Asynchronous visual target automatic tracking pre-labeling method and system for data labeling platform

Abstract

The present application relates to the technical field of data labeling, and more particularly to an asynchronous visual target automatic tracking pre-labeling method and system for a data labeling platform, which identifies typical labeling scenarios based on historical data, extracts task features, and generates efficiency influence labels; combines the labels with real-time requests, uses a hierarchical task processing mechanism to output pre-labeling results and asynchronous state records; builds a double-layer cache architecture including a resource cache layer and a result reuse layer to realize efficient reuse of media resources and inference results; analyzes the constraint boundaries of execution efficiency and resource occupation, fuses concurrent scheduling data and job priorities, and dynamically generates an asynchronous task dispatch window and an intelligent optimization strategy. This method effectively solves the system blocking problem under high concurrency, significantly reduces the repeated calculation overhead through hierarchical caching, and realizes adaptive scheduling of tasks according to multi-dimensional parameters, greatly improving the automation level, resource utilization rate, and overall processing efficiency of the labeling platform.

Description

Technical Field

[0001] This invention relates to the field of data annotation technology, and in particular to an asynchronous visual target automatic tracking and pre-annotation method and system for data annotation platforms. Background Technology

[0002] With the rapid development of artificial intelligence technology, especially the widespread application of computer vision in fields such as autonomous driving, intelligent monitoring, and robotics, the demand for high-quality visual data annotation has exploded. Among numerous annotation tasks, object detection and tracking annotation in video sequences has become one of the core bottlenecks of data annotation platforms due to its high time consumption and repetitiveness. To improve annotation efficiency, automatic pre-annotation technology has been widely introduced, which assists annotators in completing batch annotation tasks by calling pre-trained models to perform initial bounding box selection and cross-frame tracking of targets. However, existing pre-annotation systems mostly adopt synchronous processing mode, which is prone to causing competition for computing resources, response delays, and even service blockage when facing large-scale concurrent requests, seriously affecting platform availability and user experience.

[0003] Furthermore, traditional pre-annotation methods lack effective utilization of historical task knowledge. Target tracking tasks in the same or similar scenarios repeatedly execute the complete inference process, leading to wasted computational resources. Although some systems attempt to introduce caching mechanisms to reuse intermediate results, they typically only achieve single-level data temporary storage, failing to implement hierarchical management based on media resources, the reuse value of inference results, and access frequency. This makes it difficult to balance the contradiction between cache hit rate and storage overhead. Simultaneously, existing systems generally adopt first-come, first-served or static priority strategies for task scheduling, failing to fully consider the differences in task characteristics under different annotation scenarios (such as video resolution, frame rate, target density, and tracking difficulty) and their impact on overall annotation efficiency. This results in high-value tasks not receiving timely responses and low resource utilization.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide an asynchronous visual target automatic tracking and pre-annotation method and system for data annotation platforms. It aims to solve the technical problems of existing pre-annotation technologies, such as high concurrency blocking caused by synchronous processing, resource duplication and waste caused by the lack of hierarchical caching mechanism, and difficulty in balancing efficiency and state tracking in multiple scenarios due to static task scheduling strategies.

[0006] To achieve the above objectives, this invention provides an asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms, the method comprising: Acquire historical annotation task data from the data annotation platform, identify typical annotation scenarios, construct annotation scenario models, extract task feature parameters for target tracking pre-annotation under typical annotation scenarios, and generate corresponding annotation efficiency impact labels; The annotation efficiency affects the labels, and the real-time tracking pre-annotation request data initiated by the annotators is taken as input. The data is processed based on the hierarchical task processing mechanism, and the target tracking pre-annotation results and asynchronous task status records are output. A hierarchical caching architecture is set up to map the media resources and inference results that are pre-annotated and reused in historical tracking to the hierarchical caching architecture, and to build a two-layer caching structure of resource caching layer and result reuse layer. Based on the analysis of the hierarchical caching structure, the constraint boundary between the execution efficiency and resource consumption of annotation tasks is analyzed. Combining the concurrent task scheduling data of the annotation platform and the priority parameters of the annotators' jobs, an asynchronous task dispatch scheduling window set is output to generate a pre-annotation task intelligent optimization strategy.

[0007] Optionally, the process of acquiring historical annotation task data from the data annotation platform, identifying typical annotation scenarios, and constructing an annotation scenario model includes: By acquiring historical annotation task data generated by the data annotation platform during the historical annotation process, the historical annotation task data is constructed into an annotation task data matrix according to the annotation task type, resource size and submission time order; The annotation task data matrix is ​​divided into blocks based on the cache capacity hierarchical partitioning mechanism to obtain multiple historical annotation task units, and the media resource fragments and task attribute fragments corresponding to each historical annotation task unit are extracted. For each task attribute fragment, multidimensional task features are extracted. The multidimensional task features extracted from the same task attribute fragment are combined to form a task feature vector. A set of task feature vectors is constructed using the task feature vectors of all task attribute fragments. Cluster analysis is performed on the task feature vector set to identify representative typical annotation scenarios, including static image sequence annotation, VLA video annotation, and multi-camera cross-frame annotation. Based on the multi-dimensional task features extracted from each typical annotation scenario, an annotation scenario model is constructed.

[0008] Optionally, the process of extracting task feature parameters for pre-annotation of target tracking in typical annotation scenarios and generating corresponding labels affecting annotation efficiency includes: For the various typical annotation scenarios included in the annotation scenario model, based on the statistically representative multidimensional task features in the task feature vector set, task feature parameters are extracted as descriptive terms affecting annotation efficiency. Each type of annotation efficiency impact description is input into a preset cache benefit rule set. The cache benefit rule set is constructed based on the resource download time, storage consumption, and annotation acceleration effect samples of historical pre-annotated tasks through correlation analysis or logistic regression inference, and outputs the cache benefit intensity coefficient range and its resource reuse priority. By identifying the cache benefit intensity range for various typical annotation scenarios, resource reuse priority is bound to the corresponding typical annotation scenarios, and labels affecting annotation efficiency are generated.

[0009] Optionally, the process of taking the labeling efficiency-affecting labels and the real-time tracking pre-labeling request data initiated by the labelers as input, processing them based on a hierarchical task processing mechanism, and outputting the target tracking pre-labeling results and asynchronous task status records is as follows: The parameter verification and spatiotemporal registration processing are performed on the real-time tracking pre-annotation requests initiated by the annotation front-end. The real-time request data includes reference box coordinates, target tracking identifiers, camera identifiers, frame range, media resource information, and annotation task identifiers. Based on the hierarchical task processing mechanism, real-time request data is decoupled from business and algorithm layers to form hierarchical task processing segments; each task processing segment is associated and bound with the corresponding labeled efficiency impact tags to construct an asynchronous task sample set. Based on the asynchronous task sample set and each type of hierarchical processing link, multiple task processing segments are fused and analyzed in the order of business reception-state storage-algorithm call-result write-back. The flow characteristics of task parameters in each processing link are extracted, task state nodes are constructed, and they are associated with each other according to the order of processing links to form an asynchronous task state trajectory. The task state trajectory is then analyzed based on deep learning combined with multi-dimensional task features.

[0010] Optionally, the process of analyzing the task state trajectory based on deep learning and multi-dimensional task features includes: Identify abnormal task flow intervals, extract the gradient of task feature parameters, queuing time, and resource hit threshold points for each interval, and extract the recoverable waiting time within the corresponding scheduling stage, which is recorded as the task safety period. Record the gradient, queuing time, threshold points, and task safety period as abnormal task data, and record the task flow trajectory corresponding to the abnormal task interval as the asynchronous task status trajectory. Finally, output the pre-annotated frame-level results of target tracking and the complete asynchronous task status record.

[0011] Optionally, the process of setting up a hierarchical caching architecture, mapping historically tracked and pre-annotated reused media resources and inference results to the hierarchical caching architecture respectively, and constructing a two-layer caching structure of a resource caching layer and a result reuse layer includes: Media resources and pre-annotated inference results are associated with virtual cache nodes at different levels in the hierarchical caching architecture. Based on the reuse frequency and size of media resources and inference results, a multi-level mapping chain for pre-annotated cache is established, and the multi-level mapping chain is used as the basic cache space axis. Based on the resource URL characteristics, the media resources reused by each pre-annotation task are aggregated, and a public resource cache node corresponding to the resource identifier is constructed. The pre-annotation inference results that are repeatedly hit are bound to the corresponding result cache node. The access frequency and size of media resources, along with the model version and task type of the pre-labeled inference results, are recorded as attribute tags for cache nodes; the cache expiration time of media resources and pre-labeled inference results is used as the expiration tag for the current cache node and marked to the corresponding cache node. The cache nodes are analyzed based on the total cache capacity to identify the upper threshold of cache space occupancy and the lower threshold of cache eviction, and to form a cache capacity constraint boundary. Based on the basic cache space axis, cache nodes are arranged according to reuse priority and spatiotemporal order. Associative edges are constructed between cache nodes, and cache nodes are connected through these edges to build a two-layer cache structure. Cache attribute tags, expiration tags, and cache capacity constraint boundaries are bound to each corresponding cache node and associated edge, and mapped to the cache management system of the annotation platform to build a complete two-layer cache structure.

[0012] Optionally, the process of analyzing the constraint boundary between the execution efficiency and resource consumption of annotation tasks based on the hierarchical caching structure, combining the concurrent task scheduling data of the annotation platform and the annotationer's job priority parameters, outputting an asynchronous task dispatch scheduling window set, and generating a pre-annotation task intelligent optimization strategy includes: Obtain the set of cache nodes, the set of cache expiration tags, and the cache capacity constraint boundary in the hierarchical cache graph; identify the cache node group that has a risk of conflict between the cache capacity constraint boundary and the cache requirement of the newly added pre-labeled task; and determine the corresponding cache eviction risk period. Based on the annotation worker's job priority parameters, the cache nodes are sorted, and the cache nodes that are in the cache eviction risk period and have low reuse priority are marked as cache groups that can be evicted. The releaseable space fragments in their cache space are extracted to construct an adjustable cache space window set. Based on the concurrent task scheduling data of the annotation platform, spatial segments that meet the task efficiency requirements in the adjustable cache space window set are identified, and a set of dynamic scheduling windows as candidates is output. The cache groups that can be eliminated, the spatial segments that can be optimized, and the set of dynamic scheduling windows are sent to the annotation platform scheduling center. By combining the cache groups that can be eliminated, the spatial segments that can be optimized, and the set of dynamic scheduling windows, a corresponding pre-annotation intelligent optimization scheme is generated.

[0013] Furthermore, to achieve the above objectives, the present invention also provides an asynchronous visual target automatic tracking and pre-annotation system for data annotation platforms, the system comprising: The scenario modeling module is used to acquire historical annotation task data from the data annotation platform, identify typical annotation scenarios, construct annotation scenario models, extract task feature parameters for target tracking pre-annotation under typical annotation scenarios, and generate corresponding annotation efficiency impact labels. The hierarchical processing module takes the labels affecting annotation efficiency and the real-time tracking pre-annotation request data initiated by the annotators as input, processes them based on the hierarchical task processing mechanism, and outputs the target tracking pre-annotation results and asynchronous task status records. The dual-layer caching module is used to set up a hierarchical caching architecture, mapping the media resources and inference results of historical tracking and pre-annotation to the hierarchical caching architecture respectively, and constructing a dual-layer caching structure of resource caching layer and result reuse layer; The intelligent scheduling module is used to analyze the constraint boundary between the execution efficiency and resource consumption of annotation tasks based on the hierarchical caching structure. It combines the concurrent task scheduling data of the annotation platform with the priority parameters of the annotators' jobs to output a set of asynchronous task dispatch scheduling windows and generate intelligent optimization strategies for pre-annotation tasks.

[0014] Furthermore, to achieve the above objectives, the present invention also provides an asynchronous visual target automatic tracking pre-annotation device for a data annotation platform. The device includes: a memory, a processor, and an asynchronous visual target automatic tracking pre-annotation program for a data annotation platform stored in the memory and executable on the processor. The asynchronous visual target automatic tracking pre-annotation program for a data annotation platform is configured to implement the steps of the asynchronous visual target automatic tracking pre-annotation method for a data annotation platform as described in any of the above descriptions.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an asynchronous visual target automatic tracking pre-annotation program for a data annotation platform, wherein when the asynchronous visual target automatic tracking pre-annotation program for a data annotation platform is executed by a processor, it implements the steps of the asynchronous visual target automatic tracking pre-annotation method for a data annotation platform as described above.

[0016] This invention provides an asynchronous visual target automatic tracking pre-annotation method for data annotation platforms. The method constructs typical annotation scenario models and generates labels affecting annotation efficiency, achieving refined modeling of different task complexities and resource requirements, thus improving the system's perception and response capabilities to pre-annotation task characteristics. It employs a hierarchical task processing mechanism, fusing real-time requests with historical experience labels, supporting asynchronous, pipelined task execution, effectively alleviating system congestion in high-concurrency scenarios, and improving platform stability and response efficiency. By establishing a dual-layer cache structure of resource caching and result reuse, it achieves hierarchical storage and efficient reuse of media data and inference results, significantly reducing redundant computation overhead, improving pre-annotation speed, and saving computing resources. Furthermore, by combining hierarchical cache analysis and task scheduling data, it dynamically identifies resource constraint boundaries and generates asynchronous task dispatch windows and intelligent optimization strategies, achieving task scheduling optimization based on actual load and business priorities, enhancing the system's adaptability and overall throughput. In summary, this method significantly improves the automation level, resource utilization, and large-scale task processing efficiency of data annotation platforms while ensuring pre-annotation accuracy. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms according to the present invention; Figure 2 This is a structural block diagram of an embodiment of the asynchronous visual target automatic tracking and pre-annotation system for data annotation platforms according to the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms according to the present invention. An embodiment of the asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms according to the present invention is presented.

[0021] In one embodiment, the asynchronous visual target automatic tracking and pre-annotation method for a data annotation platform includes: Step S100: Obtain historical annotation task data from the data annotation platform, identify typical annotation scenarios, construct annotation scenario models, extract task feature parameters for target tracking pre-annotation under typical annotation scenarios, and generate corresponding annotation efficiency impact labels.

[0022] Historical annotation task data can be an unstructured log collection accumulated by the data annotation platform during its past operations, containing video metadata, annotation time, resource consumption, and manual correction records. This data can serve as the original sample source for constructing annotation scene models and extracting task feature parameters, supporting the digital reconstruction of historical experience. In this embodiment, obtaining historical annotation task data from the data annotation platform and identifying typical annotation scenes can involve batch extraction of historical records from a database or log system and automatic classification of scene categories using clustering algorithms. Furthermore, obtaining historical annotation task data from the data annotation platform and identifying typical annotation scenes can be achieved by using K-Means clustering based on task metadata, DBSCAN density clustering based on annotation time distribution, or topic modeling based on annotation text descriptions for semantic classification. This transforms massive amounts of heterogeneous historical data into structured scene knowledge and lays the foundation for model construction. Typical annotation scenes can be a set of task categories with similar visual features, target motion patterns, and annotation difficulty distributions, derived from cluster analysis of historical annotation task data. This can be used to abstract discrete historical tasks into modelable standard units and provide a classification basis for subsequent refined resource evaluation. For example, typical annotation scenarios can be automatically summarized by performing multi-dimensional feature extraction and unsupervised clustering algorithms on historical data to identify frequently occurring task patterns. In one exemplary embodiment, typical annotation scenarios may include, but are not limited to, high-density small target tracking scenarios, low-light long-tail target detection scenarios, and high-dynamic blurry target recognition scenarios.

[0023] A labeling scenario model can be a mathematical or logical expression entity describing the mapping relationship between task feature parameters and labeling efficiency in a typical labeling scenario. It can be used to predict the complexity and resource requirements of unknown pre-labeling tasks and replace traditional static configuration methods. Furthermore, the labeling scenario model can establish a transformation function from the feature space to the efficiency space based on historical statistical data of typical labeling scenarios and through regression analysis or machine learning fitting. In this embodiment, the labeling scenario model can collaborate with labeling efficiency to influence labels, and the model output directly determines the generated label content. Simultaneously, the labeling scenario model can collaborate with a hierarchical task processing mechanism and provide prior knowledge input for real-time requests. Task feature parameters can be a set of key variables that quantitatively characterize the pre-labeling performance under a specific labeling scenario. They can be used to transform unstructured scenario descriptions into computable numerical indicators and support automated decision-making. In a specific embodiment, constructing a labeling scenario model and extracting task feature parameters can involve selecting key feature variables for each typical scenario and training a prediction model. Furthermore, constructing labeled scene models and extracting task feature parameters can be achieved by using multiple linear regression to establish an explicit relationship between features and time consumption, using gradient boosting trees to fit nonlinear mappings, or using neural network encoders to extract implicit feature representations. This enables generalizable efficiency prediction capabilities and supports immediate evaluation of new tasks. For example, task feature parameters may include, but are not limited to, video resolution and frame rate parameters, target density and occlusion rate parameters, and motion trajectory continuity parameters.

[0024] The labeling efficiency impact tag can be a structured metadata generated based on the labeling scenario model, used to identify the expected time consumption and resource sensitivity of the current pre-labeling task. It can be used as the scheduling basis for a hierarchical task processing mechanism and achieve dynamic perception of task priority rather than fixed allocation. In this embodiment, the labeling efficiency impact tag can be generated by inputting the feature vector of real-time requests into the labeling scenario model and outputting the corresponding efficiency level or weight value after inference calculation. Furthermore, the labeling efficiency impact tag can be coordinated with real-time tracking pre-labeled request data, and the two can be fused and input into the hierarchical task processing mechanism to trigger a differentiated processing flow. For example, generating the corresponding labeling efficiency impact tag can be achieved by inputting real-time request features into the constructed scenario model and outputting standardized tags, thereby simplifying the complex model inference results into lightweight scheduling signals and reducing runtime decision-making overhead.

[0025] Step S200: The labeling efficiency-affected labels and the real-time tracking pre-labeling request data initiated by the labelers are taken as input, processed based on the hierarchical task processing mechanism, and the target tracking pre-labeling results and asynchronous task status records are output.

[0026] The real-time tracking pre-annotation request data can be a data packet containing information about the video segment to be annotated and operation instructions, initiated by the annotator in real time during the operation. It can carry the current business needs and serve as a direct input source to trigger the asynchronous pre-annotation process. The hierarchical task processing mechanism can be a logical control framework that integrates real-time requests with historical experience tags and performs asynchronous routing execution according to priority and resource status. It can decouple the request reception and inference execution processes and avoid synchronous blocking under high concurrency to improve system throughput stability. For example, the hierarchical task processing mechanism can receive tags affecting annotation efficiency and real-time requests and allocate tasks to different processing queues or pipeline stages according to a preset strategy. In this embodiment, the hierarchical task processing mechanism can work with an asynchronous task dispatch scheduling window set and determine the actual execution time of tasks based on window constraints. Simultaneously, the hierarchical task processing mechanism can collaborate with a two-layer cache structure and adjust the processing path based on cache hits. Furthermore, processing real-time tracking pre-annotation request data and tags affecting annotation efficiency based on the hierarchical task processing mechanism can involve parsing the tags and request content and routing them to the corresponding asynchronous processing channels according to preset rules. In an exemplary embodiment, the hierarchical task processing mechanism for real-time tracking of pre-labeled request data and the labeling efficiency impact of labels can be achieved by dividing tasks into high, medium and low priority queues according to label level and scheduling them separately, guiding tasks to fast paths or full inference paths based on cache hit prediction, or dynamically adjusting batch size based on resource availability for merging processing. This enables differentiated treatment and elastic execution of requests and avoids resource waste or response delays caused by a one-size-fits-all approach.

[0027] The pre-annotation results for target tracking can be structured annotation data generated by model inference, including target bounding boxes, category confidence scores, and cross-frame association IDs. This data can serve as an initial reference for annotators and reduce the workload of manual annotation from scratch. Asynchronous task status records can be persistent information that tracks and stores the progress and results of the pre-annotation task at each stage of the asynchronous execution flow in real time. This information can support front-end polling or callback notifications and ensure the continuity and observability of the user experience. In this embodiment, the output of the target tracking pre-annotation results and asynchronous task status records can be achieved by writing the results to storage and updating the status table after inference is completed, and then notifying the front-end via a message queue. This ensures the reachability and transparency of the asynchronous process results and maintains smooth human-computer collaboration.

[0028] Step S300: Set up a hierarchical caching architecture, map the media resources and inference results of historical tracking and pre-annotation to the hierarchical caching architecture respectively, and build a two-layer caching structure of resource caching layer and result reuse layer.

[0029] The hierarchical caching architecture can be a multi-level storage organization system designed for different access patterns and reuse values ​​of media data and inference results. It can balance storage costs and access speed, and avoid the problem that a single cache layer cannot handle both large file reading and high-frequency small object queries. For example, the hierarchical caching architecture can divide the cache space into independent levels based on data type and lifecycle, and configure differentiated eviction and indexing strategies. In this embodiment, the hierarchical caching architecture can collaborate with the resource cache layer and the result reuse layer, providing them with a unified addressing and management interface. Simultaneously, the hierarchical caching architecture can collaborate with constraint boundary analysis, and the cache status directly affects resource occupancy assessment. Furthermore, setting up a hierarchical caching architecture and mapping media resources and inference results to corresponding levels can initialize multi-level cache instances and configure data routing rules and lifecycle strategies, thereby establishing a hierarchical storage infrastructure and providing physical support for subsequent efficient reuse. Media resources can be the original video frame sequences, decoding intermediates, or preprocessed tensor data required for pre-annotation tasks. They can be used as input materials for inference calculations, and their loading latency directly affects end-to-end response time. The inference results can be intermediate or final products such as detection boxes, tracking IDs and feature vectors output by the model after performing forward propagation on a specific media resource. They can be directly reused under the same or similar input conditions and avoid repeated calculations.

[0030] The resource caching layer can be a dedicated cache partition for storing and quickly retrieving media resources, focusing on bandwidth optimization and large-block data reading. It can reduce disk I / O and network transmission overhead and accelerate the delivery of video data. In one specific embodiment, the resource caching layer may include, but is not limited to, memory-mapped file caching, distributed object storage caching, and local SSD read caching. The result reuse layer can be a dedicated cache partition for storing and matching inference results, focusing on semantic similarity retrieval and high-frequency access to small objects. It can be used to achieve cross-task and cross-time inference result sharing and significantly reduce GPU computing power consumption. For example, the result reuse layer may include, but is not limited to, exact hash matching caching, feature vector approximate search caching, and time-series segment cascaded caching. In this embodiment, constructing a two-layer cache structure of resource caching and result reuse can be achieved by clearly dividing two functionally independent cache domains within a layered architecture and deploying dedicated indexes. This allows for decoupling and optimization of media I / O and inference computation, with each domain employing the most suitable access strategy.

[0031] Step S400: Based on the hierarchical caching structure, analyze the constraint boundary between the execution efficiency and resource consumption of annotation tasks, combine the concurrent task scheduling data of the annotation platform and the job priority parameters of the annotators, output the set of asynchronous task dispatch scheduling windows, and generate a pre-annotation task intelligent optimization strategy.

[0032] The annotation task execution efficiency can be a metric measuring the number of effective pre-annotation tasks completed per unit time or the average single-task processing latency. It can be used as an input to the performance dimension of constraint boundary analysis and reflects the current upper limit of the system's service capacity. Resource consumption can be the actual consumption of computing and storage resources such as CPU, GPU, memory, and I / O during the execution of pre-annotation tasks. It can be used as an input to the cost dimension of constraint boundary analysis and prevent system overload. The constraint boundary can be a dynamic set of thresholds that achieve a balance between annotation task execution efficiency and resource consumption under the current system configuration. It can be used to define the safe operating range of asynchronous scheduling and avoid performance degradation caused by over-scheduling. Furthermore, the constraint boundary can be fitted in real-time to determine the feasible domain boundary by analyzing the hierarchical cache hit rate, task queue length, and hardware load feedback. In this embodiment, the constraint boundary can be coordinated with the asynchronous task dispatch scheduling window set, and window generation is strictly limited by the current constraint boundary. For example, analyzing the constraint boundary between annotation task execution efficiency and resource consumption based on the hierarchical cache structure can involve continuously monitoring the cache hit rate, queue depth, and hardware utilization and dynamically fitting a performance-cost curve. In an exemplary embodiment, the constraint boundary between the execution efficiency and resource consumption of the analysis and annotation task based on the hierarchical caching structure can be achieved by updating the boundary parameters in real time through online learning algorithms, pre-setting multiple sets of boundary templates through stress testing and switching them according to the load, or exploring the optimal boundary point through reinforcement learning. This can obtain a dynamic safety zone that adapts to the current working conditions and avoid the static threshold from failing when the load fluctuates.

[0033] Concurrent task scheduling data can be real-time telemetry information reflecting the status, queuing situation, and resource contention of all active pre-labeled tasks in the system at the current moment. This can provide load-aware context and ensure timely scheduling decisions. Annotator job priority parameters can be dynamically generated by business rules or user behavior, representing the urgency or importance of specific annotation tasks. This can ensure high-value or urgent tasks receive priority response when resources are scarce. The asynchronous task dispatch scheduling window set can be a time slice or resource quota sequence generated within constraint boundaries that allows new tasks to be injected for execution. This can be used to discretize continuous scheduling decisions into controllable execution permissions and smooth traffic peaks. Furthermore, the asynchronous task dispatch scheduling window set can combine constraint boundaries, concurrent scheduling data, and priority parameters, and dynamically calculate available dispatch opportunities using a sliding window algorithm or token bucket mechanism. In this embodiment, the asynchronous task dispatch scheduling window set can collaborate with the pre-labeled task intelligent optimization strategy, and the window set is the concrete carrier for implementing the strategy. For example, combining concurrent task scheduling data with annotator job priority parameters to output an asynchronous task dispatch scheduling window set can calculate available dispatch opportunities and quotas within constraints by comprehensively considering real-time load and business priority. This transforms abstract scheduling strategies into specific execution permissions and achieves refined traffic shaping. The pre-annotated task intelligent optimization strategy can be an adaptive rule set generated based on the current system state and business objectives, guiding task sorting, cache preheating, and resource allocation. It can be used to achieve dynamic optimization beyond fixed rules and continuously improve overall throughput and response experience. In a specific embodiment, generating the pre-annotated task intelligent optimization strategy can be based on the window set and system feedback, dynamically adjusting task sorting, cache preheating, and resource allocation rules. This enables the system to have continuous self-optimization capabilities and maintain high performance under varying loads.

[0034] Taking autonomous driving urban road video annotation as an example, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be as follows: an annotator uploads a video of an urban intersection in rainy weather and requests vehicle tracking pre-annotation. The system first extracts the resolution, frame rate, and rain / fog intensity features of the video, inputs them into the annotation scene model, and generates high-difficulty, low-cache-hit-efficiency labels. The hierarchical task processing mechanism classifies the request into a medium-priority queue and marks it as requiring complete inference. At the same time, the resource cache layer attempts to load video frames but fails due to the unique rainy scene, and the result reuse layer also has no matching historical inference. The system analyzes the current high GPU load and cache miss and determines that the constraint boundary is tightening. The generated asynchronous dispatch window delays the execution of the task by 5 seconds. After the window opens, the system performs complete inference and stores the newly generated vehicle tracking results in the result reuse layer, while updating the status record to notify the annotator. When another annotator submits a similar rainy intersection video, the result reuse layer successfully hits the target, the task is promoted to fast path processing, and the waiting time is significantly shortened.

[0035] In one embodiment, the process of acquiring historical annotation task data from a data annotation platform, identifying typical annotation scenarios, and constructing an annotation scenario model includes: By acquiring historical annotation task data generated by the data annotation platform during the historical annotation process, the historical annotation task data is constructed into an annotation task data matrix according to the annotation task type, resource size, and submission time.

[0036] The annotation task data matrix can be a structured two-dimensional or three-dimensional data table formed by arranging historical annotation task data according to three dimensions: annotation task type, resource size, and submission time order. It can be used to transform discrete and heterogeneous historical records into a standard format that can be directly processed by matrix operations and block partitioning algorithms, eliminating analytical obstacles caused by data disorder. In this embodiment, the annotation task data matrix can be generated by sorting and filling the original logs using submission time as the row index, resource size as the column index, and annotation task type as the layer index. Furthermore, the annotation task data matrix can be coordinated with a cache capacity hierarchical partitioning mechanism, and the dimensional structure of the matrix directly determines the partitioning direction and granularity of the block partitioning strategy. Constructing the annotation task data matrix by arranging historical annotation task data according to annotation task type, resource size, and submission time order can be achieved by cleaning, sorting, and dimension mapping the original historical data, and filling in missing values ​​to generate a regular matrix structure. For example, constructing a labeling task data matrix by labeling task type, resource size, and submission time order can be achieved by using in-memory PandasDataFrame to build and export it as a sparse matrix format, using a distributed computing framework to build HDF5 matrix files in parallel on a cluster, or using database window functions to directly generate materialized view tables in the storage layer. This can achieve the technical effect of establishing a standardized data base and enabling subsequent block division and feature extraction to be performed efficiently on a unified structure.

[0037] The annotation task data matrix is ​​divided into blocks based on the cache capacity hierarchical partitioning mechanism to obtain multiple historical annotation task units, and the media resource fragments and task attribute fragments corresponding to each historical annotation task unit are extracted.

[0038] The cache capacity hierarchical partitioning mechanism can be a set of partitioning rules for adaptively dividing the annotation task data matrix into blocks based on the capacity hierarchy and access characteristics of the underlying cache storage units. This can be used to align the data block division results with the physical cache structure and avoid excessively large blocks causing cache overflow or excessively small blocks causing frequent swapping in and out. In an exemplary embodiment, the cache capacity hierarchical partitioning mechanism can read system cache configuration parameters and dynamically calculate the optimal block size and boundaries for each cache layer based on the matrix data distribution characteristics. Furthermore, the cache capacity hierarchical partitioning mechanism can be coordinated with the annotation task data matrix, with the mechanism acting on the matrix to generate historical annotation task units and cooperating with historical annotation task units, where the block division result is the storage and processing boundary of the unit. The historical annotation task unit can be the smallest self-contained historical data processing block obtained after being partitioned by the cache capacity hierarchical partitioning mechanism, both logically and in storage. It can be used as the basic operation unit for subsequent feature extraction and fragment decoupling, ensuring that the processing granularity is consistent with the cache I / O granularity. In a specific embodiment, the historical annotation task unit can include, but is not limited to, complete single-task units, cross-task splicing units, and time-slice units. Media resource fragments can be subsets of raw visual data, such as video frames or image sequences, separated from historical annotation task units. They can be used to physically decouple media content from metadata and support independent caching and on-demand loading.

[0039] Task attribute fragments can be subsets separated from historical annotation task units, containing only non-media metadata such as annotation type, execution time, or correction records. These fragments can be used to centrally carry structured semantic information suitable for feature extraction and avoid mixing with large volumes of media data. Multiple historical annotation task units are obtained by partitioning the annotation task data matrix into blocks based on a cache capacity hierarchical partitioning mechanism. This can be achieved by dynamically determining the block size according to the current cache level configuration and cutting along the matrix dimensions to generate self-contained units. Furthermore, this partitioning can be accomplished by setting three-level partitioning thresholds according to the L1 / L2 / L3 cache capacity ratio, adaptively adjusting the block boundaries based on matrix data density to balance unit load, or using a Z-order curve to traverse the matrix to maintain spatial locality. This achieves the technical effect of matching data processing units with cache physical characteristics and improving the IO efficiency of subsequent fragment extraction and feature calculation. Extracting the media resource fragments and task attribute fragments corresponding to each historical annotation task unit can be done by separating media data and metadata according to a predefined schema and storing them in independent storage paths for each historical annotation task unit. Furthermore, the extraction of media resource fragments and task attribute fragments corresponding to each historical annotation task unit can be achieved through a differentiated storage strategy, thereby realizing the physical decoupling of content and semantics and supporting the technical effects of differentiated caching strategies and parallel processing.

[0040] For each task attribute fragment, multidimensional task features are extracted. The multidimensional task features extracted from the same task attribute fragment are combined to form a task feature vector. A set of task feature vectors is constructed using the task feature vectors of all task attribute fragments.

[0041] Multidimensional task features can be numerical indicators extracted from a single task attribute fragment, representing the characteristics of the task in multiple orthogonal dimensions. They can be used to transform abstract task attributes into measurable and comparable feature space coordinates. In this embodiment, multidimensional task features may include, but are not limited to, spatiotemporal complexity features, annotation interaction intensity features, and scene semantic richness features. A task feature vector can be a one-dimensional numerical array formed by concatenating all multidimensional task features extracted from the same task attribute fragment in a fixed order. It can be used to compress the multidimensional characteristics of a single task into a unified vector representation and support distance calculation and similarity measurement in the vector space. A task feature vector set can be a vector space dataset formed by aggregating the task feature vectors corresponding to all historically annotated task units. It can be used to construct the input sample space for cluster analysis and support the automatic discovery of typical annotation scenarios. For example, the task feature vector set can be obtained by traversing all historically annotated task units and sequentially extracting task feature vectors and appending them to the set container. Furthermore, the task feature vector set can be a feature aggregation carrier that collaborates with multidimensional task features and a scene classification result produced after clustering of the set with typical annotation scenarios.

[0042] Extracting multi-dimensional task features for each task attribute fragment and assembling them into a task feature vector can be achieved by applying a feature engineering pipeline to the task attribute fragments and calculating the indicators of each dimension, then concatenating them into a vector in sequence. In an exemplary embodiment, extracting multi-dimensional task features for each task attribute fragment and assembling the task feature vector can be achieved by using a manually designed feature extractor to calculate statistical indicators, using a pre-trained encoder to embed attribute text, or using a combination of statistical features and semantic embeddings to concatenate a composite vector. This achieves the technical effect of transforming unstructured attributes into a computable vector representation and supporting subsequent clustering analysis. Constructing a task feature vector set from the task feature vectors of all task attribute fragments can be achieved by traversing all units and collecting the task feature vectors to assemble them into a unified dataset. Furthermore, constructing a task feature vector set from the task feature vectors of all task attribute fragments can be achieved through batch aggregation, thereby achieving the technical effect of forming a complete sample space and providing sufficient data coverage for clustering.

[0043] Cluster analysis is performed on the task feature vector set to identify representative typical annotation scenarios, including static image sequence annotation, VLA video annotation, and multi-camera cross-frame annotation. Based on the multi-dimensional task features extracted from each typical annotation scenario, an annotation scenario model is constructed.

[0044] Static image sequence annotation can be a task mode for object detection and classification of discrete image sets with no or weak temporal correlation. It can be used to represent a business scenario with low temporal dependence and high concurrency IO requirements, and its caching and scheduling strategies focus on batch image prefetching. VLA video annotation can be a task mode for object tracking and behavior recognition of video streams with strong temporal continuity and action semantics. It can be used to represent a business scenario with high temporal dependence and computational intensity, and its caching and scheduling strategies focus on inter-frame state reuse and GPU pipeline optimization. Multi-camera cross-frame annotation can be a task mode for cross-view object association and 3D trajectory reconstruction of video sequences from multiple synchronous camera perspectives. It can be used to represent a business scenario with high spatial correlation and data bandwidth sensitivity, and its caching and scheduling strategies focus on multi-path data alignment and joint inference result caching.

[0045] Cluster analysis of task feature vector sets to identify representative typical labeled scenarios can be performed by applying clustering algorithms to unsupervised group the vector set and selecting the cluster centers or high-density regions as representative typical scenarios. In a specific embodiment, cluster analysis of task feature vector sets to identify representative typical labeled scenarios can be achieved by using K-Means clustering and selecting the nearest centroid sample as the representative, using HDBSCAN density clustering to automatically determine the number of clusters and extract core points, or using spectral clustering to divide communities based on similarity graphs and select hub nodes. This achieves the technical effect of automatically discovering hidden business patterns and avoiding the subjectivity and omissions of manually defined scenarios. A labeled scenario model is constructed by summarizing the multi-dimensional task features extracted from each typical labeled scenario. This can be achieved by statistically summarizing the multi-dimensional task features within each typical scenario cluster or by model fitting to generate the feature prototype or prediction function for that scenario. Furthermore, the construction of a labeled scenario model based on summarizing the multi-dimensional task features extracted from each typical labeled scenario can be achieved through feature aggregation. This achieves the technical effect of solidifying the clustering results into reusable scenario knowledge and supporting the efficiency evaluation and scheduling decisions of subsequent real-time requests.

[0046] For example, in the scenario of an autonomous driving multi-sensor fusion annotation platform, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be based on the platform accumulating millions of historical annotation tasks. The system first constructs a three-dimensional annotation task data matrix according to task type, resource size, and submission time. Then, based on the hierarchical configuration of GPU memory and SSD cache, the matrix is ​​divided into thousands of historical annotation task units, with each unit size adapted to L2 cache lines. Video frame sequences are separated from each unit as media resource segments, and annotation time and sensor configuration are separated as task attribute segments. Multi-dimensional task features such as spatiotemporal complexity, interaction frequency, and semantic entropy are extracted from the attribute segments and concatenated into a 128-dimensional task feature vector. HDBSCAN clustering is performed on all vectors to automatically identify three typical scenarios: static image sequence annotation, VLA video annotation, and multi-camera cross-frame annotation. Finally, the feature distribution of each scenario is summarized to construct an annotation scenario model for efficiency label generation and caching strategy matching for subsequent real-time requests.

[0047] In one embodiment, the process of extracting task feature parameters for pre-annotation of target tracking in typical annotation scenarios and generating corresponding labels affecting annotation efficiency includes: For various typical annotation scenarios included in the annotation scenario model, based on statistically representative multidimensional task features in the task feature vector set, task feature parameters are extracted as descriptive terms affecting annotation efficiency.

[0048] The description of the impact of annotation efficiency can be a statistically representative instance of task feature parameters selected from the task feature vector set that can be directly mapped to the cache benefit evaluation dimension. This can be used to transform abstract high-dimensional features into concrete evaluation inputs that can be directly consumed by the rule set and to establish a semantic bridge between task attributes and cache value. In this embodiment, the description of the impact of annotation efficiency can be obtained by performing statistical analysis on the task feature vector set and selecting feature dimensions with high variance contribution rates or strong correlation with performance indicators. Extracting task feature parameters as descriptions of the impact of annotation efficiency based on statistically representative multi-dimensional task features from the task feature vector set can be achieved by performing feature importance evaluation or dimensionality reduction analysis on the task feature vector set and selecting key feature dimensions as descriptions. Furthermore, this extraction process can be implemented by using principal component analysis to extract the feature combination with the largest variance, using mutual information to select features with the highest correlation to performance indicators, or using LASSO regression to automatically select features with non-zero coefficients. This ensures that the features of the input rule set have sufficient information and representativeness and avoids redundant or irrelevant features interfering with the evaluation accuracy.

[0049] Each type of annotation efficiency impact description is input into a preset cache benefit rule set. The cache benefit rule set is constructed based on the resource download time, storage usage, and annotation acceleration effect samples of historical pre-annotated tasks through correlation analysis or logistic regression inference, and outputs the cache benefit intensity coefficient range and its resource reuse priority.

[0050] The cache benefit rule set can be a set of logical or mathematical models built based on historical pre-annotated task samples to quantitatively evaluate the expected benefits of caching operations under different task characteristics. It can replace manual experience or simple heuristic judgment and provide an objective and calculable basis for caching decisions. In an exemplary embodiment, the cache benefit rule set can be coordinated with annotation efficiency impact descriptions to receive the descriptions as input, and can be coordinated with cache benefit intensity coefficient ranges to output the range as the evaluation result. Resource download time can be a measure of the time consumed from obtaining media resources from storage media to the computing unit in historical pre-annotated tasks, and can be used as a cost indicator to measure the necessity of caching. Storage usage can be the space occupied by media resources or inference results corresponding to historical pre-annotated tasks in the cache, and can be used as a constraint indicator to measure caching costs to balance benefits and storage overhead. Annotation acceleration effect samples can be records of the actual reduction in annotation time or the increase in processing throughput due to cache hits in historical pre-annotated tasks, and can be used as a direct performance indicator to measure caching benefits and reflect the actual contribution of caching to business value. The cache benefit intensity coefficient range can be a numerical range output by the cache benefit rule set, representing the expected cache benefit level under specific task characteristics. It can be used to discretize continuous benefit evaluation results into actionable hierarchical signals to support the formulation of differentiated strategies. For example, the cache benefit intensity coefficient range can be coordinated with resource reuse priority to determine priority assignment, and can be coordinated with annotation efficiency to affect labels as a quantitative basis for label generation.

[0051] Resource reuse priority can be a level identifier determined based on the range of cache benefit intensity coefficients, representing the priority of resources to be cached or scheduled in specific typical annotation scenarios. It can be used to guide the eviction strategy of hierarchical caching and the execution order of task scheduling to ensure that high-value resources receive preferential treatment. Inputting each type of annotation efficiency impact description into a preset cache benefit rule set can be achieved by batch feeding the corresponding annotation efficiency impact descriptions into the rule set according to typical annotation scenario categories for inference, thereby realizing scenario-level batch evaluation and supporting subsequent priority binding and label generation. Constructing the cache benefit rule set through correlation analysis or logistic regression inference can be achieved by using historical sample data to train statistical models or fit logistic functions to establish a mapping from features to benefits. In a specific embodiment, this construction process can be implemented by using Pearson or Spearman correlation analysis to screen strongly correlated features and set threshold rules, using multivariate logistic regression to fit benefit probabilities and divide confidence intervals, or using decision trees to summarize explicit rule sets, thereby constructing an objective evaluation model in a data-driven manner and avoiding subjective bias to improve rule generalization ability. The output cache benefit intensity coefficient range and its resource reuse priority can be calculated by the rule set on the input descriptor and return the benefit range and corresponding priority level, thereby transforming the complex evaluation results into structured decision signals for direct use by downstream modules.

[0052] By identifying the cache benefit intensity range for various typical annotation scenarios, resource reuse priority is bound to the corresponding typical annotation scenarios, and labels affecting annotation efficiency are generated.

[0053] Among them, binding resource reuse priority with corresponding typical annotation scenarios and generating annotation efficiency impact labels can be done by associating each typical annotation scenario with its evaluated resource reuse priority and encapsulating it into a standardized label structure, so that the labels carry quantified cache value information to support subsequent refined decisions on hierarchical caching and asynchronous scheduling.

[0054] Taking an intelligent security video annotation platform as an example, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be used for typical annotation scenarios of pedestrian tracking in low light at night that have been identified. The system extracts three statistically representative multi-dimensional task features from the task feature vector set: average brightness value, average pixel area of ​​the target, and inter-frame motion vector entropy, as descriptive terms affecting annotation efficiency. These descriptive terms are input into a cache benefit rule set pre-built based on historical tasks. This rule set analyzes the resource download time, storage occupation, and annotation acceleration effect samples of past nighttime videos and uses logistic regression to fit and derive a benefit model. The rule set outputs that the cache benefit intensity coefficient of this scenario is in the range of 0.75 to 0.92 and the corresponding resource reuse priority is high. The system binds the high priority with the pedestrian tracking scenario in low light at night and generates an annotation efficiency impact label containing this priority information. Subsequently, when an annotator initiates a similar video pre-annotation request, the system loads the video resources of the task into the resource cache layer first based on this label and assigns a higher execution weight during task scheduling, thereby maximizing the response speed and annotation efficiency of high-value scenarios under limited storage resources.

[0055] In one embodiment, the process of taking the labeling efficiency-affecting labels and the real-time tracking pre-labeling request data initiated by the labelers as input, processing it based on a hierarchical task processing mechanism, and outputting the target tracking pre-labeling results and asynchronous task status records is as follows: The real-time tracking pre-annotation request initiated by the annotation front-end is processed for parameter verification and spatiotemporal registration. The real-time request data includes reference box coordinates, target tracking identifier, camera identifier, frame range, media resource information, and annotation task identifier.

[0056] The reference bounding box coordinates can be the position and size parameters of the initial bounding box of the target specified in the real-time tracking pre-annotation request within the image plane. These can be used as the starting input for the target tracking algorithm and to define the region of interest for pre-annotation. The target tracking identifier can be a unique ID identifying the instance of the target to be tracked, ensuring that the asynchronously returned results correctly correspond to the specific target object on the front-end annotation interface. The camera identifier can be the sensor device number or logical channel number identifying the acquired media resources, supporting data routing and calibration parameter loading in multi-camera scenarios to ensure accurate spatiotemporal registration. The frame range can be the start and end sequence numbers or timestamp intervals of the video frames to be pre-annotated specified in the real-time request, defining the temporal boundaries of the algorithm to avoid invalid computation and resource waste. The media resource information can be a collection of metadata describing the storage path, format, resolution, etc., of the video or image sequence to be processed, providing addressing basis for the resource caching layer to support efficient data loading and preprocessing. The annotation task identifier can be a unique global ID identifying the current annotation session or job batch, used to associate business context and permission verification, and to support task-level status tracking and result aggregation. Performing parameter validation and spatiotemporal registration on real-time tracking pre-annotation requests can involve verifying the completeness and validity of request fields and aligning the spatiotemporal reference of multi-source data based on camera identifiers and frame ranges. Furthermore, this parameter validation and spatiotemporal registration can be achieved through methods such as using JSONSchema to validate parameter structures and calling calibration services for coordinate transformation, using protocol buffers to parse binary requests and look up tables to correct timestamp offsets, or using rule engines to validate business constraints and interpolate to complete missing frames. This approach ensures the quality and consistency of input data and prevents downstream processes from failing or erroneously due to dirty data.

[0057] Based on the hierarchical task processing mechanism, real-time request data is decoupled into business and algorithm layers to form hierarchical task processing segments.

[0058] The business-algorithm layered decoupling process can be a process of separating and encapsulating business logic parameters and algorithm execution parameters in real-time requests into independent schedulable units. In this embodiment, the business-algorithm layered decoupling process can operate by parsing the request data structure and splitting it into business control segments and algorithm computation segments according to responsibility boundaries, each carrying the necessary context. This can be used to break the tight coupling of synchronous call chains so that business responses and algorithm inference can flow asynchronously and in parallel. For example, the business-algorithm layered decoupling process can collaborate with layered task processing segments, which is the segment generation mechanism. The business-algorithm layered decoupling process can also collaborate with an asynchronous task sample set, with segments being bound and merged into the set. The layered task processing segment can be the smallest asynchronous execution unit bearing a single responsibility, generated after the business-algorithm layered decoupling process, and can be used as the basic granularity of asynchronous scheduling to support fine-grained resource allocation and state management. In an exemplary embodiment, the layered task processing segment may include, but is not limited to, business control segments, algorithm computation segments, and data preparation segments. By decoupling business and algorithm layers in real-time request data to form layered task processing segments, this can be achieved by parsing the request content and splitting the data according to business control and algorithm calculation responsibilities, then encapsulating these segments into independent fragments. Furthermore, this decoupling of business and algorithm layers in real-time request data to form layered task processing segments enables separation of concerns, allowing business responses and algorithm inference to be independently scheduled and scaled.

[0059] Each task processing segment is associated with its corresponding labeled efficiency impact tag to construct an asynchronous task sample set.

[0060] The asynchronous task sample set can be a structured task queue composed of multiple hierarchical task processing segments paired with their associated labeled efficiency impact tags. In a specific embodiment, the asynchronous task sample set can be obtained by traversing all processing segments, binding corresponding efficiency tags according to scene matching rules, and assembling them into a unified scheduling container. This can be used to enable each asynchronous unit to carry historical experience context to support differentiated priorities and resource strategies. The asynchronous task sample set can collaborate with hierarchical task processing segments; the set is the aggregation carrier of segments. The asynchronous task sample set can also collaborate with asynchronous task state trajectories; the sample execution processes in the set constitute the trajectory data source. Associating and binding each task processing segment with its corresponding labeled efficiency impact tag to construct the asynchronous task sample set can be achieved by matching efficiency tags according to the scene to which the segment belongs and packaging the two into a unified scheduling unit. Furthermore, associating and binding each task processing segment with its corresponding labeled efficiency impact tag to construct the asynchronous task sample set can endow each asynchronous unit with historical experience context, thereby supporting intelligent resource allocation.

[0061] Based on the asynchronous task sample set and each type of hierarchical processing link, multiple task processing segments are fused and analyzed in the order of business reception-state storage-algorithm call-result write-back. The flow characteristics of task parameters in each processing link are extracted, task state nodes are constructed, and they are associated with each other according to the order of processing links to form an asynchronous task state trajectory. The task state trajectory is then analyzed based on deep learning combined with multi-dimensional task features.

[0062] The process can be categorized into several stages. The business reception stage, which is the initial phase of an asynchronous task processing flow responsible for request validity verification, parameter parsing, and spatiotemporal registration, can filter invalid requests and standardize input to ensure the stability of subsequent stages. The state storage stage is an asynchronous processing stage that persists the task initialization state, intermediate progress, and exception information to the database, providing task observability and fault recovery capabilities to support front-end state polling. The algorithm invocation stage is an asynchronous processing stage that triggers pre-annotated model inference and executes target detection and tracking calculations, completing core visual computing tasks. The result write-back stage is an asynchronous processing stage that formats the algorithm output results and writes them to storage or pushes them to the front end, closing the task flow to ensure annotators receive pre-annotation auxiliary information in a timely manner. Flow characteristics are dynamic attributes describing the temporal sequence, dependencies, and data transformation characteristics of task parameters during transmission between processing stages, characterizing the inherent laws of task execution to provide quantitative basis for state node construction and trajectory analysis. Task state nodes are discretized snapshots representing the execution state of a task at a specific moment within a particular processing stage, serving as basic building blocks for asynchronous task state trajectories to support refined state tracking. An asynchronous task state trajectory can be a temporal sequence reflecting the entire lifecycle execution process of a single task, formed by connecting multiple task state nodes sequentially according to the processing stage. Furthermore, the asynchronous task state trajectory can be obtained by collecting flow characteristics and constructing state nodes in sequence at each stage of business reception, state storage, algorithm invocation, and result write-back. This can be used to provide a complete temporal view of task execution to support deep learning models in discovering execution patterns and identifying bottlenecks.

[0063] Asynchronous task status trajectories can collaborate with task status nodes, and the trajectory is composed of nodes. Asynchronous task status trajectories can also collaborate with multi-dimensional task features, enabling intelligent diagnosis through trajectory analysis combined with feature analysis. Multiple task processing segments can be fused and analyzed in the order of business reception-status storage-algorithm invocation-result write-back, which can involve sequentially executing each processing step and recording data flow and status changes between steps. Furthermore, fusing and analyzing multiple task processing segments in the order of business reception-status storage-algorithm invocation-result write-back can establish a standardized execution pipeline, thereby ensuring the controllability and observability of the asynchronous process. Extracting the flow characteristics of task parameters within each processing step and constructing task status nodes can involve collecting dynamic attributes such as parameter changes, time consumption, and dependencies at each step and generating status snapshots. For example, extracting the flow characteristics of task parameters within each processing stage and constructing task state nodes can be achieved through AOP aspect-based automatic capture of input and output parameters and execution duration, listening to event streams through message middleware to extract state changes, or collecting IO and CPU usage characteristics at the kernel layer through eBPF probes. This enables fine-grained digitization of the execution process and provides atomic data for trajectory construction.

[0064] Asynchronous task state trajectories are formed by associating processing stages sequentially. This can be achieved by linking state nodes generated by the same task at each stage into a sequence based on timestamps or logical numbers. Furthermore, associating processing stages sequentially allows for the reconstruction of the entire task lifecycle execution path, thus supporting end-to-end performance analysis. Analyzing task state trajectories based on deep learning combined with multi-dimensional task features involves inputting the trajectory sequence and multi-dimensional task features into a neural network model to learn execution and anomaly patterns. In an exemplary embodiment, analyzing task state trajectories based on deep learning combined with multi-dimensional task features can be achieved by using LSTM or GRU to model trajectory temporal dependencies and predict remaining time, using a Transformer encoder to capture long-range dependencies and identify bottlenecks, or using a graph neural network to model the trajectory as a DAG and optimize the critical path. This allows for the automatic discovery of patterns and bottlenecks from massive amounts of asynchronous execution data and drives the continuous evolution of scheduling strategies.

[0065] Taking a robot navigation video annotation platform as an example, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be as follows: the annotator selects pedestrians in the robot's view video from the front end and initiates a tracking pre-annotation request. The system first verifies whether the reference box coordinates are within the image range and whether the frame range is valid, and loads extrinsic parameters according to the camera identifier to complete spatiotemporal registration. Then, the request is decoupled into business control segments and algorithm calculation segments, and the annotation efficiency impact labels corresponding to the previously generated high dynamic blur target recognition scenarios are bound to them respectively to construct an asynchronous task sample set. The system executes in the order of business reception, status storage, algorithm call, and result write-back. The parameter flow characteristics of each link are collected, such as queue waiting time, GPU memory usage, inference time, etc., to construct task status nodes and connect them into an asynchronous task status trajectory. Finally, the trajectory and the multi-dimensional task features of the task are input into the pre-trained Transformer model. The analysis found that the GPU queue was too long due to the high resolution frame density in the algorithm call link. Based on this, the system dynamically adjusts the batch size and priority of subsequent similar tasks to avoid blocking accumulation.

[0066] In one embodiment, the process of analyzing the task state trajectory based on deep learning combined with multi-dimensional task features includes: Identify abnormal node ranges in the task flow.

[0067] The abnormal task flow node interval can be a sequence of consecutive state nodes identified by a deep learning model in the asynchronous task state trajectory, where the task feature parameters deviate from the normal pattern, causing execution delays or blockages. This can be used to accurately locate the specific execution stage and time period causing performance degradation, replacing the fuzzy diagnosis of global average indicators. In this embodiment, the abnormal task flow node interval can be obtained by inputting the asynchronous task state trajectory into a pre-trained anomaly detection model and outputting consecutive node segments with anomaly probabilities higher than a threshold. Furthermore, the abnormal task flow node interval can be coordinated with task anomaly data, and this interval represents the extraction range of the anomaly data. Additionally, the abnormal task flow node interval can be coordinated with the task safety period, and the interval includes a calculation window for recoverable waiting time. For example, identifying the abnormal task flow node interval can be achieved by inputting the asynchronous task state trajectory into an anomaly detection model and outputting an abnormal node sequence. In one exemplary embodiment, identifying abnormal node intervals in task flow can be achieved by using the isolated forest algorithm to detect outliers in the trajectory and aggregate them into intervals, using a variational autoencoder to reconstruct the trajectory and locate abnormal segments based on the reconstruction error, or using a temporal convolutional network to directly segment abnormal intervals in the trajectory. This can achieve the technical effect of accurate anomaly localization and avoid the insufficient granularity of global monitoring.

[0068] Extract the gradient of changes in task feature parameters, queuing time, and resource hit threshold points within the interval of abnormal task nodes.

[0069] The gradient of change can be the rate derivative of the change of task characteristic parameters with time or node number within the interval of task anomalies. It can be used to quantify the severity of anomalies to distinguish between transient jitter and persistent congestion. The queuing time can be the actual cumulative time a task waits for resource allocation in the corresponding scheduling stage within the interval of anomalies. It can be used to reflect the cumulative effect of resource contention and serve as a direct measure of the severity of congestion. The resource hit threshold can be the boundary parameter value at which the cache hit rate or resource reuse efficiency degrades from effective support to ineffective waiting within the interval of task anomalies. It can be used to identify the moment when a qualitative change occurs in the resource bottleneck, allowing the system to perceive the turning point from quantitative to qualitative change. In this embodiment, the resource hit threshold can be obtained by analyzing the inflection point relationship between resource hit rate and task time in historical anomaly samples and fitting the threshold value. Furthermore, the resource hit threshold can be coordinated with the task safety period, and the position of the threshold affects the start and end determination of the safety period. The resource hit threshold can also be coordinated with the gradient of change, and both jointly characterize the dynamic characteristics of anomalies. For example, extracting the gradient of changes in task feature parameters, queuing time, and resource hit threshold within an abnormal task node interval can be achieved by calculating the derivative of the feature parameters within the identified abnormal interval, accumulating the waiting time, and finding the inflection point of resource efficiency. In a specific embodiment, extracting the gradient of changes in task feature parameters, queuing time, and resource hit threshold within an abnormal task node interval can be achieved through multi-dimensional quantification of abnormal characteristics, thereby supporting the technical effect of differentiated response strategies.

[0070] Extract the recoverable waiting time within the scheduling stage corresponding to the abnormal node interval of the task, and record it as the task safety period.

[0071] The recoverable waiting time can be the maximum tolerable waiting time for a task within the scheduling stage corresponding to the task anomaly node interval without triggering a retry or dropping. It can be used to provide a flexible buffer window for the scheduler to avoid unnecessary task interruptions due to brief resource contention. The task safety period can be a time window defined by the recoverable waiting time, allowing tasks to continue waiting without being considered failures within the task flow anomaly node interval. It can be used to ensure the continuity of the asynchronous pipeline in anomaly states to improve system robustness. In this embodiment, the task safety period can be obtained by deriving the upper bound of the recoverable waiting time for each stage based on historical successful recovery case statistics and using it as the length of the safety period. Furthermore, the task safety period can be coordinated with the task flow anomaly node interval, and the safety period is a sub-time period within the interval. The task safety period can also be coordinated with task anomaly data, and the safety period is included as a key dimension in the anomaly dataset. For example, extracting the recoverable waiting time within the scheduling stage corresponding to the task anomaly node interval as the task safety period can be done by querying the historical recovery statistics of that scheduling stage and taking the upper bound of the recoverable waiting time as the length of the safety period. In an exemplary embodiment, the recoverable waiting time within the scheduling link corresponding to the abnormal node interval of the task can be extracted and recorded as the task safety period. This can be achieved by taking the P95 recoverable time based on percentile statistics, dynamically estimating the safety window under the current load based on reinforcement learning, or presetting a fixed safety period based on the SLA protocol and calibrating it according to the actual success rate. This can provide a flexible fault tolerance space for anomaly handling and reduce invalid retries.

[0072] The changing gradient, queuing time, critical threshold, and task safety period are recorded as task anomaly data. This task anomaly data can be a structured set of anomaly descriptions composed of the changing gradient, queuing time, resource hit critical threshold, and task safety period. It can be used to integrate multi-dimensional anomaly indicators and provide a complete diagnostic basis for subsequent scheduling optimization and strategy generation. For example, recording the changing gradient, queuing time, critical threshold, and task safety period as task anomaly data can be achieved by encapsulating these four indicators into structured data objects and associating them with corresponding tasks and anomaly intervals. In a specific embodiment, recording the changing gradient, queuing time, critical threshold, and task safety period as task anomaly data can be achieved by forming standardized anomaly descriptions, thereby facilitating unified consumption and analysis by downstream modules.

[0073] The system outputs pre-annotated frame-level results for target tracking and a complete asynchronous task status record. The pre-annotated frame-level results can be fine-grained pre-annotations generated for each frame within the requested frame range, including the target bounding box, category, and tracking ID. This meets the annotator's frame-by-frame accuracy requirements, supporting fine-grained manual verification and correction. The complete asynchronous task status record is a persistent log collection covering the status of each node in the task's entire lifecycle, anomaly data, and the final result. This provides end-to-end traceability to support fault debriefing, performance analysis, and model iteration. For example, outputting the pre-annotated frame-level results and the complete asynchronous task status record can be achieved by serializing the algorithm inference results frame by frame and writing the entire process status and anomaly data to persistent storage. In an exemplary embodiment, the output of the pre-annotated frame-level results and the complete asynchronous task status record can be implemented through closed-loop business delivery and system observation, thus achieving a technical effect that balances user experience and platform operation and maintenance needs.

[0074] Taking a highway vehicle tracking and annotation platform as an example, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be as follows: When the system is processing a pre-annotation task of a high-density traffic flow video, it can use a deep learning model to analyze the asynchronous task state trajectory and identify the abnormal task flow node interval in the algorithm call process. Within this interval, it can extract the GPU memory usage gradient change sudden increase, queuing time exceeding 8 seconds and cache hit rate dropping below 30%, which is the resource hit critical threshold point. At the same time, based on the historical data of this link, it is determined that the recoverable waiting time under the current load is 5 seconds, and a task safety period is set accordingly. The system packages the above-mentioned change gradient, queuing time, critical threshold point and safety period into task abnormal data. Since the task is still within the safety period, the scheduler does not trigger a retry but waits for 2 seconds before the GPU resources are released and the task is successfully completed. Finally, the system outputs the frame-level tracking box of each vehicle as the target tracking pre-annotation frame-level result and stores the complete state record containing abnormal data in the database for subsequent optimization of GPU scheduling strategy.

[0075] In one embodiment, a hierarchical caching architecture is set up, and the media resources and inference results of historical tracking and pre-annotation reuse are mapped to the hierarchical caching architecture respectively. The process of constructing a two-layer caching structure of resource caching layer and result reuse layer includes: Media resources and pre-annotated inference results are associated with virtual cache nodes at different levels in the hierarchical caching architecture. Based on the reuse frequency and size of media resources and inference results, a multi-level mapping chain for pre-annotated cache is established, and the multi-level mapping chain is used as the basic cache space axis.

[0076] In this system, virtual cache nodes can be storage units used in a hierarchical caching architecture for logically abstracting media resources or inference results. They do not directly correspond to physical memory addresses and can be used to decouple the logical organization and physical storage of cached data, supporting flexible multi-level mapping and attribute binding. Reuse frequency can be the statistical frequency of a specific media resource or inference result being hit and invoked in historical pre-annotation tasks. It can be used as an indicator to measure the value of cached data and guide the construction and priority ranking of multi-level mapping chains. Size can be the byte size of a media resource file or inference result after serialization. It can be used as a parameter for caching cost, and together with reuse frequency, determines the data's position on the cache space axis. A multi-level mapping chain can be a logical link established based on reuse frequency and size, representing the hierarchical storage relationship of cached data from high-frequency small objects to low-frequency large objects. It can be used to organize discrete cached objects into an ordered value gradient, supporting differentiated caching strategies. In this embodiment, the multi-level mapping chain can be obtained by performing two-dimensional analysis on historical access data and arranging it in a chain structure from high reuse rate and small size to low reuse rate and large size. Furthermore, the multi-level mapping chain can work in conjunction with the basic cache space axis to form the skeleton of the space axis, or it can work in conjunction with associated edges to connect nodes on the chain to form a topology. The basic cache space axis can be a one-dimensional or multi-dimensional coordinate system built on the basis of the multi-level mapping chain to describe the logical distribution of cache data. It can be used to provide a unified reference system for the location, sorting and association of cache nodes, giving cache management structured semantics.

[0077] Associating media resources with pre-annotated inference results to virtual cache nodes at different levels in a hierarchical caching architecture can be achieved by instantiating resources and results into virtual nodes at corresponding levels based on data type and preset hierarchical rules. This establishes logical cache units and lays the foundation for subsequent structured organization. Establishing a multi-level mapping chain for the pre-annotated cache based on the reuse frequency and size of media resources and inference results can be achieved by statistically analyzing the reuse frequency and size of historical access data and sorting them according to value gradients to form a chain structure. In an exemplary embodiment, establishing a multi-level mapping chain for the pre-annotated cache based on the reuse frequency and size of media resources and inference results can be achieved by using a two-dimensional binning method to divide the data into high-reuse, small-volume and low-reuse, large-volume intervals and linking them; using Pareto front analysis to select non-dominated points to construct the mapping chain; or using a weighted scoring function to combine reuse and volume to generate a one-dimensional sorting chain. This enables value grading of cached data and supports differentiated storage strategies. Using the multi-level mapping chain as the basic cache space axis can be achieved by defining the order relationship of the mapping chain as the logical coordinate axis of the cache space, thereby providing a unified reference system for cached data organization.

[0078] Based on the resource URL characteristics, the media resources reused by each pre-annotation task are aggregated, and a public resource cache node corresponding to the resource identifier is constructed. The pre-annotation inference results that are repeatedly hit are bound to the corresponding result cache node.

[0079] The resource URL features can be structured fields contained in the Uniform Resource Locator (URL) of media resources, such as path prefixes, query parameters, and hash values, which can be used as a basis for identifying identical or similar media resources and supporting the construction of public resource cache nodes. Public resource cache nodes can be shared cache entities formed by aggregating media resources shared by multiple pre-annotated tasks based on resource URL features. This can be used to eliminate duplicate storage of identical media resources and reduce IO and storage overhead. In this embodiment, public resource cache nodes can be obtained by parsing the resource URLs in each task request, extracting common features, deduplicating, and merging to generate unique cache nodes. Furthermore, public resource cache nodes can collaborate with result cache nodes to establish logical binding through associated edges, or collaborate with attribute tags of cache nodes to enable nodes to carry attributes such as access frequency and size. Result cache nodes can be cache entities specifically storing pre-annotated inference results and bound to specific media resources or task contexts. This can be used to support the accurate reuse of inference results and avoid duplicate calculations under the same input conditions. Based on resource URL characteristics, media resources reused by each pre-annotation task are aggregated to construct a common resource cache node. This can be achieved by parsing the URL to extract common fields and merging resources with the same characteristics into a single node, thereby eliminating redundant storage and improving media resource loading efficiency. Duplicate pre-annotation inference results are bound to corresponding result cache nodes. This can be done by matching existing result nodes with input conditions and model versions or creating new nodes and establishing binding relationships, enabling accurate reuse of inference results and reducing redundant calculations.

[0080] The access frequency and size of media resources, along with the model version and task type of the pre-labeled inference results, are recorded as attribute tags for cache nodes; the cache expiration time of media resources and pre-labeled inference results is used as the expiration tag for the current cache node and marked to the corresponding cache node.

[0081] The model version can be an identifier of the algorithm model used to generate the pre-annotated inference results, including version number, hash, or configuration digest. It can be used to ensure the semantic correctness of inference result reuse and prevent misuse of cross-version results. The attribute markers of cache nodes can be a set of metadata attached to the cache node, describing its business semantics and physical characteristics. This can be used to enable cache management to go beyond a simple LRU strategy and support intelligent decision-making based on multi-dimensional attributes. The cache expiration time can be the maximum duration or absolute time point that a cache node is allowed to retain. It can be used to control the lifecycle of cached data and prevent outdated data from occupying space for extended periods. The expiration tag can be a metadata identifier that explicitly marks the cache expiration time on the cache node. This can be used to allow the expiration policy to be directly read and executed by the cache management system without additional queries. Recording the access frequency, size, and corresponding parameters of the model version and task type of the pre-annotated inference results as attribute markers for the cache node allows relevant metadata to be written into the node attribute fields during node creation or updates, thereby endowing the cache node with business semantics and supporting intelligent caching decisions. The cache expiration time of media resources and pre-annotated inference results is used as the expiration tag of the current cache node and marked to the corresponding cache node. The expiration time can be calculated according to the data type and strategy and written into the node metadata, so as to realize self-descriptive management of the cache lifecycle.

[0082] The cache nodes are analyzed based on the total cache capacity to identify the upper threshold of cache space occupancy and the lower threshold of cache eviction, thus forming a cache capacity constraint boundary.

[0083] The upper threshold for cache space occupancy can be the maximum proportion or absolute value of the total cache capacity that can be occupied by stored nodes. This can be used to prevent excessive cache expansion that could lead to excessive system memory or disk pressure. The lower threshold for cache eviction can be the target capacity proportion or absolute value that must be released when a cache cleanup operation is triggered. This can be used to avoid frequent eviction operations and provide buffer space to maintain cache stability. The cache capacity constraint boundary can be a safe operating range for cache capacity defined by the upper threshold for cache space occupancy and the lower threshold for cache eviction. This can be used to ensure that the cache system can efficiently utilize resources without causing performance degradation or crashes. In this embodiment, the cache capacity constraint boundary can be obtained by dynamically or statically setting the upper and lower threshold values ​​based on the total system resources and historical load fluctuations. Furthermore, the cache capacity constraint boundary can work in conjunction with cache nodes to constrain node admission and eviction, or it can work in conjunction with the cache management system as a hard constraint for the global scheduling strategy. Analyzing and identifying the upper threshold for cache space occupancy and the lower threshold for cache eviction based on the total cache capacity to form the cache capacity constraint boundary can be done by combining historical system resources and load data to set or dynamically adjust the upper and lower threshold values. For example, the upper threshold for cache space occupancy and the lower threshold for cache eviction can be identified by analyzing cache nodes based on the total cache capacity and forming a cache capacity constraint boundary. This can be achieved by setting a static boundary with a fixed percentage, dynamically adjusting the boundary by using a sliding window to count load peaks, or adjusting the boundary in real time based on OOM risk using a feedback control algorithm. This allows for defining a safe operating range for the cache and balancing utilization and stability.

[0084] Based on the basic cache space axis, cache nodes are arranged according to reuse priority and spatiotemporal order. Associative edges are constructed between cache nodes, and cache nodes are connected through these edges to build a two-layer cache structure. Cache attribute tags, expiration tags, and cache capacity constraint boundaries are bound to each corresponding cache node and associated edge, and mapped to the cache management system of the annotation platform to build a complete two-layer cache structure.

[0085] The associated edges can be logical relationship lines connecting cache nodes according to reuse priority and spatiotemporal order based on the basic cache space axis. These edges can be used to construct the topology between cache nodes and support traversal, dependency analysis, and batch operations. The cache management system can be a software module within the annotation platform responsible for uniformly managing the two-layer cache structure, executing cache strategies, and monitoring cache status. It can provide a global cache view and control interface and automate cache lifecycle management. Constructing associated edges between cache nodes according to reuse priority and spatiotemporal order based on the basic cache space axis can connect adjacent nodes along the space axis in descending priority order and spatiotemporal proximity, thereby constructing a cache topology and supporting efficient traversal and dependency management. Connecting cache nodes through associated edges to construct a two-layer cache structure can assemble all nodes and edges into a graph structure, distinguishing between resource and result layer subgraphs, thus forming a complete two-layer cache logical architecture. Binding cache attribute tags, expiration tags, and cache capacity constraint boundaries to the corresponding cache nodes and associated edges and mapping them to the annotation platform's cache management system can register nodes, edges, and their metadata into the cache management system's internal data structure, thereby achieving platform-level integration and automated management of the cache structure.

[0086] Taking a smart retail customer behavior analysis and annotation platform as an example, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be used to process a large number of store camera video pre-annotation tasks. The system aggregates video segments according to the store ID and camera number in the resource URL to build a public resource cache node. At the same time, it binds the pedestrian detection results generated by the YOLOv8 model to the result cache node and annotates the model version as v8.2 and the task type as in-store customer flow statistics. Based on historical data, the high-frequency access small-volume thumbnails and the low-frequency access original high-definition videos are respectively located at both ends of the multi-level mapping chain to form a basic cache space axis. The system sets the upper threshold of cache occupancy to 80% and the lower threshold of eviction to 60% to form a cache capacity constraint boundary. Nodes are connected along the space axis according to reuse priority to form an associated edge to build a two-layer cache structure. When the access volume of a store video surges due to promotional activities, its corresponding node is preferentially retained along the associated edge, and the expired label ensures that the old promotional video is automatically cleaned up after 7 days. All node attributes, expired labels and capacity boundaries are mapped to the cache management system to achieve global intelligent scheduling, which significantly reduces the overhead of duplicate video decoding and inference.

[0087] In one embodiment, the process of analyzing the constraint boundary between the execution efficiency and resource consumption of annotation tasks based on the hierarchical caching structure, combining the concurrent task scheduling data of the annotation platform and the annotationer's job priority parameters, outputting an asynchronous task dispatch scheduling window set, and generating a pre-annotation task intelligent optimization strategy includes: Obtain the set of cache nodes, the set of cache expiration tags, and the cache capacity constraint boundary in the hierarchical cache graph. Identify the cache node groups that have a risk of conflict between the cache capacity constraint boundary and the cache requirements of newly added pre-labeled tasks, and determine their corresponding cache eviction risk periods.

[0088] The hierarchical cache graph can be a global state model representing nodes, associated edges, and their attributes and constraints in a two-layer cache structure in graph form. It can provide topology awareness of the cache state and support conflict identification and spatial window calculation. In this embodiment, the hierarchical cache graph can synchronize the cache node set, associated edges, and metadata from the cache management system in real time to construct an in-memory graph representation. Furthermore, the hierarchical cache graph can be a container that collaborates with the cache node set and serves as a node, or it can collaborate with an adjustable cache spatial window set and produce a set of windows through analysis. The cache node set can be an aggregation of all virtual cache nodes in the hierarchical cache graph, including resource cache nodes and result cache nodes, and can be used as the basic set of operation objects for conflict identification and priority ranking. The cache expiration tag set can be an aggregation of all cache node expiration tags in the hierarchical cache graph and characterizes the lifecycle state of each node. It can be used to support the judgment of cache eviction risk periods and avoid mistakenly deleting high-value data that has not expired. The cache requirement for newly added pre-labeled tasks can be the estimated storage amount of media resources and inference results required by the pre-labeled tasks to be scheduled. It can be used as an input variable for conflict detection and to quantify the pressure of new tasks on the cache system. Conflict risk can be the possibility that the cache demand of newly added pre-labeled tasks exceeds the available margin of the current cache capacity constraint. It can be used to trigger proactive resource adjustment mechanisms and prevent tasks from failing or being delayed due to insufficient resources. Cache node groups can be subsets of cache nodes identified in the conflict risk assessment as potentially participating in eviction or adjustment. They can be used to narrow down the scope of subsequent sorting and marking operations and improve processing efficiency.

[0089] Cache eviction risk periods can be the time windows during which cache node groups are expected to undergo eviction operations due to approaching expiration tags or increased capacity pressure. This can be used to transform static capacity issues into dynamic risks in a spatiotemporal dimension and support precise space reclamation. For example, cache eviction risk periods can be obtained by combining the set of expired tags with historical eviction patterns to predict the high-probability periods for node reclamation. Obtaining the set of cache nodes, the set of expired cache tags, and the cache capacity constraint boundaries in the hierarchical cache graph can be achieved by pulling the latest graph snapshot from the cache management system and parsing the node, tag, and boundary data. This achieves the technical effect of establishing a global cache status view and providing complete input for subsequent analysis. Identifying cache node groups at risk of conflict between cache capacity constraint boundaries and newly added pre-labeled task cache requirements can be done by comparing the new requirements with boundary reserves and locating potentially affected subsets of nodes. In an exemplary embodiment, identifying cache node groups at risk of conflict between cache capacity constraints and the caching requirements of newly added pre-labeled tasks can be achieved by using a threshold comparison method to filter nodes with less available space than required, using a graph traversal algorithm to propagate the pressure impact along associated edges, or using a machine learning model to predict conflict probabilities and truncate high-risk nodes. This achieves the technical effect of proactively discovering potential resource bottlenecks and transforming passive response into proactive prevention. Determining the cache eviction risk period corresponding to a cache node group can be achieved by combining expired tags and historical eviction patterns to estimate the time window for node reclamation. This achieves the technical effect of spatializing capacity issues in time and supporting precise reclamation rather than blind eviction.

[0090] Based on the annotation worker's job priority parameters, the cache nodes are sorted, and the cache nodes that are in the cache eviction risk period and have low reuse priority are marked as cache groups that can be evicted. The releaseable space fragments in their cache space are extracted to construct an adjustable cache space window set.

[0091] The revocable cache group can be a subset of cache nodes with low reuse priority that are within the cache eviction risk period and marked as objects that can be safely reclaimed. This ensures that resource reclamation does not affect high-value services while simultaneously reducing the cost of erroneous evictions by utilizing time windows. In one specific embodiment, the revocable cache group can be obtained by sorting cache node groups according to the annotator's job priority parameter and selecting low-priority nodes within the risk period. Furthermore, the revocable cache group can work in conjunction with releasable space fragments and serve as a source of space fragments, or it can work in conjunction with a pre-annotation intelligent optimization scheme and serve as a core input element of the scheme. Releasable space fragments can be continuous or discrete storage units extracted from the revocable cache group that can be used by new tasks within a specific time period. This can be used to transform abstract eviction decisions into concrete available resource quantities and support the construction of scheduling windows. The adjustable cache space window set can be a time-series set composed of multiple releasable space fragments organized according to their available time periods, representing the elastic supply capability of cache resources. This can be used to provide a dynamic view of cache resources and enable the scheduler to allocate the correct space at the correct time. For example, the adjustable cache space window set can be obtained by aggregating all freeable space fragments and aligning them along the timeline to form a window sequence. Furthermore, the adjustable cache space window set can be in conjunction with a dynamically scheduled window set and serve as a candidate pool for the latter, or it can be in conjunction with a group of evictable caches, with the windows consisting of the space freed up by that group.

[0092] Sorting cache nodes based on the priority parameters of the labelers can be done by arranging the nodes in the cache node group in descending order of priority. This ensures that high-value business data is prioritized for retention when resources are scarce. Prioritizing cache nodes with lower reuse priority that are within the cache eviction risk period as eligible cache groups involves selecting nodes from the sorted nodes that are in the risk period and have a priority below a threshold for marking. This balances business value with resource reclamation and avoids inadvertently damaging critical data. Extracting releaseable space fragments from the cache space of eligible cache groups to construct an adjustable cache space window set involves traversing the eligible cache groups, extracting their occupied storage units, and assembling them into windows according to available time periods. This transforms eviction decisions into schedulable resource provisioning.

[0093] Based on the concurrent task scheduling data of the annotation platform, spatial segments that meet the task efficiency requirements in the adjustable cache space window set are identified, and a set of dynamic scheduling windows as candidates is output. The cache groups that can be eliminated, the spatial segments that can be optimized, and the set of dynamic scheduling windows are sent to the annotation platform scheduling center. By combining the cache groups that can be eliminated, the spatial segments that can be optimized, and the set of dynamic scheduling windows, a corresponding pre-annotation intelligent optimization scheme is generated.

[0094] The task efficiency requirement can be the minimum cache bandwidth, hit rate, or space continuity index required by the pre-labeled task to ensure response timeliness and throughput. This can be used as a filtering condition to ensure that the allocated cache space truly meets business performance requirements. The dynamic scheduling window set can be a subset of available resource windows selected from the adjustable cache space window set that simultaneously meets the task efficiency requirement and timing constraints. This can be used to provide precise resource licenses that can be directly used for task dispatch and avoid allocating invalid or inefficient space. In this embodiment, the dynamic scheduling window set can traverse the adjustable cache space window set and verify whether each window meets the task efficiency requirement, retaining qualified windows. Furthermore, the dynamic scheduling window set can be used in conjunction with the adjustable cache space window set as a subset thereof, or it can be used in conjunction with the pre-labeled intelligent optimization scheme as a direct basis for scheduling decisions. Optimizable space fragments can be specific storage units selected from the dynamic scheduling window set for actual task dispatch. These can be used as the implementation carrier of the intelligent optimization scheme and achieve precise matching of resources and tasks. The annotation platform scheduling center can be a central control module responsible for receiving optimization elements and executing task dispatch, resource allocation, and strategy distribution. It can be used to transform analysis results into actual scheduling actions and close the optimization process. The pre-annotation intelligent optimization scheme can be a structured strategy package that combines a set of expendable cache groups, optimizable spatial fragments, and dynamic scheduling windows to guide the scheduling of the current batch of tasks. It can be used to achieve multi-dimensional collaborative optimization of resources and tasks and improve overall throughput and response efficiency.

[0095] Identifying space segments that meet task efficiency requirements within an adjustable cache window set based on concurrent task scheduling data from the annotation platform can be achieved by verifying the bandwidth and continuity of each window against the task efficiency requirements. Furthermore, this process can be accomplished by using a rule engine to verify efficiency indicators item by item, using simulated execution to estimate the actual performance of the window, or using reinforcement learning to evaluate the matching degree between the window and the task. This ensures that allocated resources are truly effective and avoids ineffective scheduling. Outputting space segments that meet task efficiency requirements as a candidate dynamic scheduling window set can be achieved by collecting verified space segments into a dynamic scheduling window set, thus generating precise resource licenses that can be directly used for scheduling. Sending the expendable cache group, optimizable space segments, and dynamic scheduling window set to the annotation platform scheduling center can be achieved by serializing these three elements and pushing them to the scheduling center via an internal message bus. This decouples analysis and execution and supports asynchronous optimization loops. Combining eliminable cache groups, optimizable space segments, and dynamic scheduling window sets to generate corresponding pre-labeled intelligent optimization schemes can enable the scheduling center to integrate these three types of elements and generate a structured scheme that includes task resource mapping, elimination instructions, and window allocation. This can achieve the technical effect of forming an executable optimization strategy and driving the system to adaptively adjust.

[0096] For example, in the scenario of an industrial quality inspection video annotation platform, the asynchronous visual target automatic tracking pre-annotation method for data annotation platforms in this embodiment can be as follows: When the platform suddenly receives a large number of high-precision defect detection pre-annotation requests, the system obtains a hierarchical cache map and finds that the cache demand for new tasks is approaching the capacity limit. It identifies that there is a risk of conflict in the cache node group storing old batches of normal product videos. Combined with the expired tags, it is determined that these nodes will enter the elimination risk period within the next ten minutes. According to the priority of the annotators' work, it is found that these nodes correspond to low-priority historical tasks and are marked as cache groups that can be eliminated. The two gigabytes of video memory space occupied by them are extracted to construct an adjustable cache space window set available for the next ten minutes. According to the current concurrent scheduling data, the segments with more than 500 megabytes of continuous space and the bandwidth meets the standard are selected as a dynamic scheduling window set. The cache groups that can be eliminated, the space segments that can be optimized, and the dynamic scheduling window set are sent to the scheduling center. The scheduling center generates an intelligent optimization plan to immediately reclaim the low-priority old video cache and allocate the new requests to the 500 megabyte window and delay non-urgent tasks until the window ends, thereby smoothly coping with sudden loads without interrupting high-priority quality inspection tasks.

[0097] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides an asynchronous visual target automatic tracking and pre-annotation system for data annotation platforms, the system comprising: The scene modeling module 10 is used to acquire historical annotation task data from the data annotation platform, identify typical annotation scenarios, construct annotation scenario models, extract task feature parameters for target tracking pre-annotation under typical annotation scenarios, and generate corresponding annotation efficiency impact labels. The hierarchical processing module 20 is used to take the labeling efficiency-affecting labels and the real-time tracking pre-labeling request data initiated by the labelers as input, process them based on the hierarchical task processing mechanism, and output the target tracking pre-labeling results and asynchronous task status records. The dual-layer cache module 30 is used to set up a layered cache architecture, mapping the media resources and inference results of historical tracking and pre-annotation reuse to the layered cache architecture respectively, and constructing a dual-layer cache structure of resource cache layer and result reuse layer; The intelligent scheduling module 40 is used to analyze the constraint boundary between the execution efficiency and resource consumption of annotation tasks based on the hierarchical caching structure. It combines the concurrent task scheduling data of the annotation platform with the priority parameters of the annotators' jobs to output a set of asynchronous task dispatch scheduling windows and generate intelligent optimization strategies for pre-annotation tasks.

[0098] Other embodiments or specific implementations of the asynchronous visual target automatic tracking and pre-annotation system for data annotation platforms described in this invention can be referred to the above-described method embodiments, and will not be repeated here.

[0099] Furthermore, to achieve the above objectives, the present invention also provides an asynchronous visual target automatic tracking pre-annotation device for a data annotation platform. The device includes: a memory, a processor, and an asynchronous visual target automatic tracking pre-annotation program for a data annotation platform stored in the memory and executable on the processor. The asynchronous visual target automatic tracking pre-annotation program for a data annotation platform is configured to implement the steps of the asynchronous visual target automatic tracking pre-annotation method for a data annotation platform as described in any of the above descriptions.

[0100] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an asynchronous visual target automatic tracking pre-annotation program for a data annotation platform, wherein when the asynchronous visual target automatic tracking pre-annotation program for a data annotation platform is executed by a processor, it implements the steps of the asynchronous visual target automatic tracking pre-annotation method for a data annotation platform as described above.

[0101] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. An asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms, characterized in that, The method includes: Acquire historical annotation task data from the data annotation platform, identify typical annotation scenarios, construct annotation scenario models, extract task feature parameters for target tracking pre-annotation under typical annotation scenarios, and generate corresponding annotation efficiency impact labels; The annotation efficiency affects the labels, and the real-time tracking pre-annotation request data initiated by the annotators is taken as input. The data is processed based on the hierarchical task processing mechanism, and the target tracking pre-annotation results and asynchronous task status records are output. A hierarchical caching architecture is set up to map the media resources and inference results that are pre-annotated and reused in historical tracking to the hierarchical caching architecture, and to build a two-layer caching structure of resource caching layer and result reuse layer. Based on the analysis of the hierarchical caching structure, the constraint boundary between the execution efficiency and resource consumption of annotation tasks is analyzed. Combining the concurrent task scheduling data of the annotation platform and the priority parameters of the annotators' jobs, an asynchronous task dispatch scheduling window set is output to generate a pre-annotation task intelligent optimization strategy.

2. The asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms as described in claim 1, characterized in that, The process of acquiring historical annotation task data from the data annotation platform, identifying typical annotation scenarios, and constructing annotation scenario models includes: By acquiring historical annotation task data generated by the data annotation platform during the historical annotation process, the historical annotation task data is constructed into an annotation task data matrix according to the annotation task type, resource size and submission time order; The annotation task data matrix is ​​divided into blocks based on the cache capacity hierarchical partitioning mechanism to obtain multiple historical annotation task units, and the media resource fragments and task attribute fragments corresponding to each historical annotation task unit are extracted. For each task attribute fragment, multidimensional task features are extracted. The multidimensional task features extracted from the same task attribute fragment are combined to form a task feature vector. A set of task feature vectors is constructed using the task feature vectors of all task attribute fragments. Cluster analysis is performed on the task feature vector set to identify representative typical annotation scenarios, including static image sequence annotation, VLA video annotation, and multi-camera cross-frame annotation. Based on the multi-dimensional task features extracted from each typical annotation scenario, an annotation scenario model is constructed.

3. The asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms as described in claim 2, characterized in that, The process of extracting task feature parameters for pre-annotation of target tracking in typical annotation scenarios and generating corresponding labels affecting annotation efficiency includes: For the various typical annotation scenarios included in the annotation scenario model, based on the statistically representative multidimensional task features in the task feature vector set, task feature parameters are extracted as descriptive terms affecting annotation efficiency. Each type of annotation efficiency impact description is input into a preset cache benefit rule set. The cache benefit rule set is constructed based on the resource download time, storage consumption, and annotation acceleration effect samples of historical pre-annotated tasks through correlation analysis or logistic regression inference, and outputs the cache benefit intensity coefficient range and its resource reuse priority. By identifying the cache benefit intensity range for various typical annotation scenarios, resource reuse priority is bound to the corresponding typical annotation scenarios, and labels affecting annotation efficiency are generated.

4. The asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms as described in claim 3, characterized in that, The process of taking the labeling efficiency-affected labels and the real-time tracking pre-labeling request data initiated by the labelers as input, processing them based on a hierarchical task processing mechanism, and outputting the target tracking pre-labeling results and asynchronous task status records is as follows: The parameter verification and spatiotemporal registration processing are performed on the real-time tracking pre-annotation requests initiated by the annotation front-end. The real-time request data includes reference box coordinates, target tracking identifiers, camera identifiers, frame range, media resource information, and annotation task identifiers. Based on the hierarchical task processing mechanism, real-time request data is decoupled from business and algorithm layers to form hierarchical task processing segments; each task processing segment is associated and bound with the corresponding labeled efficiency impact tags to construct an asynchronous task sample set. Based on the asynchronous task sample set and each type of hierarchical processing link, multiple task processing segments are fused and analyzed in the order of business reception-state storage-algorithm call-result write-back. The flow characteristics of task parameters in each processing link are extracted, task state nodes are constructed, and they are associated with each other according to the order of processing links to form an asynchronous task state trajectory. The task state trajectory is then analyzed based on deep learning combined with multi-dimensional task features.

5. The asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms as described in claim 4, characterized in that, The process of analyzing the task state trajectory based on deep learning and multi-dimensional task features includes: Identify abnormal task flow intervals, extract the gradient of task feature parameters, queuing time, and resource hit threshold points for each interval, and extract the recoverable waiting time within the corresponding scheduling stage, which is recorded as the task safety period. Record the gradient, queuing time, threshold points, and task safety period as abnormal task data, and record the task flow trajectory corresponding to the abnormal task interval as the asynchronous task status trajectory. Finally, output the pre-annotated frame-level results of target tracking and the complete asynchronous task status record.

6. The asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms as described in claim 5, characterized in that, The process of setting up a hierarchical caching architecture, mapping historically tracked and pre-annotated reused media resources and inference results to the hierarchical caching architecture respectively, and constructing a two-layer caching structure of a resource caching layer and a result reuse layer includes: Media resources and pre-annotated inference results are associated with virtual cache nodes at different levels in the hierarchical caching architecture. Based on the reuse frequency and size of media resources and inference results, a multi-level mapping chain for pre-annotated cache is established, and the multi-level mapping chain is used as the basic cache space axis. Based on the resource URL characteristics, the media resources reused by each pre-annotation task are aggregated, and a public resource cache node corresponding to the resource identifier is constructed. The pre-annotation inference results that are repeatedly hit are bound to the corresponding result cache node. The access frequency and size of media resources, along with the model version and task type of the pre-labeled inference results, are recorded as attribute tags for cache nodes; the cache expiration time of media resources and pre-labeled inference results is used as the expiration tag for the current cache node and marked to the corresponding cache node. The cache nodes are analyzed based on the total cache capacity to identify the upper threshold of cache space occupancy and the lower threshold of cache eviction, and to form a cache capacity constraint boundary. Based on the basic cache space axis, cache nodes are arranged according to reuse priority and spatiotemporal order. Associative edges are constructed between cache nodes, and cache nodes are connected through these edges to build a two-layer cache structure. Cache attribute tags, expiration tags, and cache capacity constraint boundaries are bound to each corresponding cache node and associated edge, and mapped to the cache management system of the annotation platform to build a complete two-layer cache structure.

7. The asynchronous visual target automatic tracking and pre-annotation method for data annotation platforms as described in claim 6, characterized in that, The process of analyzing the constraint boundary between the execution efficiency and resource consumption of annotation tasks based on the hierarchical caching structure, combining the concurrent task scheduling data of the annotation platform and the annotationer's job priority parameters, outputting an asynchronous task dispatch scheduling window set, and generating a pre-annotation task intelligent optimization strategy includes: Obtain the set of cache nodes, the set of cache expiration tags, and the cache capacity constraint boundary in the hierarchical cache graph; identify the cache node group that has a risk of conflict between the cache capacity constraint boundary and the cache requirement of the newly added pre-labeled task; and determine the corresponding cache eviction risk period. Based on the annotation worker's job priority parameters, the cache nodes are sorted, and the cache nodes that are in the cache eviction risk period and have low reuse priority are marked as cache groups that can be evicted. The releaseable space fragments in their cache space are extracted to construct an adjustable cache space window set. Based on the concurrent task scheduling data of the annotation platform, spatial segments that meet the task efficiency requirements in the adjustable cache space window set are identified, and a set of dynamic scheduling windows as candidates is output. The cache groups that can be eliminated, the spatial segments that can be optimized, and the set of dynamic scheduling windows are sent to the annotation platform scheduling center. By combining the cache groups that can be eliminated, the spatial segments that can be optimized, and the set of dynamic scheduling windows, a corresponding pre-annotation intelligent optimization scheme is generated.

8. An asynchronous visual target automatic tracking and pre-annotation system for data annotation platforms, characterized in that, The system includes: The scenario modeling module is used to acquire historical annotation task data from the data annotation platform, identify typical annotation scenarios, construct annotation scenario models, extract task feature parameters for target tracking pre-annotation under typical annotation scenarios, and generate corresponding annotation efficiency impact labels. The hierarchical processing module takes the labels affecting annotation efficiency and the real-time tracking pre-annotation request data initiated by the annotators as input, processes them based on the hierarchical task processing mechanism, and outputs the target tracking pre-annotation results and asynchronous task status records. The dual-layer caching module is used to set up a hierarchical caching architecture, mapping the media resources and inference results of historical tracking and pre-annotation to the hierarchical caching architecture respectively, and constructing a dual-layer caching structure of resource caching layer and result reuse layer; The intelligent scheduling module is used to analyze the constraint boundary between the execution efficiency and resource consumption of annotation tasks based on the hierarchical caching structure. It combines the concurrent task scheduling data of the annotation platform with the priority parameters of the annotators' jobs to output a set of asynchronous task dispatch scheduling windows and generate intelligent optimization strategies for pre-annotation tasks.

9. An asynchronous visual target automatic tracking and pre-annotation device for a data annotation platform, characterized in that, The device includes: a memory, a processor, and an asynchronous visual target auto-tracking pre-annotation program for a data annotation platform stored in the memory and executable on the processor, the asynchronous visual target auto-tracking pre-annotation program for a data annotation platform being configured to implement the steps of the asynchronous visual target auto-tracking pre-annotation method for a data annotation platform as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an asynchronous visual target automatic tracking pre-annotation program for a data annotation platform, which, when executed by a processor, implements the steps of the asynchronous visual target automatic tracking pre-annotation method for a data annotation platform as described in any one of claims 1 to 7.

Images