Data annotation method and device and electronic equipment

Abstract

The invention relates to a data labeling method and device and electronic equipment, and the method comprises the steps: obtaining lane line point cloud data, a user view angle position and user view angle change in response to a labeling task request; based on three-dimensional space coordinates corresponding to the lane line point cloud data, determining spatial distribution characteristics, local density, curvature change conditions and reflection intensity characteristics, and then adjusting a slice granularity threshold value; based on the spatial distribution characteristics and a slice granularity threshold value, performing spatial block slicing processing on the lane line point cloud data to generate a point cloud slice set; based on the point cloud slice set, establishing a point cloud slice database containing a space-time four-dimensional index; loading a visual point cloud slice from a point cloud slice database based on the user visual angle position and the user visual angle change; based on the space-time four-dimensional index, the annotation association is established, and the visual point cloud slices are annotated, so that the problems of memory bottleneck, rendering lagging and annotation breakage in point cloud processing are solved, and collaborative improvement of precision, efficiency and system performance is realized.

Description

Technical Field

[0001] This invention relates to the field of indoor data annotation technology, and more specifically, to a data annotation method, apparatus, and electronic device. Background Technology

[0002] With the rapid development of autonomous driving, high-precision map production, and intelligent transportation systems, 4D point cloud data collected by LiDAR (LiDAR) (which adds a millisecond-level time dimension to traditional three-dimensional spatial coordinates) has become a core data source for building high-precision environmental perception models. Among these, lane lines, as a key element of road structure, are fundamental to ensuring the safety of autonomous vehicle positioning, path planning, and decision-making. Currently, such annotation tasks mainly rely on data annotation cloud platforms, where professionals interactively annotate point cloud data through a visual interface.

[0003] However, existing technologies face the following prominent technical bottlenecks when handling large-scale, temporal 4D point cloud lane line annotation tasks: excessive memory consumption and poor rendering performance. Traditional point cloud visualization and annotation tools (such as solutions based on open-source frameworks like Potree) typically employ a full data loading mode. When dealing with large-scale urban scene point clouds with millions or even tens of millions of points per frame, the complete data must be loaded into the client's memory at once. Real-world testing shows that this operation often results in memory consumption exceeding 32GB, easily triggering browser out-of-memory (OOM) errors and crashes. The system crash rate in complex scenes can reach over 80%. Simultaneously, when rendering such large-scale point clouds using WebGL on the web, even with general LOD (Level of Detail) technology, the frame rate is generally below 5 FPS, causing severe stuttering during view scaling, translation, and rotation, resulting in a poor annotation experience and significantly impacting annotation efficiency. Furthermore, temporal annotation breaks and consistency assurance are difficult. Most existing annotation platforms treat 4D point cloud data as a series of independent 3D point cloud frames, lacking an effective spatiotemporal correlation mechanism. Annotators must manually annotate lane lines frame by frame on the timeline, which is not only labor-intensive but also makes it difficult to ensure the continuity and consistency of annotation results over time. Vehicle movement and changes in viewpoint can cause the same lane line to appear differently in different frames, and manual annotation is prone to problems such as misalignment, breakage, or inconsistent attributes. Although manual review can be performed later for calibration, it is inefficient and cannot fundamentally solve the technical challenge of cross-frame semantic coherence. Data scheduling is not intelligent and the interactive experience is poor. Existing solutions mostly use static or simple cropping based on a fixed viewpoint, failing to deeply integrate with the user's real-time interactive behavior. For example, when the user pans or rotates the viewpoint, the system cannot predict their intention and preload the data that will enter the viewpoint in advance, resulting in significant loading delays and waiting with each viewpoint switch. In addition, for data far from the current viewpoint or non-critical areas, there is a lack of effective dynamic unloading and caching management mechanisms, resulting in continuous invalid occupation of memory resources. There is a lack of scene-adaptive optimization for lane line annotation, and general point cloud processing frameworks have not been deeply optimized for the specific task of lane line annotation. For example, in the spatial organization of point clouds, the differences in detail requirements between lane and non-lane areas were not considered, and a uniform subdivision granularity was used. This either resulted in insufficient detail in critical areas or a large amount of redundant data in non-critical areas, affecting processing and transmission efficiency. Furthermore, data compression and transmission were not differentiated according to the importance of the content.

[0004] In summary, existing technologies have significant shortcomings in terms of system performance, annotation efficiency, result consistency, and user experience when handling large-scale 4D point cloud lane line annotation tasks on cloud platforms. Therefore, there is an urgent need for a new 4D point cloud data processing method that can achieve efficient memory management, intelligent data scheduling, and temporal consistency guarantees, while being optimized for annotation tasks, to support the growing demand for high-quality annotation data in high-precision applications such as autonomous driving. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a data annotation method, apparatus and electronic device to solve the prominent problems existing in the prior art, such as excessive memory consumption, time-series annotation breakage and difficulty in ensuring consistency, unintelligent data scheduling and poor interactive experience, low rendering efficiency and scene adaptation.

[0006] One aspect of the present invention provides a data annotation method, the method comprising the following steps: In response to the annotation task request, lane line point cloud data, user view position, and user view changes are acquired; the lane line point cloud data includes the three-dimensional spatial coordinates corresponding to the point cloud data; and based on the three-dimensional spatial coordinates of the lane line point cloud data, the spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics of the lane line point cloud data are determined. Based on the point cloud density, curvature variation and reflection intensity characteristics corresponding to the lane line point cloud data, the slice granularity threshold is dynamically adjusted. Based on the spatial distribution characteristics of the lane line point cloud data and the slice granularity threshold, a spatial segmentation algorithm is used to perform spatial block slicing on the lane line point cloud data to generate a point cloud slice set. Based on the point cloud slice set, a point cloud slice database containing a spatiotemporal four-dimensional index is established; and based on the user's viewpoint position and the changes in the user's viewpoint, visible point cloud slices within the visible area are dynamically loaded from the point cloud slice database. Based on the spatiotemporal four-dimensional index in the point cloud slice database, a cross-time frame annotation association is established using the temporal consistency module, and the visible point cloud slice is annotated based on the annotation association.

[0007] Another aspect of the present invention provides a data annotation apparatus, the apparatus comprising: The acquisition module is used to respond to the annotation task request and acquire lane line point cloud data, user view position, and user view changes; the lane line point cloud data includes the three-dimensional spatial coordinates corresponding to the point cloud data; and based on the three-dimensional spatial coordinates of the lane line point cloud data, the spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics of the lane line point cloud data are determined. The adjustment module is used to dynamically adjust the slice granularity threshold based on the point cloud density, curvature change and reflection intensity characteristics corresponding to the lane line point cloud data. The slicing module is used to perform spatial segmentation slicing of the lane line point cloud data based on the spatial distribution characteristics of the lane line point cloud data and the slicing granularity threshold, and to generate a point cloud slice set by using a spatial segmentation algorithm. The loading module is used to establish a point cloud slice database containing a spatiotemporal four-dimensional index based on the point cloud slice set; and dynamically load visible point cloud slices within the visible area from the point cloud slice database based on the user's view position and the changes in the user's view. The annotation module is used to establish annotation associations across time frames based on the spatiotemporal four-dimensional index in the point cloud slice database, using the temporal consistency module, and to annotate the visible point cloud slices based on the annotation associations.

[0008] Another aspect of the present invention provides an electronic device, comprising: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the data annotation method described in any one of the above descriptions.

[0009] Another aspect of the present invention provides a computer-readable storage medium that, when the instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to perform the data annotation method described in any one of the above descriptions.

[0010] This invention provides a data annotation method, apparatus, and electronic device. Responding to annotation task requests, the system simultaneously acquires lane line point cloud data containing three-dimensional spatial coordinates, along with user viewpoint position and change information. Based on the point cloud coordinates, it extracts multi-dimensional features such as spatial distribution, local density, curvature changes, and reflection intensity. Subsequently, the system dynamically adjusts the slice granularity threshold according to these features, achieving an adaptive segmentation strategy of fine-grained segmentation for lane areas and coarse-grained segmentation for non-lane areas. Based on this, combining spatial distribution features and granularity thresholds, a spatial segmentation algorithm is used to intelligently divide the point cloud into blocks, generating a structured point cloud slice set, and constructing a point cloud slice database with a spatiotemporal four-dimensional index. During the interaction phase, point cloud slices within the visible and predicted areas are dynamically loaded from the database based on the real-time user viewpoint position and changes, achieving a smooth and lag-free browsing experience. Finally, utilizing the database's spatiotemporal index and temporal consistency module, annotation associations across time frames are established, allowing annotations completed by the user in the current frame to automatically propagate to corresponding slices in adjacent frames. This ensures efficient annotation while maintaining the continuity and consistency of annotation results in the temporal dimension. This method systematically solves the problems of memory bottleneck, rendering lag, and annotation fragmentation in large-scale 4D point cloud processing, achieving a synergistic improvement in accuracy, efficiency, and system performance. Attached Figure Description

[0011] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating a data annotation method according to an exemplary embodiment; Figure 2 This is a schematic diagram of a process for dynamically adjusting the slice granularity threshold according to an exemplary embodiment; Figure 3 This is a schematic diagram of a process for generating a point cloud slice set according to an exemplary embodiment; Figure 4 This is a schematic diagram of a process for establishing a point cloud slice database according to an exemplary embodiment; Figure 5 This is a schematic diagram of a process for dynamically loading visible point cloud slices within a visible area from a point cloud slice database, according to an exemplary embodiment. Figure 6 This is a schematic diagram of a process for preloading a visible point cloud slice corresponding to an upcoming field of view, according to an exemplary embodiment. Figure 7 This is a schematic diagram of a process for establishing annotation associations across time frames, provided according to an exemplary embodiment. Figure 8 This is a schematic diagram of a process for detecting and resolving annotation conflicts according to an exemplary embodiment; Figure 9 This is a schematic diagram of a data annotation device provided according to an exemplary embodiment. Detailed Implementation

[0013] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0015] It should be noted that the terms "first," "second," etc., in the specification and drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0016] Figure 1 This is a flowchart illustrating a data annotation method according to an exemplary embodiment. This specification provides method operation steps as shown in the embodiments or flowcharts, but based on conventional or non-inventive labor, more or fewer operation steps may be included. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual system or server product execution, the method can be executed in the order shown in the embodiments or drawings, or in parallel (e.g., in a parallel processor or multi-threaded processing environment). Specifically, as shown... Figure 1 As shown, using a data annotation cloud platform as the execution entity, an embodiment of a data annotation method according to this application is introduced. The method may include: S101: In response to the annotation task request, acquire lane line point cloud data, user view position, and changes in user view; the lane line point cloud data includes the three-dimensional spatial coordinates corresponding to the point cloud data; and based on the three-dimensional spatial coordinates of the lane line point cloud data, determine the spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics of the lane line point cloud data.

[0017] In a specific embodiment, the annotation task request can be an instruction initiated by a user (usually a data annotator) through the client interface of the data annotation cloud platform to manually or semi-automatically annotate a specific batch of lane line point cloud data. This request includes a task identifier, the associated point cloud dataset index, and possible initial annotation parameters, serving as the starting signal for triggering the entire background data processing flow. The lane line point cloud data can be a raw point cloud dataset collected by sensors such as LiDAR, primarily used for road lane line identification and annotation. Each data point contains at least three-dimensional spatial coordinates and typically also includes information such as reflection intensity and color. In this application, it specifically refers to data that requires lane line feature extraction and annotation. The user's viewpoint position can be the spatial coordinates of the user's virtual camera in the point cloud 3D visualization environment. Optionally, the user's viewpoint position defines the user's observation point in the 3D scene and is a core parameter for determining which point cloud data are within the current visible range. It is usually represented by the X, Y, and Z coordinates of the viewpoint. Changes in the user's viewpoint can be dynamic changes in the user's viewpoint position and direction, typically including the translational velocity, rotational angular velocity, and movement direction vector. It reflects the user's interaction intent and browsing trajectory, serving as a key input for the system to predict the area the user might focus on next, thus enabling data preloading; spatial distribution characteristics can be a quantitative description of the macroscopic distribution patterns of point cloud data in three-dimensional space. For example, the overall boundary of the point cloud, its center position, and the density distribution trend in different regions; local density can be the number of points contained in a unit volume within a specific point or a small neighborhood in the point cloud. It is a key feature for distinguishing different objects or regions; curvature variation can be a feature of the degree of curvature and rate of change of the local surface of the point cloud. By calculating the surface curvature fitted to a point in the point cloud and its neighboring points, regions with significant geometric changes, such as lane line edges, road shoulders, and obstacle corners, can be identified; reflection intensity characteristics can be the echo intensity information received by the lidar sensor, reflecting the target surface's ability to reflect laser light. Specifically, in response to a user's annotation task request initiated on the data annotation cloud platform, the system simultaneously acquires the raw lane line point cloud data to be processed, as well as the user's current interactive state, viewpoint position, and changes in viewpoint. The lane line point cloud data is essentially a massive collection of points containing three-dimensional spatial coordinates. Subsequently, instead of directly manipulating the raw data, the system first extracts deep features. Optionally, based on the three-dimensional coordinates of each point in the lane line point cloud data, it calculates its spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics. First, global statistical analysis determines the spatial distribution characteristics of the point cloud, including its overall boundary, centroid, and macroscopic density trends, providing a basis for the top-level design of spatial organization. Second, for each point or its small neighborhood, local density is calculated to numerically characterize the density of the region, serving as a primary indicator to distinguish high-value annotation areas (such as lanes) from background areas. Furthermore, through local surface fitting, the curvature changes of the point cloud are analyzed to accurately locate areas where the geometry has significantly changed (such as lane edges and shoulders), areas requiring extremely high annotation accuracy. At the same time, the system directly reads or normalizes the reflection intensity characteristics provided by the sensor.

[0018] S103: Dynamically adjust the slice granularity threshold based on the point cloud density, curvature variation, and reflection intensity characteristics corresponding to the lane line point cloud data.

[0019] In one specific embodiment, after point cloud feature extraction, the system enters a crucial strategy formulation stage. This involves dynamically adjusting the slice granularity threshold based on the point cloud density, curvature variation, and reflection intensity characteristics corresponding to the lane line point cloud data. Specifically, the system performs a comprehensive analysis of the scene space to identify feature combinations in different regions: for regions with high point cloud density, dramatic curvature variations (representing geometric edges and details), and significant reflection intensity characteristics (representing specific materials), the system determines them as critical regions requiring high-precision characterization (such as lane lines and complex intersections), and dynamically assigns them a smaller slice granularity threshold to ensure that subsequent generated slice units are sufficiently fine to retain all critical details. Conversely, for regions with sparse point clouds, gentle curvature, and ordinary reflection intensity, the system determines them as non-critical or background regions, assigning them a larger slice granularity threshold to generate coarser-grained slices.

[0020] Figure 2 This is a schematic diagram of a process for dynamically adjusting a slice granularity threshold according to an exemplary embodiment. In an optional embodiment, such as Figure 2 As shown, the above-mentioned dynamic adjustment of the slice granularity threshold based on the point cloud density, curvature variation, and reflection intensity characteristics corresponding to the lane line point cloud data includes: S201: Obtain a preset granularity threshold; the preset granularity threshold includes a first slice granularity threshold and a second slice granularity threshold, wherein the first slice granularity threshold is less than the second slice granularity threshold; S203: Perform lane line detection on the lane line point cloud data to determine the lane line region and non-lane line region; and set the granularity threshold of the lane line region to the first slice granularity threshold, and set the granularity threshold of the non-lane line region to the second slice granularity threshold.

[0021] In one specific embodiment, the system obtains a preset granularity threshold benchmark, which includes a first slice granularity threshold with a smaller value for high-precision processing and a second slice granularity threshold with a larger value for high-efficiency processing, both satisfying the core rule that the first threshold is less than the second threshold. Then, the system executes a lane line detection task, which comprehensively utilizes the aforementioned multi-dimensional features such as point cloud density, curvature variation, and reflection intensity characteristics. Through rule-based judgment or model inference, the entire point cloud scene is divided into clearly defined lane line regions and non-lane line regions. Finally, the system performs direct mapping according to preset rules: the slice granularity threshold of the lane line region is uniformly set to the first slice granularity threshold to ensure that this key region is cut into sufficiently small units, preserving all geometric and semantic details of the lane lines; simultaneously, the slice granularity threshold of the non-lane line region is uniformly set to the second slice granularity threshold, thereby generating larger data blocks in this region and effectively reducing the overall data volume.

[0022] In the above embodiments, the combination of preset differential thresholds and lane line region detection achieves both accuracy and efficiency in the slicing strategy. The system applies a fine-grained first threshold to critical lane line regions, ensuring the detail integrity and annotation accuracy of the data in these regions; simultaneously, a coarse-grained second threshold is applied to non-critical regions, significantly reducing redundant data. This semantic-based static rule mapping achieves an optimal balance between accuracy and efficiency in data organization with low computational overhead, laying a lightweight and reliable data foundation for subsequent smooth interaction.

[0023] S105: Based on the spatial distribution characteristics and slice granularity threshold of lane line point cloud data, a spatial segmentation algorithm is used to perform spatial block slicing processing on lane line point cloud data to generate a point cloud slice set.

[0024] In a specific embodiment, based on the spatial distribution characteristics of the lane line point cloud obtained in the aforementioned steps and the dynamically adjusted slice granularity threshold, the system initiates the core data reorganization process, that is, it uses a spatial segmentation algorithm to perform spatial block slicing processing on the lane line point cloud data, and finally generates a structured point cloud slice set.

[0025] Figure 3This is a schematic diagram of a process for generating a point cloud slice set according to an exemplary embodiment. In an optional embodiment, such as Figure 3 As shown, based on the spatial distribution characteristics and slice granularity threshold of lane line point cloud data, the above-mentioned spatial segmentation algorithm is used to perform spatial block slicing processing on the lane line point cloud data, generating a point cloud slice set including: S301: Calculate the optimal segmentation depth and determine the target segmentation depth based on spatial distribution characteristics; S303: Based on the target segmentation depth and the target slice granularity threshold, recursively perform octree partitioning on the three-dimensional space to generate a point cloud slice set.

[0026] In a specific embodiment, the target segmentation depth can be the maximum recursive level that the segmentation algorithm should ideally achieve (without considering local granularity threshold differences). Optionally, the target segmentation depth is a global, preventative control parameter. Specifically, based on the spatial distribution characteristics of the lane line point cloud data, especially its overall data scale and density distribution, the system calculates an optimal segmentation depth and determines it as the target segmentation depth for octree recursive partitioning. Subsequently, the system initiates recursive partitioning: starting from the largest cube representing the entire point cloud space, it divides it into eight layers. At each partitioning decision point, the algorithm checks two stopping conditions in parallel: first, whether the current recursive level has reached the target segmentation depth; second, whether the physical size of the current sub-cube to be partitioned is less than or equal to the slice granularity threshold corresponding to the location of the sub-cube's center point. As long as either condition is met, the partitioning terminates immediately, and the current sub-cube and all point clouds within it are solidified into a final slice unit.

[0027] In the above embodiments, efficient and controllable spatial segmentation is achieved by combining global depth control and local granularity constraints. The target segmentation depth is determined based on the overall data scale to prevent excessive or insufficient recursion, ensuring the overall management efficiency of the slice set. Simultaneously, a dynamically adjusted granularity threshold is used as a local stopping condition to ensure that sufficiently fine slices are generated in critical regions to preserve details, while coarser slices are generated in non-critical regions to reduce data volume. The synergy of these two approaches enables the octree partitioning process to achieve both macroscopic efficiency and local accuracy, providing a structurally optimal point cloud slice set for constructing a high-performance spatiotemporal database.

[0028] Figure 4 This is a schematic diagram of a process for establishing a point cloud slice database according to an exemplary embodiment. In an optional embodiment, after generating the point cloud slice set, as... Figure 4 As shown, the above method also includes: S401: Classify the point cloud slice set to obtain lane area slices and non-lane area slices, and compress the lane area slices using the first compression ratio to obtain the first compressed slice set. S403: Based on the point cloud density of the non-lane region and the distance from each slice in the non-lane region to the center of the lane, the importance of the slices in the non-lane region is judged to determine the compression level; and the non-lane region is compressed based on the second compression ratio corresponding to the compression level to obtain a second compressed slice set; the second compression ratio is greater than the first compression ratio.

[0029] S405: Establish a point cloud slice database based on the first compressed slice set and the second compressed slice set.

[0030] In one specific embodiment, the system, based on existing semantic classification results, explicitly divides the point cloud slice set into lane region slices and non-lane region slices. For lane region slices, which are crucial to the core quality of annotation, the system uses a low, conservative first compression ratio for uniform compression to generate a first compressed slice set. This compression strategy aims to preserve key attributes such as the geometric details and reflection intensity of lane lines to the maximum extent, ensuring the accuracy of subsequent annotation operations. For non-lane region slices, the system performs a more refined importance judgment, comprehensively evaluating the point cloud density of each slice (the higher the density, the richer the environmental information it contains) and its spatial distance to the lane centerline (the closer the distance, the stronger the potential correlation with the driving scenario), and calculates a dynamic importance score for it. Based on this score, the system dynamically assigns a compression level to each non-lane area slice and uses a higher second compression ratio corresponding to that level for compression, thereby generating a second compressed slice set. The second compression ratio is greater than the first compression ratio, which means that the non-lane area data will undergo a higher compression rate to reduce its volume. Furthermore, the system integrates the first compressed slice set with high compression fidelity with the intelligently compressed second compressed slice set, and uses them together as input to construct the final point cloud slice database.

[0031] In the above embodiments, a differentiated compression strategy significantly optimizes storage and transmission efficiency while ensuring data quality. A low compression ratio is applied to critical lane areas to preserve geometric details to the greatest extent possible and ensure annotation accuracy; for non-lane areas, a higher compression ratio is dynamically allocated based on their point cloud density and distance from the lane center, intelligently compressing secondary information. This mechanism significantly reduces the overall data volume while having almost no impact on the core task performance, effectively alleviating the storage pressure and network transmission load on the cloud platform, and improving the economy of system resource utilization and the speed of data distribution.

[0032] S107: Based on the point cloud slice set, establish a point cloud slice database containing a spatiotemporal four-dimensional index; and dynamically load visible point cloud slices within the visible area from the point cloud slice database based on the user's view position and changes in the user's view.

[0033] In one specific embodiment, based on the generated, structured set of point cloud slices, the system enters the data organization and service-oriented phase. First, a point cloud slice database containing a spatiotemporal four-dimensional index is established. Each slice in the database possesses precise spatiotemporal coordinates, and the database supports efficient and complex queries such as "querying all slices within a spatial range near a certain point in time." Subsequently, the system switches to real-time service mode, dynamically loading visible point cloud slices within the visible area from the point cloud slice database based on the user's viewpoint position and changes in viewpoint.

[0034] Figure 5 This is a schematic diagram illustrating a process for dynamically loading visible point cloud slices within a visible region from a point cloud slice database, according to an exemplary embodiment. In an optional embodiment, such as Figure 5 As shown, the above-mentioned dynamic loading of visible point cloud slices within the visible area from the point cloud slice database based on the user's viewpoint position and changes in the user's viewpoint includes: S501: Based on the user's viewpoint position, use the view frustum clipping technique to determine the current set of slices to be loaded; S503: Based on the speed and direction of changes in the user's viewpoint, establish a viewpoint movement prediction model, and based on the viewpoint movement prediction model, preload the visible point cloud slices corresponding to the upcoming field of view from the current set of slices to be loaded.

[0035] In one specific embodiment, an instant loading system is initiated, utilizing frustum clipping techniques from computer graphics based on the user's viewpoint position. This system transforms the user's viewpoint into a concrete 3D frustum space model, accurately representing the visible range of the human eye or camera. Subsequently, the system rapidly performs geometric intersection tests between the bounding boxes of each slice in the point cloud slice database and the current frustum using efficient spatiotemporal indexing. All slices determined to intersect are immediately marked and added to the current set of slices to be loaded, triggering high-priority network requests and decoding processes to ensure that the scene within the user's current field of view is rendered completely and in real-time. Simultaneously, a predictive preloading system operates in parallel. This system analyzes the speed and direction of changes in the user's viewpoint in real time to build a lightweight viewpoint movement prediction model. This model, by analyzing short-term motion trends, calculates the spatial region to which the user's field of view is most likely to expand in the next very short time (e.g., the next few frames). Based on this prediction, the system does not wait passively, but actively filters out the subset of slices located in the predicted future visible area from the already identified "current set to be loaded", and puts these slices into a background preloading queue for advance scheduling.

[0036] In the above embodiments, a dual mechanism of "instant loading + intelligent preloading" enables smooth, latency-free interaction with large-scale point cloud data. View frustum culling accurately determines the currently needed data, ensuring complete image presentation; the prediction model based on changes in the user's perspective pre-schedules slices about to enter the field of view, effectively eliminating lag caused by network transmission and decoding. The combination of these two mechanisms provides users with an instant response experience similar to native applications when panning and rotating the viewpoint, significantly improving the overall efficiency and immersiveness of annotation operations.

[0037] Figure 6 This is a schematic flowchart illustrating the preloading of a visible point cloud slice corresponding to an upcoming field of view, according to an exemplary embodiment. In an optional embodiment, such as... Figure 6 As shown, the above-mentioned viewpoint movement prediction model preloads the visible point cloud slices corresponding to the upcoming field of view region from the current set of slices to be loaded, including: S601: Get the slice center of the current set of slices to be loaded; S603: Based on the slice center and the user's view position, determine the loading priority of each slice in the current set of slices to be loaded; and preload based on the loading priority.

[0038] In one specific embodiment, after determining the set of candidate slices about to enter the field of view, the system does not perform unordered loading, but rather implements intelligent scheduling by calculating loading priorities. First, the system obtains the center coordinates of each slice in the current set of slices to be loaded. This center point represents the average position of the slice in three-dimensional space. Then, based on the slice center and the real-time updated user viewpoint position, the system assigns a quantified loading priority score to each slice using a priority calculation model. The core of this model lies in comprehensively evaluating two key spatial relationships: one is the Euclidean distance from the slice to the user's viewpoint. The closer the distance, the higher the probability of it being seen and the stronger its urgency, thus the greater its priority contribution. The other is the orientation of the slice relative to the user's current line of sight. This is measured by calculating the cosine of the angle between the line of sight vector and the vector from the viewpoint to the slice center. The smaller the angle, the closer the slice is to the user's visual focus movement path, and the higher its priority contribution. The system combines these two factors (often in the form of a weighted sum) to calculate the final priority of each slice. Finally, the preloading execution engine sorts all candidate slices in descending order according to this priority score and initiates asynchronous loading requests strictly in this order.

[0039] In the above embodiments, an optimal allocation of preloaded resources is achieved by introducing a priority scheduling mechanism based on spatial relationships. The system calculates priorities based on the distance and directional angle between the slice center and the user's viewpoint, ensuring that slices that are closer and closer to the line of sight are loaded first. This intelligent sorting allows limited network bandwidth and front-end cache resources to be concentrated on serving the data most likely to be observed immediately, thereby maximizing the hit rate and timeliness of preloading in complex dynamic interactions, further consolidating a smooth and seamless user experience.

[0040] S109: Based on the spatiotemporal four-dimensional index in the point cloud slice database, establish cross-time frame annotation associations using the temporal consistency module, and annotate the visible point cloud slices based on the annotation associations.

[0041] In one specific embodiment, based on a pre-built point cloud tile database containing a spatiotemporal four-dimensional index, the system achieves a crucial leap from discrete data processing to intelligent and coherent annotation through its temporal consistency module. The core of this process lies in establishing annotation associations across time frames. When a user completes lane line annotation on a point cloud tile in the currently visible area within a specific time frame, the temporal consistency module is immediately activated. It deeply utilizes the spatiotemporal four-dimensional index in the point cloud tile database to perform an efficient spatiotemporal joint query: quickly retrieving all relevant tiles that spatially overlap with the current annotation area but are in adjacent frame sequences in the temporal dimension. Based on the principle of spatiotemporal continuity, the system determines that these tiles represent different observation states of the same physical lane line over continuous time, thereby automatically constructing an "annotation logical group" connecting multiple time points and the same spatial entity. After establishing this association, the system enters the intelligent annotation execution phase. When the user browses to a new time point or a viewpoint shift causes a new point cloud tile to enter the visible area, the system submits the tile and its spatiotemporal coordinates to the temporal consistency module for matching in real time. If a match is successful, meaning the slice is found to belong to an established annotation association group, the system can automatically or with minimal user confirmation, intelligently propagate and render the completed annotation results (including geometric shape and attribute information) within the association group onto the new slice.

[0042] Figure 7 This is a schematic diagram illustrating a process for establishing annotation associations across time frames, according to an exemplary embodiment. In an optional embodiment, such as... Figure 7 As shown, the above-mentioned spatiotemporal four-dimensional index based on the point cloud slice database, using the temporal consistency module to establish cross-time frame annotation associations includes: S701: Construct a four-dimensional spatiotemporal index tree with spatial coordinates and timestamps as dimensions; S703: Based on a four-dimensional spatiotemporal index tree, the annotations completed by the user on the target slice in the current time frame are automatically propagated to the corresponding slices of the preset frame values ​​before and after, establishing the annotation association between the target slice in the current time frame and the corresponding slice of the preset frame value; the preset frame value is a real-time frame value that is dynamically adjusted according to the vehicle speed and scene complexity.

[0043] In one specific embodiment, the system constructs a four-dimensional spatiotemporal index tree based on the metadata of each slice in the point cloud slice database, with spatial coordinates (X, Y, Z) and timestamps (T) as joint dimensions. Optionally, the metadata of the point cloud slices is automatically constructed during the process of generating slice sets through spatial segmentation. Specifically, when determining each final slice through octree partitioning, the system synchronously calculates and encapsulates multi-dimensional information, including the spatial range and center coordinates directly extracted from the cube boundary, statistical analysis of the number, density, and reflection intensity characteristics of point clouds within the slice, association with the timestamps carried by the original point clouds to determine the time range, recording the hierarchy, parent nodes, and adjacent slice relationships in the tree structure, and adding post-processing identifiers such as region classification and compression parameters. This four-dimensional spatiotemporal index tree organizes massive slice data in an orderly manner in a spatiotemporally continuous domain, enabling efficient composite queries that "quickly return all matching slices given a spatial range and time interval." When a user completes the annotation operation on a target slice within a certain current time frame, the temporal consistency module immediately starts automatic propagation based on this index tree. The core of this system is the calculation of a key dynamic parameter: the pre-set frame value (ΔT). Optionally, this pre-set frame value is not fixed but dynamically generated based on real-time scene information. Specifically, the system comprehensively considers the vehicle's speed (the faster the speed, the larger the ΔT value, to cover a wider time span to match the spatial displacement caused by motion) and the complexity of the current scene (in complex scenarios such as overpasses and merging areas, the ΔT value may be more conservative to reduce false propagation), calculating the optimal propagation time window radius in real time. Then, using the spatial bounding box of the target slice as the center and [current timestamp - ΔT, current timestamp + ΔT] as the time window, the system initiates a precise range query to the four-dimensional spatiotemporal index tree, instantly retrieving all corresponding slices within this time spatiotemporal range. The system then automatically and intelligently maps the complete annotations (geometric shapes and attributes) completed by the user on the current target slice to each corresponding slice found in the query, and establishes explicit annotation association records for these slices in the database.

[0044] In the above embodiments, by constructing a four-dimensional spatiotemporal index and a dynamic propagation mechanism, a leap in automation and intelligence of annotation work is achieved. The index tree supports millisecond-level cross-spatiotemporal slice retrieval, providing an efficient foundation for association establishment; while the preset frame values ​​dynamically adjusted according to vehicle speed and scene complexity ensure that annotation propagation maintains accuracy even in motion and complex environments. This mechanism transforms annotation from repetitive frame-by-frame labor into an intelligent process of one-time definition and automatic synchronization, greatly improving efficiency and fundamentally guaranteeing the spatial consistency and continuity of annotation results over time.

[0045] Figure 8 This is a schematic diagram of a process for detecting and resolving annotation conflicts according to an exemplary embodiment. In an optional embodiment, such as Figure 8 As shown, the above method also includes: S801: In the process of establishing annotation associations, bidirectional difference detection is used to identify annotation conflicts and obtain the identification results; S803: When the identification result indicates that a conflict has been detected, weighted fusion is performed based on the confidence level of the historical annotation results; Alternatively, generate conflict resolution prompts for the user; Alternatively, actively select the labeling results with higher confidence from the conflict labeling results.

[0046] In a specific embodiment, during the establishment and propagation of annotation associations, bidirectional difference detection is performed. Optionally, bidirectional difference detection examines consistency from two directions: first, forward detection, which evaluates the matching degree between the newly propagated annotation results and the target frame point cloud data, as well as its compatibility with existing annotations in that frame; second, backward detection, which verifies the logical coherence between the newly generated annotations and existing associated annotations at earlier time points. The detection content covers spatial location deviations, geometric shape distortions, and semantic attribute contradictions, etc. Potential annotation conflicts are identified through quantitative calculations, and clear identification results are generated. If the identification results indicate a conflict, the system initiates a multi-level resolution strategy. First, it attempts the most efficient automated fusion, which involves weighting the historical confidence levels of the annotation results from each conflicting party (this confidence level integrates factors such as annotation source, operational granularity, and historical acceptance) to automatically generate a new annotation version with higher consistency that combines the advantages of all parties. If the conditions for automated fusion are insufficient or the result is highly uncertain, the system downgrades to a human-machine collaborative resolution mode, proactively generating clear conflict resolution prompts for the user, presenting conflict details in a visual manner, and guiding the user to review and adjudicate. As a supplement or under explicit rules, the system can also execute a conservative best-preference strategy, directly adopting the party with the highest confidence level among the conflicting annotations as the final result.

[0047] In the above embodiments, the introduction of an intelligent conflict handling mechanism ensures the robustness of automated annotation association and the high reliability of the final data. Bidirectional difference detection proactively identifies potential spatial or semantic inconsistencies that may arise during cross-frame propagation. Subsequently, the system provides multi-level resolution strategies: prioritizing silent repair using confidence-based automated weighted fusion; if this is not feasible, generating prompts to guide user decision-making, or conservatively adopting the higher-confidence result. This process effectively prevents error propagation while maintaining the advantages of automation efficiency, ensuring the inherent consistency and quality of the labeled dataset.

[0048] This invention provides a data annotation method that, in response to annotation task requests, synchronously acquires lane line point cloud data containing three-dimensional spatial coordinates, as well as user viewpoint position and change information. Based on the point cloud coordinates, it extracts multi-dimensional features such as spatial distribution, local density, curvature changes, and reflection intensity. Subsequently, the system dynamically adjusts the slice granularity threshold according to these features, achieving an adaptive segmentation strategy of fine-grained segmentation for lane areas and coarse-grained segmentation for non-lane areas. Based on this, combining spatial distribution features and granularity thresholds, a spatial segmentation algorithm is used to intelligently divide the point cloud into blocks, generating a structured point cloud slice set, and constructing a point cloud slice database with a spatiotemporal four-dimensional index. During the interaction phase, point cloud slices within the visible and predicted areas are dynamically loaded from the database based on the real-time user viewpoint position and changes, achieving a smooth and lag-free browsing experience. Finally, utilizing the database's spatiotemporal index and temporal consistency module, annotation associations across time frames are established, allowing annotations completed by the user in the current frame to automatically propagate to corresponding slices in adjacent frames, thus ensuring the continuity and consistency of annotation results in the temporal dimension while efficiently completing the annotation. This method systematically solves the problems of memory bottleneck, rendering lag, and annotation fragmentation in large-scale 4D point cloud processing, achieving a synergistic improvement in accuracy, efficiency, and system performance.

[0049] Figure 9 This is a schematic diagram of a data annotation apparatus according to an exemplary embodiment. The following describes an embodiment of a data annotation apparatus according to this application, specifically, as follows: Figure 9 As shown, the device includes: The acquisition module 901 is used to respond to the annotation task request and acquire lane line point cloud data, user view position, and user view changes; the lane line point cloud data includes the three-dimensional spatial coordinates corresponding to the point cloud data; and based on the three-dimensional spatial coordinates of the lane line point cloud data, determine the spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics of the lane line point cloud data. The adjustment module 903 is used to dynamically adjust the slice granularity threshold according to the point cloud density, curvature change and reflection intensity characteristics corresponding to the lane line point cloud data. The slicing module 905 is used to perform spatial segmentation slicing on the lane line point cloud data based on the spatial distribution characteristics of the lane line point cloud data and the slicing granularity threshold, and to generate a point cloud slice set by using a spatial segmentation algorithm. The loading module 907 is used to establish a point cloud slice database containing a spatiotemporal four-dimensional index based on the point cloud slice set; and dynamically load visible point cloud slices within the visible area from the point cloud slice database based on the user's view position and the changes in the user's view. The annotation module 909 is used to establish a cross-time frame annotation association based on the spatiotemporal four-dimensional index in the point cloud slice database, using the temporal consistency module, and to annotate the visible point cloud slice based on the annotation association.

[0050] In an optional embodiment, the slicing module 905 includes: A depth calculation unit is used to calculate the optimal segmentation depth and determine the target segmentation depth based on the spatial distribution characteristics. The partitioning unit is used to recursively partition the three-dimensional space into an octree based on the target segmentation depth and the target slice granularity threshold, thereby generating the point cloud slice set.

[0051] In an optional embodiment, the adjustment module 903 includes: A threshold acquisition unit is used to acquire a preset granularity threshold; the preset granularity threshold includes a first slice granularity threshold and a second slice granularity threshold, wherein the first slice granularity threshold is smaller than the second slice granularity threshold. The lane line detection unit is used to perform lane line detection on the lane line point cloud data, determine the lane line region and the non-lane line region, and set the granularity threshold of the lane line region to the first slice granularity threshold, and set the granularity threshold of the non-lane line region to the second slice granularity threshold.

[0052] In an optional embodiment, the loading module 907 includes: The set determination unit is used to determine the set of slices to be loaded based on the user's viewpoint position and using the view frustum clipping technique. The model loading unit is used to establish a viewpoint movement prediction model based on the speed and direction of the user's viewpoint change, and preload the visible point cloud slices corresponding to the upcoming field of view from the current set of slices to be loaded based on the viewpoint movement prediction model.

[0053] In an optional embodiment, the model loading unit described above includes: The center acquisition subunit is used to acquire the slice center of the currently loaded slice set; The preloading subunit is used to determine the loading priority of each slice in the current set of slices to be loaded based on the slice center and the user's view position; and to perform preloading based on the loading priority.

[0054] In an optional embodiment, the annotation module 909 includes: Tree building unit, used to construct a four-dimensional spatiotemporal index tree with spatial coordinates and timestamps as dimensions; The association unit is used to automatically propagate the annotations completed by the user on the target slice in the current time frame to the corresponding slices of the preset frame values ​​before and after, based on the four-dimensional spatiotemporal index tree, and establish the annotation association between the target slice in the current time frame and the corresponding slice of the preset frame value; the preset frame value is a real-time frame value that is dynamically adjusted according to the vehicle speed and scene complexity.

[0055] In an optional embodiment, the annotation module 909 further includes: The identification unit is used to identify label conflicts during the process of establishing label associations by using bidirectional difference detection, and to obtain the identification results. A conflict handling unit is used to perform weighted fusion based on the confidence level of historical annotation results when the identification result indicates that a conflict has been detected. Alternatively, generate conflict resolution prompts for the user; Alternatively, actively select the labeling results with higher confidence from the conflict labeling results.

[0056] In an optional embodiment, after generating the point cloud slice set, the above apparatus further includes: A classification unit is used to classify the point cloud slice set to obtain lane area slices and non-lane area slices, and to compress the lane area slices using a first compression ratio to obtain a first compressed slice set. The compression level determination unit is used to determine the importance of slices in the non-lane region based on the point cloud density of the non-lane region and the distance from each slice in the non-lane region to the center of the lane, and to compress the non-lane region based on the second compression ratio corresponding to the compression level to obtain a second compressed slice set; the second compression ratio is greater than the first compression ratio.

[0057] The database establishment unit is used to establish the point cloud slice database based on the first compressed slice set and the second compressed slice set.

[0058] In an exemplary embodiment, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the data annotation method as described in the embodiments of this disclosure.

[0059] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the data annotation method of the present disclosure embodiments.

[0060] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the data annotation methods provided in the various optional implementations described above.

[0061] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0062] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the foregoing claims.

[0063] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims. The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A data annotation method, characterized in that, The method includes: In response to the annotation task request, lane line point cloud data, user view position, and user view changes are acquired; the lane line point cloud data includes the three-dimensional spatial coordinates corresponding to the point cloud data; and based on the three-dimensional spatial coordinates of the lane line point cloud data, the spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics of the lane line point cloud data are determined. Based on the point cloud density, curvature variation and reflection intensity characteristics corresponding to the lane line point cloud data, the slice granularity threshold is dynamically adjusted. Based on the spatial distribution characteristics of the lane line point cloud data and the slice granularity threshold, a spatial segmentation algorithm is used to perform spatial block slicing on the lane line point cloud data to generate a point cloud slice set. Based on the point cloud slice set, a point cloud slice database containing a spatiotemporal four-dimensional index is established; and based on the user's viewpoint position and the changes in the user's viewpoint, visible point cloud slices within the visible area are dynamically loaded from the point cloud slice database. Based on the spatiotemporal four-dimensional index in the point cloud slice database, a cross-time frame annotation association is established using the temporal consistency module, and the visible point cloud slice is annotated based on the annotation association.

2. The method according to claim 1, characterized in that, Based on the spatial distribution characteristics of the lane line point cloud data and the slice granularity threshold, a spatial segmentation algorithm is used to perform spatial block slicing processing on the lane line point cloud data to generate a point cloud slice set, including: Based on the aforementioned spatial distribution characteristics, the optimal segmentation depth is calculated, and the target segmentation depth is determined. Based on the target segmentation depth and the target slice granularity threshold, the three-dimensional space is recursively divided into octrees to generate the point cloud slice set.

3. The method according to claim 1, characterized in that, The step of dynamically adjusting the slice granularity threshold based on the point cloud density, curvature variation, and reflection intensity characteristics corresponding to the lane line point cloud data includes: Obtain a preset granularity threshold; the preset granularity threshold includes a first slice granularity threshold and a second slice granularity threshold, wherein the first slice granularity threshold is smaller than the second slice granularity threshold; Lane line detection is performed on the lane line point cloud data to determine lane line regions and non-lane line regions; the granularity threshold of the lane line regions is set as the first slice granularity threshold, and the granularity threshold of the non-lane line regions is set as the second slice granularity threshold.

4. The method according to claim 1, characterized in that, The step of dynamically loading visible point cloud slices within the visible area from the point cloud slice database based on the user's viewpoint position and changes in the user's viewpoint includes: Based on the user's viewpoint position, the set of slices to be loaded is determined using the view frustum clipping technique; Based on the speed and direction of the user's perspective change, a perspective movement prediction model is established, and based on the perspective movement prediction model, the visible point cloud slices corresponding to the upcoming field of view area are preloaded from the current set of slices to be loaded.

5. The method according to claim 4, characterized in that, The preloading of visible point cloud slices corresponding to the upcoming field of view region from the currently loaded slice set based on the viewpoint movement prediction model includes: Get the slice center of the currently loaded slice set; Based on the slice center and the user's viewpoint position, the loading priority of each slice in the current set of slices to be loaded is determined; and preloading is performed based on the loading priority.

6. The method according to claim 1, characterized in that, The step of establishing cross-time frame annotation associations based on the spatiotemporal four-dimensional index in the point cloud slice database using the temporal consistency module includes: Construct a four-dimensional spatiotemporal index tree with spatial coordinates and timestamps as dimensions; Based on the four-dimensional spatiotemporal index tree, the annotations completed by the user on the target slice in the current time frame are automatically propagated to the corresponding slices of the preset frame values ​​before and after, establishing the annotation association between the target slice in the current time frame and the corresponding slice of the preset frame value; the preset frame value is a real-time frame value that is dynamically adjusted according to the vehicle speed and scene complexity.

7. The method according to claim 6, characterized in that, The method further includes: In the process of establishing annotation associations, bidirectional difference detection is used to identify annotation conflicts and obtain the identification results; If the identification result indicates that a conflict has been detected, a weighted fusion is performed based on the confidence level of the historical annotation results; Alternatively, generate conflict resolution prompts for the user; Alternatively, actively select the labeling results with higher confidence from the conflict labeling results.

8. The method according to claim 1, characterized in that, After generating the point cloud slice set, the method further includes: The point cloud slice set is classified to obtain lane area slices and non-lane area slices, and the lane area slices are compressed using a first compression ratio to obtain a first compressed slice set. Based on the point cloud density of the non-lane region and the distance from each slice in the non-lane region to the center of the lane, the importance of the slices in the non-lane region is determined to establish a compression level; and the non-lane region is compressed based on the second compression ratio corresponding to the compression level to obtain a second compressed slice set; the second compression ratio is greater than the first compression ratio. The point cloud slice database is established based on the first compressed slice set and the second compressed slice set.

9. A data annotation device, characterized in that, The device includes: The acquisition module is used to respond to the annotation task request and acquire lane line point cloud data, user view position, and user view changes; the lane line point cloud data includes the three-dimensional spatial coordinates corresponding to the point cloud data; and based on the three-dimensional spatial coordinates of the lane line point cloud data, the spatial distribution characteristics, local density, curvature changes, and reflection intensity characteristics of the lane line point cloud data are determined. The adjustment module is used to dynamically adjust the slice granularity threshold based on the point cloud density, curvature change and reflection intensity characteristics corresponding to the lane line point cloud data. The slicing module is used to perform spatial segmentation slicing of the lane line point cloud data based on the spatial distribution characteristics of the lane line point cloud data and the slicing granularity threshold, and to generate a point cloud slice set by using a spatial segmentation algorithm. The loading module is used to establish a point cloud slice database containing a spatiotemporal four-dimensional index based on the point cloud slice set; and dynamically load visible point cloud slices within the visible area from the point cloud slice database based on the user's view position and the changes in the user's view. The annotation module is used to establish annotation associations across time frames based on the spatiotemporal four-dimensional index in the point cloud slice database, using the temporal consistency module, and to annotate the visible point cloud slices based on the annotation associations.

10. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the data annotation method as described in any one of claims 1 to 8.

Images