Cross-slice non-physical slicing and quality inspection method and system based on hybrid label

Abstract

The invention provides a cross-slice non-physical slicing and quality inspection method and system based on hybrid tags, and relates to the technical field of automatic driving data engineering, and the method comprises the steps: carrying out the time alignment and source standardization processing of event type and fragment type hybrid tags; based on the standardized tags, performing parameter-adaptive window slicing on the event-type tags, and inserting start-stop marks in physical boundaries of the fragment-type tags to generate cross-fragment non-physical scene indexes and hash signatures; based on the index and label data, multi-dimensional quality inspection scores including time continuity, course consistency, causal coverage and source consistency are calculated; performing weighted fusion and closed-loop decision proofreading based on confidence and time sequence precision on the standardized label data with unqualified consistency quality inspection; and finally outputting a complete data packet containing the index file, the quality inspection report and the digital signature. According to the invention, high-quality and high-efficiency production and traceable management of automatic driving scene data are realized.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving data engineering, and in particular to a method and system for cross-slice non-physical slicing and quality inspection based on hybrid labels. Background Technology

[0002] The development of autonomous driving technology relies on the efficient processing and quality control of massive amounts of onboard sensor data. Raw data needs to go through screening, slicing, annotation, and quality inspection before it can be transformed into effective data that can be used for model training and testing.

[0003] Currently, a typical method for filtering autonomous driving data includes: receiving and storing raw data; generating metadata through deep learning model inference; inputting the metadata into a logical filter for filtering, and outputting an index of metadata to be retained; generating a subset of data to be labeled based on this index, which is then manually labeled and reviewed before being stored in the database. This method aims to obtain high-quality labeled data at a lower cost, thereby accelerating model iteration and scenario generalization. However, existing technologies still have significant limitations in specific aspects of data slicing and quality assurance. Most systems use physical slicing methods with fixed durations or based on simple event triggers, making it difficult to fully retain the semantics of coherent driving scenarios across slice boundaries. Furthermore, there is a lack of unified management of information such as the confidence level and temporal accuracy of labels from multiple sources, including automatic algorithms and manual annotations, and the conflict handling process is inadequate. In addition, existing quality inspection methods are mostly limited to format and simple logical checks, lacking the ability to conduct deep semantic consistency assessments from dimensions such as temporal continuity and motion consistency.

[0004] Overall, existing technologies still have significant shortcomings in the core aspects of data processing. Fixed rule slicing can easily destroy the semantic integrity of the scene; multi-source label management is crude and lacks the ability to resolve conflicts; the quality inspection dimension is single and it is difficult to identify deep inconsistencies in data; and the scene data organization method is not conducive to efficient storage and accurate traceability. Summary of the Invention

[0005] This invention aims to at least solve the aforementioned technical problems existing in the prior art. To this end, the first aspect of this invention proposes a method for cross-slice non-physical slicing and quality inspection based on hybrid labels, the method comprising:

[0006] Acquire tag data, which includes event-type tags and fragment-type tags. Align the tags with a time reference and standardize their source information according to the tag type to obtain standardized tag data. The source information includes tag source, confidence level, and time series accuracy.

[0007] Based on the standardized label data, adaptive window slicing is performed on the event-type labels, and boundary marking processing is performed on the fragment-type labels at the physical slice boundaries to obtain a cross-slice non-physical scene index.

[0008] Based on the standardized label data and the cross-slice non-physical scene index, a consistency quality inspection is performed to determine the quality inspection score; the quality inspection score includes at least one of the following indicators: temporal continuity score, heading consistency score, causal coverage score, and source consistency score.

[0009] For standardized label data that fails the consistency quality inspection, the conflicting labels are weighted and fused and the decision is checked based on the confidence level and time series accuracy in the standardized label data to obtain the conflict checking result.

[0010] Based on the cross-slice non-physical scene index, the quality inspection score, and the conflict correction result, an index file, a quality inspection report, and an index signature are generated and output.

[0011] This method effectively overcomes the semantic fragmentation problem caused by traditional fixed-duration slicing by standardizing event-type and fragment-type labels and generating cross-slice non-physical scene indexes based on this standardization, ensuring the semantic integrity of long-cycle or cross-boundary driving scenarios. By introducing multi-dimensional quality inspection scores to conduct deep semantic consistency assessment of scene data, low-quality data can be identified and filtered more accurately. Simultaneously, the closed-loop verification process for multi-source label conflicts improves the rationality and reliability of data fusion decisions, and the final output index signature enhances the traceability and version management capabilities of the data product.

[0012] Optionally, the adaptive window slicing of the event-type label includes:

[0013] Determine the length of the event preceding window and the length of the event following window based on at least one of the parameters: vehicle speed, road curvature, or road slope.

[0014] When a limited field of view is detected, the length of the window preceding the event and the length of the window following the event are adjusted based on a preset magnification factor.

[0015] This method adaptively determines the event window based on dynamic parameters such as vehicle speed and road conditions, and makes compensation adjustments when the field of view is limited. This ensures that the segmented scene fragments can cover key causal chains and effectively remove a large number of irrelevant frames, significantly improving the effectiveness and storage efficiency of scene data.

[0016] Optionally, the boundary marking process for the fragment-type label at the physical slice boundary includes:

[0017] An end-of-segment marker is written at the end of the first physical slice, and a start-of-segment marker is written at the beginning of the second physical slice; the second physical slice is temporally continuous with the first physical slice.

[0018] Based on the segment end marker and the segment start marker, continuity on the timeline is ensured.

[0019] This method clearly defines the boundaries of fragmented scenes by inserting special start and end markers at the boundaries of physical slices, ensuring seamless connection of scenes across multiple physical storage units on the timeline and providing a reliable foundation for subsequent continuous playback or simulation reproduction.

[0020] Optionally, obtaining the cross-slice non-physical scene index includes:

[0021] A non-physical index is constructed based on a list of identifiers for physical slices and an array of offsets pointing to the original data frame intervals;

[0022] The index signature is obtained by performing a hash operation on the index information, which includes a list of slice identifiers, an offset array, a time interval range, a source information digest, and a quality inspection score.

[0023] This method constructs a non-physical index using an identifier list and an offset array, avoiding the physical copying and handling of the original data, thus greatly saving storage space and I / O overhead. By generating a unique signature through hash operations on the core index information, it provides crucial technical support for data asset integrity verification, copyright management, and traceability auditing.

[0024] Optionally, the process of performing consistency quality inspection and determining the quality inspection score includes:

[0025] The temporal continuity score is calculated based on the time offset between adjacent tags;

[0026] The heading consistency score is calculated based on the change in heading angle between adjacent frames;

[0027] The causal coverage score is calculated based on the ratio of the number of key causal frames to the total number of frames within the slice window;

[0028] The source consistency score is obtained by weighting the consistency of tags from different sources in terms of timestamps and spatial locations.

[0029] This method constructs a quality inspection system from multiple dimensions such as temporal continuity, consistency of motion trajectory, logical causal coverage, and multi-source consistency, realizing a comprehensive and quantitative evaluation of scene data quality from surface format to deep semantics, making up for the shortcomings of traditional quality inspection methods with their single dimension.

[0030] Optionally, the following steps are also included:

[0031] Based on the comparison result between the quality inspection score and the preset threshold, it is determined whether the quality inspection has passed;

[0032] If the quality inspection fails, adjust the slice window parameters or boundary markers and re-perform the quality inspection.

[0033] If the data still fails the re-inspection, the slice mode will be downgraded to a mode based on event points and fixed-duration buffers, or the corresponding data will be marked as requiring manual review.

[0034] This method establishes multi-level quality inspection thresholds and corresponding rollback mechanisms (such as parameter adjustment, downgrading slicing mode, and marking for review), forming a flexible and progressive quality inspection process. This maximizes data utilization while ensuring the baseline quality of output data, and automatically directs difficult cases to human intervention, improving the overall intelligence and robustness of the processing flow.

[0035] Optionally, for the standardized label data that fails the consistency quality inspection, the conflicting labels are weighted, fused, and corrected based on the confidence level and time series accuracy in the standardized label data, including:

[0036] Align the timestamps of multiple conflicting tags;

[0037] The aligned tag information is weighted and fused using the confidence level and the time series accuracy as weights.

[0038] Calculate the consistency score of the fused labels;

[0039] Decisions are made based on the consistency score, and the results of the decisions include retaining a single label, synthesizing a new label, or setting the label status to pending.

[0040] Conflict cases with a decision result of suspension or a consistency score below the second threshold will be added to the review queue.

[0041] This method provides a standardized closed-loop processing flow for multi-source label conflicts. Through time alignment, confidence weighting, fusion decision and review queue mechanism, it transforms the original conflict resolution process that relied on human experience into a quantifiable and traceable automatic or semi-automatic process, which significantly improves the efficiency, consistency and decision quality of conflict handling.

[0042] A second aspect of this invention proposes a cross-slice non-physical slicing and quality inspection system based on hybrid labels, the system comprising:

[0043] The tag acquisition and processing module is used to acquire tag data, which includes event-type tags and fragment-type tags. The tag data is aligned with the time base and standardized with the source information according to the tag type to obtain standardized tag data.

[0044] The slice index module is used to adaptively slice the event-type labels according to the standardized label data, and to perform boundary marking processing on the fragment-type labels at the physical slice boundaries to obtain a cross-slice non-physical scene index.

[0045] The quality inspection scoring module is used to perform consistency quality inspection based on the standardized label data and the cross-slice non-physical scene index, and determine the quality inspection score.

[0046] The conflict resolution module is used to perform weighted fusion and decision-making verification on standardized label data that fails the consistency quality inspection, based on the confidence level and time series accuracy in the standardized label data, to obtain the conflict resolution result.

[0047] The output generation module is used to generate and output an index file, a quality inspection report, and an index signature based on the cross-slice non-physical scene index, the quality inspection score, and the conflict correction result.

[0048] Optionally, the slice index module is specifically used for:

[0049] Determine the length of the event preceding window and the length of the event following window based on at least one of the parameters: vehicle speed, road curvature, or road slope.

[0050] When a limited field of view is detected, the length of the window preceding the event and the length of the window following the event are adjusted based on a preset magnification factor.

[0051] Optionally, the quality inspection scoring module is specifically used to determine at least one of the following indicators: time continuity score, course consistency score, causal coverage score, and source consistency score. Attached Figure Description

[0052] Figure 1 A flowchart illustrating a cross-slice non-physical slicing and quality inspection method based on hybrid labels provided in an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of a tag data model for standardized tag data provided in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the process for weighted fusion and decision correction of conflicting labels provided in an embodiment of the present invention;

[0055] Figure 4 This is a schematic diagram of the structure of an index file provided in an embodiment of the present invention. Detailed Implementation

[0056] 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.

[0057] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, the use of "based on" or "according to" implies openness and inclusiveness, because processes, steps, calculations, or other actions "based on" or "according to" one or more of the stated conditions or values ​​may in practice be based on additional conditions or beyond the stated values.

[0058] This invention provides a method for cross-slice non-physical slicing and quality inspection based on hybrid labels, such as... Figure 1 As shown, the method may include the following steps:

[0059] Step 101: Obtain tag data, which includes event-type tags and fragment-type tags. Based on the tag type, the tag data is aligned with the time base and standardized with the source information to obtain standardized tag data.

[0060] The source information includes the label source, confidence level, and time series accuracy.

[0061] Specifically, the raw label data mainly includes two categories: event-type labels generated by automatic detection operators (such as trigger events like vehicle entry or sudden braking), and fragment-type labels generated by manual on-vehicle annotation (such as weather conditions, road types, and other states that last for a period of time). Standardized label data is attached. Figure 2The label data model shown contains the following core fields for each label record: label_type (label category, value is event or section), time information (time t for event type, start and end time [t_start, t_end] for section type), label_source (label source, value is auto or human), confidence (confidence, normalized to the [0,1] interval), time_quality (time series accuracy, divided into high, mid, and low levels according to the deviation from the system clock), operator identifier (operator_id or anonymized human_id), review_status (review status, such as pending, approved, rejected), clip_id of the physical slice to which it belongs, and other remarks information extra.

[0062] The standardization process specifically includes: adding the algorithm version and trigger threshold information to tags from automatic operators; and adding recording device, shift, and session information to tags from manual on-vehicle labeling. Regardless of the source, the confidence level is uniformly mapped to the [0,1] scale, and the timing accuracy level is assessed based on the deviation between the timestamp and the reference time (e.g., deviation <10ms is high, 10-50ms is mid, and >50ms is low). Simultaneously, consistency constraints are enforced, such as ensuring that the time logic of the same label_id is strictly prohibited from being reversed; for potential duplicate writes, a merging strategy of "version number + timestamp priority" is adopted, ultimately outputting standardized tag data with a unified structure and complete information.

[0063] Step 102: Based on the standardized label data, perform adaptive window slicing on event-type labels and perform boundary marking on fragment-type labels at the physical slice boundaries to obtain cross-slice non-physical scene indexes.

[0064] In one possible implementation, adaptive window slicing is performed on event-type labels, including:

[0065] Determine the length of the event preceding window and the length of the event following window based on at least one of the parameters: vehicle speed, road curvature, or road slope.

[0066] When a limited field of view is detected, the length of the window preceding the event and the length of the window following the event are adjusted based on a preset magnification factor.

[0067] Specifically, for event-based tags, instead of using fixed-duration slices, adaptive window calculation is performed. The event preceding window length L... pre With the event follow-up window length L postThe calculation formula is derived from dynamic parameters and is as follows:

[0068]

[0069]

[0070] Where v is the vehicle speed, κ is the road curvature, and road_grade is the road slope; α0, α1, α2, β0, β1, β2 are preset weighting coefficients; the clip function ensures that the calculation result falls within the preset minimum value L. min With the maximum value L max Within the interval formed.

[0071] When a limited field of view is detected (such as rain, fog, nighttime, or occlusion by a curve), a magnification factor γ greater than 1 is introduced to extend the window length to ensure that key visual information is not missed.

[0072] In one possible implementation, the fragment-type tags are subjected to boundary marking processing at the physical slice boundaries, including:

[0073] Write a segment end marker at the end of the first physical slice and a segment start marker at the beginning of the second physical slice; the second physical slice is temporally continuous with the first physical slice.

[0074] Based on segment end markers and segment start markers, continuity on the timeline is ensured.

[0075] Specifically, for fragment tags, their semantics may span multiple physical clips divided into fixed-duration segments. To address the issue of lost semantics at boundaries, boundary marking is performed at the intersection of adjacent physical clips: an "end marker" is written at the end time of Clip N, and a "start marker" is written at the start time of Clip N+1. The timing relationship is precisely controlled to ensure that t_end(clip_n) = t_start(clip_{n+1}) - Δ. Here, Δ is a very small positive time difference used to characterize continuity.

[0076] In one possible implementation, obtaining the cross-slice non-physical scene index includes:

[0077] A non-physical index is constructed based on a list of identifiers for physical slices and an array of offsets pointing to the original data frame intervals;

[0078] The index signature is obtained by hashing the index information, which includes a list of slice identifiers, an array of offsets, a time interval range, a source information digest, and a quality inspection score.

[0079] Specifically, a cross-slice non-physical scene index is constructed. This index does not copy the original media stream or bus message data, but only organizes scenes through logical references. Its core components include: a list of physical slice IDs arranged in chronological order (clip_ids[]), and an array of offsets (offsets[]) pointing to the start and end positions of the corresponding valid frame sequence within each physical slice. Finally, a hash operation (using algorithms such as SHA-256 or XXH64) is performed on all the core metadata of this index (including clip_ids[], offsets[], total time interval t_range, standardized tag source digest source_digest, and quality inspection scores qc_scores to be generated in subsequent steps) to generate a unique index_hash as an immutable signature for the scene index, used for integrity verification and traceability.

[0080] Step 103: Based on standardized label data and cross-slice non-physical scene index, perform consistency quality inspection and determine the quality inspection score.

[0081] In one possible implementation, a consistency quality check is performed to determine a quality check score, including:

[0082] Calculate the temporal continuity score based on the temporal offset between adjacent labels;

[0083] Calculate the heading consistency score based on the change in heading angle between adjacent frames;

[0084] The causal coverage score is calculated based on the ratio of the number of key causal frames to the total number of frames within the slice window.

[0085] A source consistency score is obtained by weighting the consistency of tags from different sources in terms of timestamps and spatial locations.

[0086] Specifically, a series of quantifiable quality control scores are calculated and determined for the generated scene index to perform semantic quality assessment. The quality control scores include:

[0087] Time continuity fractions The evaluation assesses the consistency and smoothness of timestamps within the scene and across slice boundaries, calculated using the following formula: ,in To assess the deviation between the actual and expected time intervals between assessment points, This is the preset time decay constant.

[0088] Heading Consistency Score The rationality and smoothness of the vehicle's motion direction are evaluated using data from the vehicle's inertial measurement unit (IMU), electronic compass, or heading angle fitted from trajectory points. The calculation formula is as follows: ,in This represents the change in heading angle between adjacent calculation points. This is the preset heading attenuation constant.

[0089] Causal Coverage Score For event-driven scenarios, assess whether the sliced ​​window completely contains the key antecedent and direct consequence frames that triggered the event. The calculation formula is as follows: The logic for determining key causal frames needs to be predefined based on the specific event type.

[0090] Source consistency score When labels from multiple sources (such as automated algorithms and manual annotations) exist within the same time period, their consistency is evaluated. This score is calculated by weighting and fusing the spatial (e.g., bounding box overlap) and semantic (e.g., category) consistency of labels from different sources after temporal alignment. The calculation formula is as follows: .

[0091] A corresponding pass / fail threshold is set for each of the above scores. If any score is lower than its corresponding threshold, a multi-level fallback mechanism is triggered: First, the slice window parameters are adjusted or the segment boundary markers are fine-tuned; if the adjustment still fails to meet the requirements, the slice mode of the scene is downgraded to a simplified mode of "event point only + fixed duration buffer"; if the downgrade still fails to meet the quality requirements, the scene is finally marked as "unavailable" and a detailed review ticket is automatically generated and included in the manual review queue for processing.

[0092] In one possible implementation, the following steps are also included:

[0093] Based on the comparison between the quality inspection score and the preset threshold, it is determined whether the quality inspection has passed.

[0094] If the quality inspection fails, adjust the slice window parameters or boundary markers and re-perform the quality inspection.

[0095] If the data still fails the re-inspection, the slice mode will be downgraded to a mode based on event points and fixed-duration buffers, or the corresponding data will be marked as requiring manual review.

[0096] Specifically, based on the consistency score of the source. For example, the quality inspection process is as follows:

[0097] like If the value is ≥0.70 and there is no semantic conflict, it is judged as a "synthetic label" and the weight from which the participation originates is recorded.

[0098] If 0.55≤ <0.70, each source has a clear priority, the label data of the priority source is retained, and the data of the other sources is marked as "overwritten";

[0099] like If the value is ≤0.55, or there is a hard conflict in data category / direction, or the label data from high-weight sources is inconsistent and there is no priority rule, it is marked as "suspended" for manual review, and conflict evidence is reported.

[0100] Step 104: For standardized label data that fails the consistency quality inspection, perform weighted fusion and decision correction on conflicting labels based on the confidence level and time series accuracy in the standardized label data to obtain conflict correction results.

[0101] Standardized label data that fails the consistency quality check is specifically multi-source conflicting labels for the same semantic unit. The candidate set of conflicting labels is selected according to the following rules:

[0102] 1. Time Neighborhood: Event-type tags require that the events be of the same type and the time difference be within ±100ms (this threshold can be adjusted to 50-200ms depending on the device synchronization quality); fragment-type tags require that their time intervals have ≥30% overlap, or the minimum overlap duration is ≥0.5 seconds.

[0103] 2. Semantic equivalence: The label is assigned to a predefined semantic equivalence class (e.g., "left lane departure" and "LDW-L" are considered equivalent). If the categories are mutually exclusive (e.g., "left lane departure" and "right lane departure"), they are directly marked as "hard conflict".

[0104] 3. Spatial nearest neighbor (coarse screening): For bounding boxes or polygons, intersection-over-union ratio (IoU) ≥ 0.20 is considered as possibly pointing to the same object; for lines or points, the nearest distance ≤ 1.5 meters is considered as possibly pointing to the same object.

[0105] 4. Source difference: If the tag source IDs are different (such as from different automatic operators, human annotators or model versions), and simultaneously meet any combination of the above time, semantic and spatial conditions, they will be included in the same conflict candidate set for subsequent processing.

[0106] In one possible implementation, for standardized label data that fails the consistency quality inspection, conflicting labels are weighted, fused, and corrected based on the confidence level and temporal accuracy of the standardized label data. Figure 3 As shown, this step includes the following procedures:

[0107] Step 401: Align the timestamps of multiple conflicting tags.

[0108] Specifically, multiple conflicting tags that are multi-source conflicting tags of the same semantic unit are finely aligned based on their timestamps.

[0109] Step 402: Weighted fusion of aligned label information using confidence level and temporal accuracy as weights.

[0110] Based on the aligned labels, the confidence and time quality levels of each label are used as weighting factors to perform weighted calculations and fusion of information from conflicting labels.

[0111] Step 403: Calculate the consistency score of the fused labels.

[0112] The consistency score is assessed similarly to the aforementioned quality inspection score, but the judgment criteria are more stringent, focusing on the local score of the current conflict set; two thresholds are set for each category of consistency score: a composite threshold. and retention threshold .

[0113] Synthesis threshold (Default value 0.7 is recommended) is used to determine "multiple sources are basically consistent and can be fused".

[0114] Retention threshold (The default value is 0.55) This value is used for single-source validity assessment, specifically for "retaining only one source".

[0115] Step 404: Make a decision based on the consistency score. The decision results include retaining the single label, synthesizing a new label, or setting the label status to pending.

[0116] Specifically, for the score of each category in the consistency score, the decision-making method is as follows:

[0117] When the local consistency score among the majority of sources is greater than or equal to Furthermore, the semantics are not conflicting. Based on the timestamp, location / shape, and source weight, a weighted fusion is performed to synthesize a new label (Fusion). The confidence score is taken as the weighted mean or upper quantile, and a "synthesis source list" is generated to record the sources participating in the fusion.

[0118] When the consistency of a single source is sufficient (≥ Furthermore, other sources show low consistency with this source, or this source is significantly superior in terms of time series accuracy / confidence, or there is a preset priority level (e.g., "manually reviewed" is higher than "automatically detected", "high time series accuracy" is higher than "medium / low time series accuracy") and the score of this priority source is greater than or equal to If so, the decision is to retain the single label for that source (Keep One), and mark other source labels as "covered / discarded" and include them in the audit log.

[0119] When the discrepancy within the set is large (spatial / temporal differences exceed the threshold or there is hard class conflict) and the consistency scores from all major sources are less than [a certain value], [the situation is as follows]. If two high-weight sources are mutually exclusive and there is no clear priority rule, the decision is to "suspend".

[0120] Step 405: Add conflict cases with a decision result of suspension or a consistency score below the retention threshold to the review queue.

[0121] Specifically, all conflict cases with a decision result of "suspended" or a consistency score below another security threshold are automatically added to a dedicated review queue, awaiting final adjudication by human experts, with detailed conflict evidence (such as screenshots of differences, data curves, source metadata, etc.) attached to the case. The entire verification process, weight allocation, decision basis, and results are fully recorded and linked to the final scenario index signature.

[0122] Step 105: Based on the cross-chip non-physical scene index, quality inspection score and conflict correction results, generate and output the index file, quality inspection report and index signature.

[0123] Specifically, all the products from the aforementioned steps are standardized, packaged, and output, with the output index file structure as follows: Figure 4 As shown, it includes:

[0124] The index file (SceneIndex) is a structure containing clip_ids[] (a list of physical slice IDs), offsets[] (an array of frame offsets), t_range (the total time range), source_digest (a digest of the tag source), qc_scores (quality control scores for each item), and the crucial index_hash (the index signature).

[0125] Quality Control Report (QCReport): A detailed document that records the calculated scores (S_t, S_ψ, C_cause, S_src), the list of detected violations (violations[]), and the final quality control decision made by the system or a human.

[0126] Index signature: also known as index_hash, serves as a unique and tamper-proof digital fingerprint of the index data packet in this scenario.

[0127] Furthermore, the method supports batch processing task flow execution, orchestrating task nodes such as "batch slicing → consistency inspection → conflict resolution → index output" using a directed acyclic graph (DAG) to ensure that each node's processing is idempotent, retryable, and has timeout control capabilities. For exceptions encountered during execution (such as data parsing failure, timestamp reversal, or slice span exceeding limits), it can automatically roll back to the previous successful node and generate error reports and a "defect sample list," ensuring the robustness of the process.

[0128] In summary, this invention constructs a standardized hybrid label data model and performs adaptive cross-slice semantic slicing and multi-dimensional deep quality inspection based on it, ultimately forming a complete data asset package consisting of index files, quality inspection reports, and digital signatures. This method effectively solves the semantic fragmentation problem caused by traditional fixed slicing. By introducing a source confidence-weighted conflict closed-loop verification mechanism, it significantly improves the reliability and decision transparency of multi-source heterogeneous label fusion. Simultaneously, the design of non-physical indexes and hash signatures greatly reduces storage and I / O overhead while ensuring the semantic continuity of the scene, and establishes strong traceability. This achieves intelligent and standardized management of autonomous driving scene data throughout the entire process from production and quality inspection to management, providing a high-quality, efficient, and reliable data foundation for downstream model training, simulation testing, and performance evaluation.

[0129] This invention provides a system for cross-slice non-physical slicing and quality inspection based on hybrid labels, comprising:

[0130] The tag acquisition and processing module is used to acquire tag data, which includes event-type tags and fragment-type tags. The tag data is aligned with the time base and standardized with the source information according to the tag type to obtain standardized tag data.

[0131] The slice index module is used to adaptively slice the event-type labels according to the standardized label data, and to perform boundary marking processing on the fragment-type labels at the physical slice boundaries to obtain a cross-slice non-physical scene index.

[0132] The quality inspection scoring module is used to perform consistency quality inspection based on the standardized label data and the cross-slice non-physical scene index, and determine the quality inspection score.

[0133] The conflict resolution module is used to perform weighted fusion and decision-making verification on standardized label data that fails the consistency quality inspection, based on the confidence level and time series accuracy in the standardized label data, to obtain the conflict resolution result.

[0134] The output generation module is used to generate and output an index file, a quality inspection report, and an index signature based on the cross-slice non-physical scene index, the quality inspection score, and the conflict correction result.

[0135] In one possible implementation, the slice index module is specifically used for:

[0136] Determine the length of the event preceding window and the length of the event following window based on at least one of the parameters: vehicle speed, road curvature, or road slope.

[0137] When a limited field of view is detected, the length of the window preceding the event and the length of the window following the event are adjusted based on a preset magnification factor.

[0138] In one possible implementation, the quality inspection scoring module is specifically used to determine at least one of the following indicators: time continuity score, course consistency score, causal coverage score, and source consistency score.

[0139] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0140] In another embodiment of the present invention, an electronic device is also provided, which includes a processor and a memory. The memory stores at least one instruction or at least one program. The at least one instruction or at least one program is loaded and executed by the processor to implement the method for cross-slice non-physical slicing and quality inspection based on hybrid tags proposed in the embodiments of the present invention.

[0141] In another embodiment of the present invention, a computer-readable storage medium is also provided, which stores at least one instruction or at least one program, wherein the at least one instruction or at least one program is loaded and executed by a processor to implement the method for cross-slice non-physical slicing and quality inspection based on hybrid labels proposed in the embodiments of the present invention.

[0142] The foregoing primarily describes the solutions provided by the embodiments of the present invention from the perspective of the device. It is understood that, in order to achieve the above functions, the device includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the algorithmic steps of the examples described in the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0143] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for cross-slice non-physical slicing and quality inspection based on hybrid labels, characterized in that, The method includes: Acquire tag data, which includes event-type tags and fragment-type tags. Based on the tag type, align the tag data with a time reference and standardize its source information to obtain standardized tag data. The source information includes tag source, confidence level, and time series accuracy. Based on the standardized label data, adaptive window slicing is performed on the event-type labels, and boundary marking processing is performed on the fragment-type labels at the physical slice boundaries to obtain a cross-slice non-physical scene index. Based on the standardized label data and the cross-slice non-physical scene index, a consistency quality inspection is performed to determine the quality inspection score; the quality inspection score includes at least one of the following indicators: temporal continuity score, heading consistency score, causal coverage score, and source consistency score. For standardized label data that fails the consistency quality inspection, the conflicting labels are weighted and fused and the decision is checked based on the confidence level and time series accuracy in the standardized label data to obtain the conflict checking result. Based on the cross-slice non-physical scene index, the quality inspection score, and the conflict correction result, an index file, a quality inspection report, and an index signature are generated and output.

2. The method according to claim 1, characterized in that, The adaptive window slicing of the event-type labels includes: Determine the length of the event preceding window and the length of the event following window based on at least one of the parameters: vehicle speed, road curvature, or road slope. When a limited field of view is detected, the length of the window preceding the event and the length of the window following the event are adjusted based on a preset magnification factor.

3. The method according to claim 1, characterized in that, The boundary marking process for the fragment-type tags at the physical slice boundaries includes: An end-of-segment marker is written at the end of the first physical slice, and a start-of-segment marker is written at the beginning of the second physical slice; the second physical slice is temporally continuous with the first physical slice. Based on the segment end marker and the segment start marker, continuity on the timeline is ensured.

4. The method according to claim 1, characterized in that, The process of obtaining the cross-slice non-physical scene index includes: A non-physical index is constructed based on a list of identifiers for physical slices and an array of offsets pointing to the original data frame intervals; The index signature is obtained by performing a hash operation on the index information, which includes a list of slice identifiers, an offset array, a time interval range, a source information digest, and a quality inspection score.

5. The method according to claim 1, characterized in that, The process of conducting consistency quality inspection and determining the quality inspection score includes: The temporal continuity score is calculated based on the time offset between adjacent tags; The heading consistency score is calculated based on the change in heading angle between adjacent frames; The causal coverage score is calculated based on the ratio of the number of key causal frames to the total number of frames within the slice window; The source consistency score is obtained by weighting the consistency of tags from different sources in terms of timestamps and spatial locations.

6. The method according to claim 1 or 5, characterized in that, Also includes: Based on the comparison result between the quality inspection score and the preset threshold, it is determined whether the quality inspection has passed; If the quality inspection fails, adjust the slice window parameters or boundary markers and re-perform the quality inspection. If the data still fails the re-inspection, the slice mode will be downgraded to a mode based on event points and fixed-duration buffers, or the corresponding data will be marked as requiring manual review.

7. The method according to claim 1, characterized in that, For standardized label data that fails the consistency quality inspection, the conflicting labels are weighted, fused, and corrected based on the confidence level and time series accuracy of the standardized label data, including: Align the timestamps of multiple conflicting tags; The aligned label information is weighted and fused using the confidence level and the time series accuracy as weights. Calculate the consistency score of the fused labels; Decisions are made based on the consistency score, and the results of the decisions include retaining a single label, synthesizing a new label, or setting the label status to pending. Conflict cases with a decision result of suspension or a consistency score below the second threshold will be added to the review queue.

8. A cross-slice non-physical slicing and quality inspection system based on hybrid labels, characterized in that, The system includes: The tag acquisition and processing module is used to acquire tag data, which includes event-type tags and fragment-type tags. The tag data is aligned with the time base and standardized with the source information according to the tag type to obtain standardized tag data. The slice index module is used to adaptively slice the event-type labels according to the standardized label data, and to perform boundary marking processing on the fragment-type labels at the physical slice boundaries to obtain a cross-slice non-physical scene index. The quality inspection scoring module is used to perform consistency quality inspection based on the standardized label data and the cross-slice non-physical scene index, and determine the quality inspection score. The conflict resolution module is used to perform weighted fusion and decision-making verification on standardized label data that fails the consistency quality inspection, based on the confidence level and time series accuracy in the standardized label data, to obtain the conflict resolution result. The output generation module is used to generate and output an index file, a quality inspection report, and an index signature based on the cross-slice non-physical scene index, the quality inspection score, and the conflict correction result.

9. The system according to claim 8, characterized in that, The slice index module is specifically used for: Determine the length of the event preceding window and the length of the event following window based on at least one of the parameters: vehicle speed, road curvature, or road slope. When a limited field of view is detected, the length of the window preceding the event and the length of the window following the event are adjusted based on a preset magnification factor.

10. The system according to claim 8, characterized in that, The quality inspection scoring module is specifically used to determine at least one of the following indicators: time continuity score, course consistency score, causal coverage score, and source consistency score.

Images