new data features and existing knowledge topology is quantitatively analyzed. The formula for the joint wave domain-topology perturbation matrix is as follows: ； in To label the residual matrix, It is a topological relation matrix. The projection mapping matrix, The Laplacian operator for the residual matrix is given, and WEMap is the wavelet boundary enhancement confidence map. and This is the balance coefficient; Based on the residual analysis, incremental learning is used to dynamically adjust the projection mapping matrix, thereby achieving incremental learning of the model.
6. A data annotation method for autonomous driving according to claim 5, characterized in that, The method of dynamically adjusting the projection mapping matrix using incremental learning specifically involves: The model before the update is used as the teacher model, and the model to be updated is used as the student model. During incremental training, the student model learns the scene topology distribution and the boundary feature distribution after wave domain optimization from the teacher model. A consistency constraint loss is constructed based on the wave domain-topology joint perturbation matrix to constrain the output differences between the student model and the teacher model in low confidence regions, boundary regions, and new scene regions. Only the high residual regions indicated by the joint perturbation matrix are updated with weights, keeping the original scene feature mapping unchanged.
7. A data annotation system for autonomous driving, used to implement the method according to any one of claims 1-6, characterized in that, include: The data acquisition and preprocessing module is used to acquire and synchronize multi-source sensor data; The initial annotation module integrates an automated annotation model for performing preliminary entity annotation on preprocessed data; The joint optimization module performs adaptive discrete wave domain-topology joint optimization on the preliminary entity annotation results; The deep active learning module evaluates unlabeled data based on information gain and automatically filters high-value samples. The human-machine collaborative review and optimization module presents the high-value samples for human review and performs algorithm optimization on the low-confidence regions. The topological residual projection and incremental learning module analyzes the residuals between new data and existing knowledge topology based on manual review results and optimization results for low-confidence regions, and updates the annotation model accordingly. The closed-loop iteration management module manages the deployment of updated models, road test monitoring, and the collection of new problem data to drive a new round of iterations.
8. A data annotation system for autonomous driving according to claim 7, characterized in that, The joint optimization module specifically includes: The wavelet transform unit is used to perform a two-dimensional discrete wavelet transform on the input labeled image and perform multi-level decomposition. The high-frequency processing unit is used to process and reconstruct the high-frequency subband coefficients obtained from the decomposition. Topological constraint cells are used to construct topological residuals and apply structure-preserving constraints to the wavelet enhancement region. The confidence assessment unit is used to calculate the confidence level of the region label and output the low-confidence region identifier.
9. A data annotation system for autonomous driving according to claim 7, characterized in that, The deep active learning module specifically includes: The information gain calculation unit calculates the prediction entropy, Bayesian uncertainty, and task-related uncertainty for each sample in the unlabeled data pool, and integrates them into a comprehensive information gain score. The sample sorting and filtering unit sorts all samples according to the comprehensive information gain score and selects a predetermined number of samples with the highest scores as the high-value samples for output.