An urban traffic video data intelligent computing and empowerment system based on a containerization framework

Abstract

The application provides a kind of urban traffic video data intelligent computing and enabling system based on containerization framework, it is related to intelligent transportation and computer vision technical field, including based on standard video transmission protocol and view library protocol, unified gathering access and hierarchical classification management to multi-source video data, build video data resource account book;Heterogeneous computing power resources are registered to computing power management pool, various AI algorithms are independently packaged as container image, according to the received AI computing task query video data resource account book, based on the resource attribute information obtained, determine the computing power scheduling priority and target AI algorithm, and then dynamically create or reuse the container instance carrying the target AI algorithm to process the video stream of the target video point;Based on resource attribute information, access control strategy is executed, and data services are published externally through standardized interfaces.The beneficial effect is to break through the barrier of multi-source video data;Realize the decoupling of computing power and algorithm;Guarantee cross-department data sharing safety controllable.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and computer vision technology, and in particular to an intelligent computing and empowerment system for urban traffic video data based on a containerized framework. Background Technology

[0002] With the continuous improvement of smart city construction and the level of precision in traffic management, video surveillance resources have become a core infrastructure for sensing urban traffic operation status and supporting emergency command and decision-making. Especially in complex business scenarios such as traffic management, port operations, security for large-scale sports events, and traffic control around commercial complexes, the demand for real-time access and intelligent analysis of cross-domain video data is becoming increasingly urgent.

[0003] Currently, urban traffic video data mainly originates from independently constructed monitoring systems by various departments. These systems typically use standard protocols such as GB / T 28181 for internal networking and the GA / T 1400 view library protocol to achieve basic structured data transmission. However, the existing urban traffic video application system still suffers from significant architectural bottlenecks, making it difficult to meet the demands of high-concurrency and intelligent processing of massive amounts of video data. First, video resources are severely isolated, and cross-departmental sharing processes are cumbersome. Due to inconsistent construction standards and closed data interfaces among different business entities such as transportation, ports, and venues, video data cannot flow effectively between heterogeneous systems. When cross-departmental access to video resources is required, it often relies on manual coordination and physical copying, failing to achieve real-time sharing based on business needs. The value of data is severely limited by administrative and technical barriers.

[0004] Secondly, video data utilization is low, and the ability to extract high-value information is insufficient. Existing systems mostly focus on the storage and passive playback of video signals, lacking a systematic AI algorithm scheduling mechanism. Core elements such as people, vehicles, and roads contained in massive amounts of video data have not been automatically and scalably extracted into structured data. At the same time, traditional computing resource management methods result in a high degree of coupling between computing resources and algorithm modules, making it difficult for algorithms to flexibly migrate between different hardware architectures. They cannot dynamically reuse computing resources according to specific business scenarios such as port loading and unloading, road congestion, and venue crowd flow, resulting in a large amount of expensive GPU / NPU computing power being wasted during idle periods, while facing computing bottlenecks during peak business periods.

[0005] Therefore, there is an urgent need for an intelligent computing and empowerment system for urban traffic video data based on a containerized framework. This system should unify the aggregation of multi-source videos through standardized protocols, dynamically schedule containerized AI algorithms based on ledger matching, and provide supporting hierarchical management and control of shared services to solve industry pain points such as video silos, coupled computing power and algorithms, and low efficiency in intelligent video processing. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides an intelligent computing and empowerment system for urban traffic video data based on a containerized framework, comprising: a multi-source aggregation and access module for unified aggregation and access of multi-source video data based on standard video transmission protocols and view library protocols; a video data full-element governance module for hierarchical and classified governance of aggregated videos, constructing a video data resource ledger containing video location resource attribute information; and a containerized video data intelligent computing module, comprising: a computing power management unit for uniformly registering heterogeneous computing power resources to a computing power management pool; an algorithm management unit for independently encapsulating various AI algorithms into container images; and a task orchestration unit for scheduling tasks according to received AI algorithms. The computing task queries the video data resource ledger to obtain the resource attribute information corresponding to the target video point pointed to by the AI ​​computing task. Then, based on the resource attribute information, it determines the computing power scheduling priority of the AI ​​computing task and the matching target AI algorithm. Then, based on the computing power scheduling priority and the real-time load status of the computing power management pool, it dynamically creates or reuses container instances that carry the target AI algorithm and distributes the video stream of the target video point to the container instances for processing. The multi-dimensional sharing and empowerment module is used to receive structured data output by the algorithm container instance and execute access control policies based on the resource attribute information to publish data services to the outside world through standardized interfaces.

[0007] Preferably, the video data full-element governance module includes a classification and hierarchical governance unit, which is used to classify and govern the aggregated video according to at least one of the following dimensions: location, business, application, and security. The corresponding video data resource ledger is constructed in a tree-like directory structure according to the dimensions of region, unit, attribute, or object.

[0008] Preferably, the video data full-element governance module further includes a data quality audit unit, which is used to perform full-link data quality audit on the structured data reported from the front-end device and output from the algorithm container. The audit includes integrity check, consistency check, standardization check and uniqueness check.

[0009] Preferably, the algorithm management unit registers AI algorithms from different algorithm providers to the algorithm management pool through a standardized algorithm interface, and configures corresponding container images for each AI algorithm. The standardized algorithm interface includes a management plane channel and a data plane channel. The management plane channel uses the REST over HTTP / 2 protocol for algorithm registration, computing power application, task status query, and configuration distribution. The data plane channel uses the gRPC bidirectional streaming protocol for transmitting video frame data to algorithm container instances and receiving structured analysis results returned by algorithm container instances. The algorithm management unit also pre-caches frequently used algorithm container images to each computing node to reduce container cold start latency.

[0010] Preferably, the resource attribute information includes location level, location dimension label, and business dimension label; determining the computing power scheduling priority and matching target AI algorithm based on the resource attribute information includes: when the location level is core level, allocating exclusive computing power resources and assigning the highest scheduling priority; when the location level is important level, allocating computing power resources using a time-sharing multiplexing mode; when the location level is general level, allocating computing power resources only when computing power resources are idle; when the location dimension label is port, matching container recognition or truck flow statistics algorithms; when the business dimension label is emergency command, increasing the video stream priority to the highest level.

[0011] Preferably, in the time-sharing multiplexing mode, the task orchestration unit allocates the same computing resources to multiple AI computing tasks based on a weighted priority time slice round-robin strategy, including: calculating a normalized weight according to the priority of each AI computing task; allocating a basic time slice for each AI computing task based on the normalized weight; executing each AI computing task sequentially according to the scheduling queue order; and moving unfinished tasks to the end of the queue to wait for the next round of scheduling after execution. When a high-priority task arrives and there are no available computing resources, the lowest priority task currently running is selected to execute the checkpoint and save it, releasing the computing resources it occupies and allocating them to the high-priority task. The preempted task is re-enqueued and marked as pending recovery.

[0012] Preferably, the access control strategy executed by the multi-shared empowerment module includes ultra-fine-grained authorization management. The dimensions of the authorization management include at least one of video device resource authorization, service resource authorization, service call quota authorization, call data volume quota authorization, and service call time and space range authorization, and each interface call generates a traceable audit log.

[0013] Preferably, it also includes a video data asset management module for asset identification and confirmation, asset valuation, asset transaction management, and asset lifecycle management of video data assets; wherein, the video data assets include video source data and structured data output by the containerized video data intelligent computing module; the asset valuation adopts a dual-track valuation model combining the market approach and the cost approach, and introduces technology correction coefficient, value density correction coefficient, date correction coefficient, and capacity correction coefficient for dynamic adjustment.

[0014] Preferably, it also includes a full-link security module, which provides at least one of the following security measures in each stage of video data access, aggregation, storage, governance, application and sharing: access control, security encryption, data anonymization, technical anti-counterfeiting, application logging and audit logs.

[0015] Preferably, the system is logically deployed in a three-tier architecture, and physically deployed in a distributed architecture consisting of one central platform, N departmental nodes, and M social nodes. The three-tier architecture includes: a bottom layer (infrastructure layer) deploying the multi-source aggregation access module and the video data full-element governance module; a middle layer (capability layer) deploying the containerized video data intelligent computing module; and a top layer (business application layer) deploying the multi-dimensional sharing and empowerment module. The distributed architecture includes: the central platform deploying the computing power management unit and the task orchestration unit to uniformly schedule computing resources across the entire domain; the N departmental nodes serving as regional computing power work nodes deploying local computing versions of the containerized video data intelligent computing module to execute video data computing tasks within their respective regions and report structured data to the central platform; and the M social nodes serving as temporary authorized access nodes, deploying only the video stream forwarding proxy of the multi-source aggregation access module to connect the video stream to the central platform or the nearest departmental node within the set video sharing time and space range.

[0016] The above technical solution has the following advantages or beneficial effects: 1) Breaking down barriers between multi-source video data and significantly improving cross-departmental video sharing capabilities: This invention sets up a multi-source aggregation access module, relying on a unified standard video transmission and view library protocol to complete standardized access for all types of video devices across the entire domain. Combined with a video data full-element governance module, it constructs a unified video resource ledger, breaking down data silos formed by the independent construction of video systems by different business entities. In actual testing scenarios, the response time for cross-departmental video access was reduced from 30 minutes of manual coordination to within 30 seconds, the video resource sharing coverage rate increased from less than 15% to 85%, and the development cycle for video interfacing between different units was significantly shortened, eliminating the need for separately customized heterogeneous device interface interfaces.

[0017] 2) Decoupling computing power from algorithms significantly improves computing power utilization and video structured extraction efficiency: This invention achieves unified management of the computing power pool and independent encapsulation of AI algorithm containers through a containerized video data intelligent computing module. It dynamically schedules and reuses algorithm containers based on video location resource ledgers to match business needs, completely solving the problems of deep binding between computing power and algorithms, resource idleness, or insufficient computing power in traditional solutions. After long-term testing with 5000 video streams, the overall utilization rate of heterogeneous computing power increased from less than 40% to over 75%. A single video stream can carry multiple AI algorithm analyses, including traffic flow, license plates, and events, with end-to-end intelligent structured processing latency controlled within 2 seconds. The standardized container image encapsulation method significantly shortens the new algorithm deployment cycle, improves AI access efficiency by approximately 80%, and enables automated, large-scale extraction of human, vehicle, and road elements from massive amounts of video, fully exploring the inherent data value of the video.

[0018] 3) Refined access control ensures secure and controllable cross-departmental data sharing: This invention's multi-dimensional sharing empowerment module matches corresponding access control policies based on video location resource attributes, achieving fine-grained authorization across multiple dimensions such as device, call frequency, data limit, and usage time. All data service calls retain complete audit logs, enabling full-chain traceability of security incidents. Compared to traditional coarse-grained departmental-level permission control, this invention refines security control to the dimensions of single user, single device, and single call, mitigating the risks of data leakage and unauthorized access while enabling cross-departmental data sharing capabilities.

[0019] 4) Reduce redundant investment in hardware and software, resulting in significant economic benefits: This invention relies on a unified heterogeneous computing power pool to achieve full-domain computing power sharing, eliminating the need for each business department to build its own AI computing power cluster. Hardware procurement investment can be reduced by 40% to 50%, and annual computing power operation and maintenance costs can be reduced by 35%. Taking the port scenario as an example, by relying on automated video OCR container recognition to replace manual container inspection, millions of yuan in labor costs can be saved annually, cargo hoisting and scheduling efficiency can be improved by 25%, and the overall operational benefits in large-scale urban transportation deployment scenarios can be significantly improved. Attached Figure Description

[0020] Figure 1 A schematic diagram of the structure of an intelligent computing and empowerment system for urban traffic video data based on a containerized framework, as a preferred embodiment of the present invention; Figure 2 A preferred embodiment of the present invention is a flowchart of containerized video computing task orchestration. Figure 3 A flowchart of a video data asset valuation algorithm is shown in a preferred embodiment of the present invention. Figure 4 A preferred embodiment of the present invention is shown in the end-to-end security architecture diagram. Figure 5 In a preferred embodiment of the present invention, a distributed deployment topology diagram of a "1+N+M" visual center is shown. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The present invention is not limited to this embodiment; other embodiments that conform to the spirit of the present invention may also fall within the scope of the present invention.

[0022] In a preferred embodiment of the present invention, in response to the pain points of existing urban traffic video systems, such as video resource silos, deep coupling of computing power and algorithms, and difficulty in extracting high-value structured information from massive amounts of video at scale, an intelligent computing and empowerment system for urban traffic video data based on a containerized framework is provided. The overall software logic is divided into a three-layer software architecture: IaaS infrastructure layer 100, PaaS capability layer 200, and SaaS business application layer 300. Each functional module is fixedly deployed at the corresponding layer, and physically it adopts a distributed topology of 1 central platform + N departmental nodes + M social nodes.

[0023] Specifically, such as Figure 1 As shown, the system's bottom layer is the IaaS infrastructure layer 100, which carries hardware access and raw video transmission-related functions, and only deploys the multi-source aggregation access module 1. The middle PaaS capability layer 200 is the system's core processing layer, carrying all core logic for video governance, containerized AI computing, data sharing, asset management, and security protection, and sequentially deploys the video data full-element governance module 2, the containerized video data intelligent computing module 3, the multi-source sharing empowerment module 4, and the video data asset management module 5. The full-link security module 6 spans the PaaS capability layer 200 and the SaaS business application layer 300, providing basic security services (access control, security encryption, data anonymization, and anti-counterfeiting technology) in the PaaS capability layer 200, and providing a security management interface (application traceability and audit log visualization management) in the SaaS business application layer 300. The top-level SaaS business application layer 300 is the business presentation layer, which, apart from carrying the security management interface of the full-link security module 6, has no independent functional modules and only calls the PaaS through standardized interfaces. The capability layer's multi-faceted sharing and empowerment module 4 acquires structured video data to support upper-layer business applications such as traffic command, port scheduling, and event passenger flow.

[0024] The data interaction flow between the layers of this three-tier software architecture is as follows: In the data uplink flow (IaaS Infrastructure Layer 100 → PaaS Capability Layer 200 → SaaS Business Application Layer 300), the IaaS Infrastructure Layer 100 video stream access gateway receives the raw video stream (H.264 / H.265 encoded) and pushes it to the PaaS Capability Layer 200 via the RTSP / GB / T28181 protocol. The PaaS Capability Layer 200 receives the video stream, completes device file matching, hierarchical classification and labeling, and writes the metadata to the database. The PaaS Capability Layer 200 subscribes to the list of video streams that need to be computed from the video asset center and creates an algorithm container Pod through the K8s scheduler. The algorithm container pulls the video stream (gRPC bidirectional stream) from the IaaS layer, performs AI inference computation, and outputs structured data (JSON format). The structured data is written to the capability layer data lake (based on Apache Iceberg table format) and pushed to the Kafka message queue. The business applications of the PaaS Capability Layer 200 subscribe to the structured data stream through the REST API provided by the PaaS Capability Layer 200 and display the business results.

[0025] In the data downlink flow (SaaS Business Application Layer 300 → PaaS Capability Layer 200 → IaaS Infrastructure Layer 100), the SaaS Business Application Layer 300 user initiates a video access request (specifying device ID and time range) through the business application interface; the request is authenticated by the PaaS Capability Layer 200 (checking whether the user has access permissions for the device); after successful authentication, the PaaS Capability Layer 200 queries the device access information (streaming media server address, device code); the PaaS Capability Layer 200 sends a stream retrieval command to the streaming media server in the IaaS Infrastructure Layer 100 (GB / T 28181 Invite message); the IaaS Infrastructure Layer 100 returns the video stream URL (RTSP / WebRTC address), and the PaaS Capability Layer 200 returns the URL to the SaaS Business Application Layer 300; the SaaS Business Application Layer 300 player directly pulls the video stream from the IaaS Infrastructure Layer 100, reducing the forwarding pressure on the PaaS Capability Layer 200.

[0026] More specifically, the multi-source aggregation access module 1, deployed in the IaaS infrastructure layer 100, is used to perform unified aggregation access of multi-source video data based on standard video transmission protocols and view library protocols.

[0027] The standard video transmission protocol includes GB / T 28181, and the view library protocol includes GA / T 1400. GB / T 28181 is used to transmit raw video streams, including H.264 / H.265 encoded real-time / historical video streams and PTZ control commands, typically used for real-time video preview, historical playback, and PTZ control. GA / T 1400 is used to transmit structured data, including view segments and structured attributes (JSON format) of faces, bodies, motor vehicles, and non-motor vehicles, typically used for image search, license plate recognition result query, and trajectory analysis.

[0028] This invention preferably employs a protocol separation, dual-channel parallel processing mode: two protocols are selected separately on the data source side according to the device capabilities, and bandwidth contention is avoided at the transport layer through VLAN isolation. The two protocols work in parallel without conflict. The multi-source video data includes at least two of the following: fixed camera video, mobile law enforcement recorder video, vehicle-mounted camera video, and drone video. During the device registration phase, the system uses a GB / T 28181 device catalog query (Catalog command) to detect whether the device supports the GA / T 1400 protocol: intelligent devices (supporting front-end intelligent analysis) simultaneously enable both protocols, transmitting two data streams in parallel; ordinary devices (supporting only video encoding) only enable GB / T 28181, with structured data generated by AI calculations on the platform side.

[0029] During multi-source aggregation access, when the same device simultaneously provides GB / T 28181 video streams and GA / T 1400 structured data, the system also performs the following data consistency checks: timestamp alignment check—the timestamp of the structured data must be within ±500ms of the timestamp of the corresponding video frame, otherwise it is marked as “timing abnormal”; spatial coordinate check—the target position coordinates in the structured data must be within the video resolution range, otherwise it is marked as “coordinate out of bounds”. For conflict scenarios where the calculation results from the front-end device and the platform side are inconsistent, the system performs the following processing: Scenario 1: The device sends video streams and structured data simultaneously, but the content is inconsistent—for example, the front-end device reports "3 vehicles detected" through GA / T 1400, but the platform-side AI algorithm identifies "5 vehicles" after analyzing the video stream. At this time, the system triggers the data quality audit process, saves both sets of data to the data lake and marks their confidence levels—the confidence level of the front-end data is set to 0.7 by default (which can be adjusted according to the historical accuracy of the device manufacturer), and the confidence level of the platform-side calculated data is set to 0.9 (based on the historical score of the algorithm evaluation module). Finally, the platform-side data with higher confidence level is used as the output result, and the front-end data is saved as the "alternative result" field; Scenario 2: Network congestion causes the GB / T 28181 video stream to be stuttered, but the GA / T 1400 structured data is normal—at this time, the system automatically reduces the video stream resolution (from 1080P to 720P) or frame rate (from 25fps to 10fps) to prioritize the smooth flow of the structured data channel, and at the same time triggers an alarm to notify the operation and maintenance personnel to check the network. If the GB / T 28181 connection fails, the system will automatically try alternative protocols such as RTSP / RTMP, retrying a maximum of 3 times with a 10-second interval; if the GA / T 1400 connection fails, the system will automatically switch to platform-side computing mode and pull video streams through GB / T 28181 to generate structured data.

[0030] The video data full-element governance module 2, deployed in the PaaS capability layer 200, is used to perform hierarchical and classified governance of aggregated videos and build a video data resource ledger containing video point resource attribute information.

[0031] In a preferred embodiment of the present invention, the video data full-element governance module 2 includes a classification and hierarchical governance unit 21, which is used to classify and govern the aggregated video according to at least one of the following dimensions: location (sports stadiums, ports, roads, etc.), business (traffic management, emergency command, etc.), application, and security.

[0032] Specifically, the system classifies devices into three levels: core (Level 1), important (Level 2), and general (Level 3). The classification directly affects subsequent processing strategies: core-level devices are given priority in GPU resource allocation, support parallel computing of multiple algorithms, store all data for 90 days with permanent storage during critical periods, and perform hourly snapshot quality checks, suitable for main road intersections and accident-prone sections; important-level devices are allocated computing power on demand, support time-sharing reuse mode, store all data for 30 days, and perform quality checks every 4 hours, suitable for secondary road intersections and commercial areas; general-level devices are processed only when computing power is idle, store data for only 7 days or on demand, and perform daily quality checks, suitable for side roads and community roads.

[0033] The category label directly determines the algorithm calling strategy: when the location dimension is a port, the container recognition algorithm and the truck flow statistics algorithm are automatically loaded; when the business dimension is emergency command, the video stream is automatically prioritized to the highest level and pushed to the emergency command center in real time; when the location dimension is a sports venue and the business dimension is passenger flow management, the high-density crowd counting algorithm and the abnormal gathering detection algorithm are automatically loaded.

[0034] The corresponding video data resource ledger is structured into a tree-like directory by region, unit, attribute, or object, laying the foundation for subsequent data exchange. The video data full-element governance module 2 receives the standardized video stream output from the IaaS layer multi-source aggregation access module 1, completes the governance, and then sends it upward to the containerized video data intelligent computing module 3 in the capability layer for AI analysis.

[0035] In a preferred embodiment of the present invention, the video data full-element governance module 2 further includes a data quality audit unit 22, which is used to perform full-link data quality audit on the structured data reported from the front-end device and output from the algorithm container. The audit includes integrity check, consistency check, standardization check and uniqueness check.

[0036] The integrity check verifies whether the required fields of the structured data in the view are completely filled and whether there are any missing data records: the field-level check requires that the vehicle detection results must include six core fields, such as license plate number, vehicle type, color, timestamp, device ID, and coordinate location. If any field is empty, it is judged as an incomplete field; the record-level check judges the record missing rate by comparing the video duration with the number of structured data records. The qualified standard is that the field integrity rate is ≥95% and the record missing rate is ≤10%.

[0037] Consistency checks are used to verify whether the data of the same data object is consistent across different data sources or at different times: cross-source consistency requires that the license plate number of the same vehicle must be consistent in the detection results of two adjacent intersections, and the time difference between passing must conform to the travel time of the road segment; temporal consistency requires that the timestamps of data collected continuously by the same device must monotonically increase; logical consistency requires that vehicle attributes are not allowed to change abruptly in the same trajectory, and the passing standard is a consistency conflict rate ≤ 5%.

[0038] The normative inspection is used to verify whether the data format, encoding rules, and value range comply with standards and specifications such as GA / T 1400: the license plate number must comply with the Chinese license plate format, the timestamp must comply with the ISO 8601 format, and the vehicle speed must be within a reasonable range of 0-180km / h. The passing standard is a normative violation rate of ≤3%.

[0039] Uniqueness checks are used to verify whether there are duplicate records in the data: the primary key of each structured data must be unique in the database, the same vehicle should not be checked repeatedly on the same device at the same time, and the pass standard is a duplicate record rate of ≤2%.

[0040] In a further optimized configuration, the data quality audit unit 22 performs data scans at different frequencies based on different equipment levels. Upon discovering quality issues, the system automatically attempts to repair them: for integrity issues, it attempts to infer missing fields from historical data; for consistency issues, it triggers a manual review process; for standardization issues, it automatically converts the format or marks them as needing correction; for uniqueness issues, it automatically removes duplicates and retains high-confidence records. For issues that cannot be automatically repaired, a quality report is generated and pushed to data governance personnel. The repaired data is rewritten into the data lake, and the original dirty data is moved to the quality issue data archive and stored for 30 days.

[0041] The containerized video data intelligent computing module 3, also deployed in the PaaS capability layer 200, includes a computing power management unit 31, an algorithm management unit 32, and a task orchestration unit 33. For example... Figure 2 The diagram shows the orchestration process of containerized video computing tasks, which fully demonstrates the internal data flow and computing power scheduling of the PaaS capability layer 200 containerized video data intelligent computing module 3. The left side of the diagram is the platform-side scheduling entry, the right side is the K8s computing power cluster execution end, and the middle is the algorithm image repository support link. Figure 2The starting point on the left platform side is the video data access interface. After receiving the GB / T 28181 / RTSP video stream, it enters the task request parsing stage, parsing the computational requirements and parameters (reading the point resource attributes in the video data resource ledger), and then enters the algorithm matching and judgment branch. In the algorithm selection stage, the system uses the "algorithm supermarket" (i.e., the visual service catalog of the algorithm management pool) to classify and display registered AI algorithms by vertical field, sub-scenario, and source vendor, supporting multi-dimensional queries such as algorithm identifier code, name, version number, input / output specifications, and resource requirements. The algorithm matching result queries the "algorithm metadata database" through the REST / gRPC interface to obtain the list of available algorithms and the corresponding algorithm image address, and then pulls the target algorithm Docker image from the Harbor image repository through the gRPC protocol. After successful matching, a computing power allocation request is simultaneously initiated to the computing power management unit 31. The computing power management unit 31 reads the real-time load of the global computing power pool, executes resource scheduling decisions (GPU / CPU / memory allocation), generates a task orchestration plan (DAG dependency graph construction), and enters the process. Figure 2 The K8s cluster execution logic on the right first completes the containerized task creation (Docker image pulling) and K8s Pod scheduling (NodeSelector / affinity), and then starts the time-sharing multiplexing scheduling logic (GPU time-sharing multiplexing / MIG splitting). During the scheduling process, computing power load indicators are collected in real time through Prometheus and updated to the platform side every 15 seconds. After the task is completed, the task status is determined: if it is not completed, the checkpoint is saved and it is put back into the scheduling queue; if all tasks are completed, GPU / CPU computing power resources are reclaimed. The AI ​​computing output structured data (JSON / Protobuf format) is written down to the video asset ledger database through gRPC bidirectional data flow.

[0042] The above Figure 2 The overall presentation of the collaborative scheduling process of each unit in the containerized video data intelligent computing module 3 is presented. The specific implementations of the computing power management unit 31, the algorithm management unit 32, and the task orchestration unit 33 are explained below.

[0043] The computing power management unit 31 is used to uniformly register heterogeneous computing power resources to the computing power management pool. These heterogeneous computing power resources include at least two of the following: cloud GPU clusters (NVIDIA A100 / V100 / T4), edge computing nodes (Jetson Orin / Atlas 300), and domestically produced NPUs (Ascend 910 / 310). The computing power management pool maintains the following core attributes for each type of computing power resource: computing power identifier (unique resource ID, computing power type, manufacturer, physical location), computing power capacity (total computing power, available computing power, VRAM / RAM capacity), operating status (real-time temperature, power consumption, utilization percentage, fault status), and task queue (current number of queued tasks, estimated waiting time). The computing power management unit 31 uses Prometheus + Grafana to achieve real-time dynamic monitoring of the operating status and consumption of computing power resources, with a data collection granularity of once every 15 seconds.

[0044] The algorithm management unit 32 is used to independently package various AI algorithms into container images. This invention uses Docker+Kubernetes (K8s) as the containerization infrastructure: the container runtime uses Docker Engine 19.03 or later, with containerd as the underlying runtime. Each AI algorithm is packaged into an independent Docker image, which contains algorithm inference code, model weight files, dependent runtime environments (Python / C++ Runtime), and standardized input / output interface adaptation layers. The container orchestration layer uses Kubernetes v1.26+ as the container orchestration engine to manage the container lifecycle across nodes (central cloud + edge nodes). The central platform deploys K8s Master nodes, and the visual hub platforms and edge nodes of various traffic management departments deploy K8s Worker nodes to form a unified managed container cluster. A Harbor private image repository (v2.0+) is built to provide version management, security scanning, image signing, and access control for algorithm images. Istio is used as the service mesh layer to realize traffic management, load balancing, circuit breaking and degradation, and observability of inter-service communication.

[0045] In a preferred embodiment of the present invention, the algorithm management unit 32 registers AI algorithms from different algorithm providers to the algorithm management pool through a standardized algorithm interface, and configures corresponding container images for each AI algorithm.

[0046] Specifically, the algorithms are categorized and managed according to vertical fields (transportation, security, industry, etc.), sub-scenarios (traffic flow analysis, crowd density, container recognition, license plate recognition, abnormal behavior detection, etc.), and source manufacturers. The algorithms and computing power are deeply decoupled, and the algorithm modules can be updated and iterated independently without affecting the underlying computing power scheduling.

[0047] The standardized algorithm interface includes a management plane channel and a data plane channel. The management plane channel uses the REST over HTTP / 2 protocol (JSON format) for algorithm registration, computing power application, task status query and configuration distribution. The data plane channel uses the gRPC bidirectional streaming protocol (Protocol Buffers format) for transmitting video frame data to the algorithm container instance and receiving structured analysis results returned by the algorithm container instance.

[0048] Furthermore, the key interfaces of the management plane channel include: an algorithm registration interface (algorithm vendors submit algorithm identifiers, names, version numbers, input / output specifications, resource requirements, container image addresses, etc., and the platform returns a globally unique identifier), a computing power application interface (business systems submit task identifiers, algorithm identifiers, video stream addresses, scheduling modes, priorities, resource requirements, etc., and the platform returns container instance IDs and running node IPs), a task status query interface (returns the current task status, execution progress percentage, and real-time processing speed), and a computation result callback interface (algorithm containers actively push structured results back to the platform). The gRPC bidirectional streaming service of the data plane channel establishes a full-duplex data pipeline between the platform and the algorithm containers: in the uplink stream, the platform continuously sends video frame data packets containing frame sequence numbers, timestamps, image data, encoding formats, resolutions, and camera identifiers to the algorithm containers; in the downlink stream, the algorithm containers continuously send analysis results containing corresponding frame sequence numbers, timestamps, lists of detected target objects, algorithm identifiers, and confidence scores to the platform. This invention uses Protocol Buffers as the data serialization format. The binary encoding has a transmission efficiency that is about 60% higher than that of the JSON text format, and the strong type constraints avoid runtime field parsing errors.

[0049] In addition, the algorithm management unit 32 is also used to pre-cache frequently used algorithm container images to each computing node to reduce container cold start latency. According to actual tests, the container cold start latency is ≤3 seconds (optimized by image layered caching and node pre-pull mechanism), the container hot start latency is ≤500 milliseconds, the maximum number of concurrent computing tasks is ≥5000 / central platform, and the maximum number of concurrent tasks per edge node is ≥200.

[0050] In a preferred embodiment of the present invention, the resource attribute information includes location level, location dimension label, and business dimension label; determining the computing power scheduling priority and matching the target AI algorithm based on the resource attribute information includes: when the location level is core level, allocating exclusive computing power resources and assigning the highest scheduling priority; when the location level is important level, allocating computing power resources using a time-sharing multiplexing mode; when the location level is general level, allocating computing power resources only when computing power resources are idle; when the location dimension label is port, matching container recognition or truck flow statistics algorithms; when the business dimension label is emergency command, increasing the video stream priority to the highest level.

[0051] Based on this, the task orchestration unit 33 queries the video data resource ledger according to the received AI computing task, obtains the resource attribute information corresponding to the target video point pointed to by the AI ​​computing task, then determines the computing power scheduling priority and the matching target AI algorithm of the AI ​​computing task based on the resource attribute information, and then dynamically creates or reuses container instances that carry the target AI algorithm based on the computing power scheduling priority and the real-time load status of the computing power management pool, and distributes the video stream of the target video point to the container instance for processing.

[0052] Specifically, the system maintains three core scheduling queues: a task scheduling queue to store instances of computing tasks to be executed, each task carrying a task ID, algorithm ID, priority (1-10, 10 being the highest), target video stream URL, scheduling mode flag, resource requirement specifications, and maximum execution time; a computing power resource pool to maintain all available computing power nodes and their current load status; and an algorithm container image cache to pre-cache container images of frequently used algorithms on each computing node to reduce cold start latency.

[0053] AI computing task scheduling modes include at least two of the following: real-time computing mode, time-sharing multiplexing mode, and offline batch processing mode. In real-time computing mode, tasks are preemptively scheduled with the highest priority (9-10), exclusively utilizing computing resources. This is suitable for business scenarios requiring real-time alerts, such as real-time congestion detection and abnormal event warnings. In near-real-time batch processing mode, tasks are non-preemptively scheduled with medium priority (5-8), with shared resource pool allocation. This is suitable for near-real-time analysis scenarios such as traffic flow statistics and passenger flow density analysis. In time-sharing multiplexing mode, multiple tasks share the same computing resources, using time-slice round-robin scheduling based on weighted priority. This is suitable for scenarios where multiple analysis algorithms run sequentially on the same video stream. In offline batch processing mode, tasks are best-efforts scheduled with the lowest priority (1-4), utilizing idle computing power to perform offline batch processing tasks such as historical traffic data modeling and feature pattern mining.

[0054] In a preferred embodiment of the present invention, under time-sharing multiplexing mode, task orchestration unit 33 allocates the same computing resources to multiple AI computing tasks based on a weighted priority time slice round-robin (WPRR) strategy, including: calculating normalized weights based on the priority of each AI computing task, allocating a base time slice for each AI computing task based on the normalized weights, executing each AI computing task sequentially according to the scheduling queue order, and moving unfinished tasks to the end of the queue to wait for the next round of scheduling after execution. The algorithm flow is as follows: Step 1, queue initialization: arrange all tasks to be scheduled in descending order of priority, and arrange tasks with the same priority in ascending order of waiting time to form a scheduling queue; Step 2, base time slice calculation: calculate the base time slice length q_base for each task. i= Q_total × X i (Q_total is the total scheduling period, defaulting to 60 seconds, X) i (for normalized weights), if q_base i If the time slice is less than the minimum time slice (default 5 seconds), the minimum time slice is used; Step 3, Preemption Check: During the scheduling cycle, the arrival events of high-priority tasks (p≥9) are monitored in real time. If a high-priority task arrives and there are no available computing resources, the lowest priority task currently running is selected to perform a checkpoint saving, releasing the computing resources it occupies and allocating them to the high-priority task. The preempted task is re-enqueued and marked as pending recovery; Step 4, Time Slice Round Robin Execution: Time slices are allocated to each task in the order of the scheduling queue: computing resources are allocated to the target algorithm and container instances are created / reused. The execution time of the algorithm computation task is q_base. i If a task is completed within its time slice, resources are released and the task is marked as completed. If a task is not completed and reaches the maximum continuous execution time (default 300 seconds), a checkpoint is executed to save the task, and the task is moved to the end of the queue. If a task is not completed but has not reached the maximum execution time, a checkpoint is executed to save the task, and the task is moved to the end of the queue to wait for the next round of scheduling. Step 4 is repeated until all tasks are completed or all computing resources are reclaimed. Step 5: Dynamic weight adjustment: After each complete scheduling cycle, the weights of each task are dynamically adjusted according to the task waiting time decay factor a = 1 + ln(1 + waitTime / avgWaitTime) and the task completion urgency factor b = remainingTime / totalAllocatedTime. The adjusted weight X' i = X i ×a×b, after being renormalized, is applied to the next scheduling cycle.

[0055] When a high-priority task arrives and there are no available computing resources, the lowest priority task currently running is selected to perform a checkpoint save, releasing the computing resources it occupies and allocating them to the high-priority task. The preempted task is then re-enqueued and marked as pending recovery.

[0056] Furthermore, regarding resource contention: a multi-level feedback queue strategy is adopted for computing power resource contention, where high-priority tasks are first placed into a high-level queue to obtain more time slices, while low-priority tasks can be promoted to a high-level queue when computing power is idle; a token bucket algorithm is used for network bandwidth contention to limit and guarantee the data transmission rate of each video stream, ensuring the bandwidth requirements of critical video streams; and storage I / O contention adopts IO priority grading, with real-time computing tasks having higher I / O priority than offline batch processing tasks.

[0057] Regarding maximizing computing power utilization: When switching algorithms in time-division multiplexing for the same video stream, if the next algorithm container image is already cached on this node and the target computing power type is the same, the existing container runtime environment is reused, and only the algorithm model loading is switched, reducing the switching latency from seconds to milliseconds; During the idle computing power gaps of real-time computing tasks (such as the gaps while waiting for the next video frame to arrive), idle computing power is automatically allocated to tasks in the offline batch processing task queue to ensure that computing power is not idle.

[0058] Also deployed in PaaS capability layer 200, the multi-dimensional sharing enabling module 4 is used to receive structured data output by algorithm container instances and execute access control policies based on resource attribute information to publish data services to the outside world through standardized interfaces.

[0059] In a preferred embodiment of the present invention, the access control policy executed by the multi-shared empowerment module 4 includes ultra-fine-grained authorization management. The dimensions of authorization management include at least one of video device resource authorization, service resource authorization, service call quota authorization, call data volume quota authorization, and service call time and space range authorization, and each interface call generates a traceable audit log.

[0060] Specifically, video capability services (source video services, structured data services, etc.) are uniformly packaged, registered, cataloged, and uploaded to form a standardized service catalog. Based on a unified empowerment bus and unified identity authentication, third-party business systems are registered, authorized, and maintained. The system performs full-process monitoring and statistical analysis of empowerment services to achieve quantitative presentation of empowerment effectiveness. The multi-faceted shared empowerment module 4 receives structured data output from the containerized video data intelligent computing module 3 and opens it up to various business application systems deployed in the business application layer (SaaS layer) in the form of standardized API interfaces.

[0061] In a preferred embodiment of the present invention, a video data asset management module 5 is further deployed on the PaaS capability layer 200 for asset identification and confirmation, asset valuation, asset transaction management and asset lifecycle management of video data assets; wherein, the video data assets include video source data and structured data output by the containerized video data intelligent computing module; the asset valuation adopts a dual-track valuation model combining the market approach and the cost approach, and introduces technology correction coefficient, value density correction coefficient, date correction coefficient and capacity correction coefficient for dynamic correction.

[0062] Specifically, such as Figure 3The diagram shows the flowchart of the video data asset valuation algorithm of this invention, illustrating the dual-track valuation model of the market approach and cost approach, as well as the calculation and comprehensive evaluation process of various correction coefficients. The process is divided into three segments: parallel valuation branches, correction coefficient calculation, and comprehensive valuation output. After the video data asset valuation is initiated, the data type classification node first distinguishes between raw video assets and structured datasets, which then proceed to the market approach valuation path and the cost approach valuation path, respectively. The market-based valuation branch extracts comparable sample price benchmarks by screening traffic video data transaction cases in the same industry. Specifically, it executes the following steps: selecting an active trading market (data exchange / data trading platform), selecting similar data types for reference (video data / traffic data / location data), collecting comparable transaction cases (no fewer than 3 comparable cases), and calculating the market price benchmark (using an arithmetic average or weighted average). The cost-based valuation branch calculates the total cost of video acquisition, storage, and AI analysis, identifying the depreciation and loss of hardware and algorithm functions. Specifically, it executes the following steps: calculating data formation costs (acquisition / labeling / storage / governance costs), determining replacement costs (cost of reacquisition under current technological conditions), identifying value loss (functional depreciation / economic depreciation), and calculating the cost-based valuation (replacement cost minus value loss). The outputs of both branches are simultaneously incorporated into the correction coefficient calculation block. This block contains four types of dynamic correction factors: a technology correction coefficient α (based on algorithm recognition accuracy and data integrity), a value density correction coefficient β (based on the ratio of effective information to total data), a date correction coefficient γ (based on data timeliness decay), and a capacity correction coefficient δ (based on the number of video channels and storage scale). These four coefficients are simultaneously superimposed on the market valuation and cost valuation for weighted adjustment. After weighting, the comprehensive valuation is calculated using the formula V = w1×V_market + w2×V_cost ×α×β×γ×δ (where V represents the comprehensive valuation of the video data asset; w1 represents the market approach valuation weight; V_market represents the market valuation output from the market approach assessment branch; w2 represents the cost approach valuation weight, and w1+w2=1; V_cost represents the cost valuation output from the cost approach assessment branch). The final output is a standardized video asset valuation report with asset range and confidence level, which is then stored in the asset ledger. Figure 3 The valuation process shown in the system solves the problem that traditional video data cannot be quantified and circulated, and provides a standardized pricing basis for video data from departments such as ports and transportation management to enter the data element trading market.

[0063] Specifically, asset identification and ownership confirmation is used to sort out video data resources across the entire domain and identify asset elements such as basic information, business information, management information, and value information; asset transaction management is used to record data asset transaction contracts over time and supports the generation of detailed transaction ledgers by region, business, and industry; asset lifecycle management covers the entire process of identification, ownership confirmation, application, inventory, change, and disposal, realizing value-added management of video data assets.

[0064] In a preferred embodiment of the present invention, a full-link security module 6 spanning the PaaS capability layer 200 and the SaaS business application layer 300 is also included. This module provides basic security services (access control, security encryption, data anonymization, and anti-counterfeiting technology) at the capability layer and a security management interface (application logging and audit log visualization management) at the SaaS layer. This provides at least one of the following security measures in the various stages of video data access, aggregation, storage, governance, application, and sharing: access control, security encryption, data anonymization, anti-counterfeiting technology, application logging, and audit logs.

[0065] Specifically, such as Figure 4 The diagram shown illustrates the end-to-end security architecture of this invention. It corresponds to the end-to-end security module 6 in the capability layer, horizontally divided into four major security protection domains: network layer, platform layer, application layer, and audit layer. At the bottom, it horizontally connects the complete data lifecycle, including video data access, data aggregation, data storage, data governance, data application, data sharing, and asset disposal. Among these, network layer protection (corresponding to...) Figure 4 The "Network Security Protection" domain employs TLS 1.3 transport encryption (supporting forward security, ensuring that even if long-term keys are leaked, historical session keys cannot be cracked; supporting 1-RTT handshake optimization to reduce connection establishment latency), VPC network isolation (implementing east-west traffic control through VPC / security groups to manage lateral communication traffic between containers and services, preventing lateral penetration), DDoS traffic scrubbing, and WAF web firewall mechanisms to protect the infrastructure layer video streaming transmission link; platform layer protection (corresponding to...) Figure 4 The "Platform Security Protection" domain deploys national cryptographic SM4 static storage encryption (national cryptographic SM4 algorithm, static data encryption), role-based access control (RBAC) basic permissions (role-based permission model, least privilege principle), and attribute-based access control (ABAC) fine-grained dynamic authorization (attribute-based permission model, fine-grained dynamic authorization), managing capability layer computing power and video storage resources; application layer protection (corresponding to...). Figure 4The "Application Security Protection" domain supports dynamic and static data masking (dynamic / static masking, K-anonymization), video digital watermarking for tracing (visible / blind watermarking, data tracing), and OAuth2.0 interface authentication (OAuth2.0 / JWT Token, API gateway rate limiting), managing external data service output; and audit layer protection (corresponding to...). Figure 4 The "full audit coverage" domain is used to build an ELK log collection system (full-link operation log, ELK log collection and analysis), which records device access, algorithm scheduling, data calls and asset transaction operations in the entire chain, and is equipped with User Behavior Analysis (UBA) abnormal behavior detection and compliance audit report generation capabilities (meeting the requirements of Level 3 Information Security Protection / Data Security Law). Figure 4 The bottom section shows the overall data lifecycle chain, with four major security protection domains covering all stages of the chain. Each operation retains traceable logs, achieving 100% coverage for security incident tracing. Based on... Figure 4 The four-layer protection architecture shown, combined with the multi-dimensional authorization mechanism of the multi-shared empowerment module 4, completely avoids the risks of unauthorized video access, data leakage and tampering from the bottom layer transmission to the upper layer business.

[0066] In a preferred embodiment of the present invention, the system is logically deployed in a three-layer architecture, and physically deployed in a distributed architecture consisting of one central platform, N departmental nodes, and M social nodes. The three-layer architecture includes: a bottom layer (infrastructure layer) deploying a multi-source aggregation access module and a video data full-element governance module; a middle layer (capability layer) deploying a containerized video data intelligent computing module; and a top layer (business application layer) deploying a multi-faceted sharing and empowerment module. The distributed architecture includes: a central platform deploying a computing power management unit and a task orchestration unit to uniformly schedule computing resources across the entire domain; N departmental nodes serving as regional computing power work nodes, deploying local computing versions of the containerized video data intelligent computing module to execute video data computing tasks within their respective regions and report structured data to the central platform; and M social nodes serving as temporary authorized access nodes, deploying only video stream forwarding proxies for the multi-source aggregation access module to connect video streams to the central platform or nearby departmental nodes within a defined video sharing time and space range.

[0067] Specifically, such as Figure 5 The diagram shown is a distributed deployment topology diagram of the "1+N+M" visual hub of this invention, illustrating the deployment relationship of one central platform, N traffic management department nodes, and M social unit aggregation nodes, as well as the data reporting and distribution paths. It is divided into three layers of node links from top to bottom, clearly marking the bidirectional transmission channels for data reporting and command distribution, wherein: The first layer includes a central platform, fully equipped with a complete set of IaaS, capability, and SaaS modules. It serves as the unified management unit 31 for global computing power and the task orchestration unit 33 for the K8sMaster node. It aggregates video and structured data from all nodes and possesses core capabilities such as a full video asset catalog, a unified algorithm marketplace, global computing power scheduling, a data asset valuation engine, and the highest level of access control. It generates a unified video asset catalog, a global computing power scheduling plan, and a video asset valuation ledger for the entire city.

[0068] The second layer contains N visual hub nodes for traffic management departments (N corresponding to traffic detachments in each district, ports, and airports). Each department node is lightweightly deployed with an IaaS layer multi-source aggregation access module 1, a capability layer video data full-element governance module 2, and a containerized video data intelligent computing module 3. As a K8s Worker edge computing node, it only processes video data within its jurisdiction. Structured data is uploaded to the central platform along the data reporting link (including video metadata reporting, event alarm reporting, and statistical information aggregation). The central platform's scheduling instructions and algorithm update packages are pushed layer by layer to the department nodes along the data distribution / empowerment link (including algorithm model distribution, scheduling instruction distribution, and analysis result empowerment). In the case of network outage, the department nodes can independently complete local video governance and AI computing.

[0069] The third layer contains M social aggregation nodes (M corresponds to commercial complexes, sports stadiums, and logistics parks). It only deploys IaaS layer video stream forwarding agents and has no capability layer computing modules. It only uploads video streams to the nearest department node within the authorized time and space range and supports short-term temporary sharing permission configuration for events within 48 hours.

[0070] Figure 5 The diagram illustrates a two-way data link: the M social node → N departmental nodes → central platform is the data reporting uplink, used to push raw video and AI-structured vehicle / box number data. The uplink also includes video metadata reporting, event alarm reporting, and statistical information aggregation. The central platform → N departmental nodes → M social node is the instruction downlink, used to issue scheduling rules, algorithm images, and shared authorization strategies. The downlink also includes algorithm model issuance, scheduling instruction issuance, and analysis result empowerment.

[0071] Relying on Figure 5 The distributed topology shown allows for lightweight deployment of tens of thousands of video feeds across different areas in a city, eliminating the need for a large-scale computing power room built uniformly throughout the city. This saves 40%-50% on hardware investment and reduces annual computing power maintenance costs by 35%. Taking a port scenario as an example, M social nodes upload video to N port department nodes, complete container OCR calculations locally, and then aggregate the data to the central platform. This can reduce manual container inspection costs by approximately 1.2 million yuan per year and improve hoisting scheduling efficiency by approximately 25%.

[0072] Example 1: Cross-departmental video data fusion computing in a port container truck dispatching scenario Taking the trucking and shipping scheduling of a port operator and its upstream and downstream related enterprises as a typical application scenario, the specific implementation process of the technical solution of this invention is explained: Step 1: Multi-source video data aggregation and access The system connects to fixed cameras in various operating areas of the port yard and cameras of the joint yard operating entity via the GB / T 28181 protocol, and simultaneously accesses structured data of vehicle views via the GA / T 1400 protocol. Data from mobile law enforcement recorders is uniformly accessed through the mobile video governance module, and all data is aggregated and managed at the IaaS layer.

[0073] Step 2: Classification and Management of Video Data The video data full-element governance module classifies and labels the accessed port video points according to the location dimension (loading and unloading area, storage yard area, gate area) and automatically generates application tags (container scheduling, remote container inspection, road traffic flow); it also manages the data in a hierarchical manner according to equipment level and passenger flow density, and prioritizes the allocation of computing resources to key operation areas.

[0074] Step 3: Containerized AI Algorithm Task Orchestration and Execution In the containerized video data intelligent computing module, the system administrator creates the following three types of containerized computing tasks for this business scenario: Task A (Improving the efficiency of remote container lifting): Call on the algorithm container of the AI ​​vendor that supports the target detection algorithm to perform real-time analysis of the crane operation video, identify the position and attitude of the container, and output the lifting path optimization suggestions; Task B (Remote Container Inspection): Call the algorithm container based on image OCR technology to automatically identify the container number, output the container number information in a structured manner, and compare it with the port business system in real time. Task C (Traffic Flow Analysis Around the Yard): Based on shared video image data from multiple joint operating entities around the yard, the task calls a traffic flow statistics algorithm container to perform real-time statistics on the entry and exit flow of container trucks, combines historical data models to predict peak entry times, and outputs scheduling suggestions.

[0075] The three types of tasks mentioned above share the same set of edge computing resources through a time-sharing scheduling mode. That is, during the off-peak period of nighttime operations, computing resources are preferentially allocated to task C (historical data batch processing and analysis); during the peak period of daytime operations, computing resources are preferentially guaranteed to ensure the real-time computing needs of task A and task B.

[0076] Time-sharing scheduling execution process: The above three types of tasks share the same set of edge computing resources (2×NVIDIA A100 + 4×NVIDIA T4 GPU servers), which are dynamically allocated through the WPRR time-sharing scheduling algorithm: During peak daytime operation period (08:00-20:00): computing resources are prioritized for task A (real-time computing, exclusively using 1×A100) and task B (near real-time, allocated 2×T4), while task C uses the remaining 1×A100 + 2×T4 to run in a basic time slice of 8 seconds. Within a total scheduling cycle of 60 seconds: Task A is allocated 60 seconds (exclusive), Task B is allocated 60 seconds (exclusive 2×T4), and Task C is allocated 48 seconds (based on weight calculation: 6 / (9+7+6) × 60 × adjustment factor); During off-peak nighttime operations (20:00-08:00 the next day): Task A is downgraded to priority 5 (only anomaly detection is required), releasing the A100 computing power to Task C for offline batch processing analysis (backtracking 30 days of historical traffic flow data modeling), while Task B continues to run. At this time, Task C obtains a larger time slice (approximately 40 seconds / round); Special cases—emergency event triggering: When the port TOS system detects a large number of container trucks entering the site in a concentrated manner (more than 50 trucks within 15 minutes), Task C automatically increases its priority to 8, increasing its time slice to 20 seconds / round, ensuring that scheduling suggestions are generated within 2 minutes.

[0077] Step 4: Output and Empowerment of High-Value Data Services The calculation results are released to the public by the multi-source sharing empowerment module in the form of standardized API interfaces. After the port scheduling system and the joint yard operation enterprise system pass identity authentication and obtain the corresponding service call quota authorization, they can call high-value structured data such as container number identification results and vehicle flow analysis data in real time. The system automatically records the access log of each service call for subsequent statistical analysis of empowerment effectiveness.

[0078] Step 5: Video Data Asset Management The box number recognition result dataset and traffic flow time series dataset generated in this scenario are entered into the data asset ledger as data assets. They are valued according to the market approach and the cost approach. A value density correction coefficient is introduced (based on reasonable inference from the data document: the value density correction coefficient of OCR structured data is preferably set to the range of 1.2~1.5) to correct the final valuation. Standardized transaction contracts and ledger records are generated in the video data asset management module.

[0079] Example 2: Passenger Flow and Parking Management Scenario at Sports Event Venues During a sports event (accommodating approximately 50,000 spectators), a system analyzed passenger and vehicle flow data from video data (approximately 80 cameras) across the main passageways of the venue. This data, combined with video data from a nearby state-owned shopping mall group (approximately 120 parking lot and road cameras, with temporary equipment resource authorization obtained through the video empowerment center, valid for 48 hours before and after the event), was used to develop a regional parking management plan. Simultaneously, based on passenger flow analysis from video data of the subway transportation group's underground transportation area (approximately 30 subway station cameras), a crowd control plan was predicted.

[0080] This embodiment highlights the following technical features: Cross-departmental temporary authorization mechanism: The mall group applies for temporary equipment resource authorization through the diversified sharing and empowerment module. The authorization dimensions include: spatial scope (2-kilometer area around the venue), time scope (48 hours before and after the event), and call quota (100,000 times / day). Dynamic deployment of containerized algorithms: For event scenarios, the system dynamically deploys "high-density crowd counting algorithm container" (priority=10, real-time calculation mode) and "parking space occupancy detection algorithm container" (priority=8, near real-time mode) in the K8s cluster, and automatically reclaims resources after the event ends; Real-time early warning push: When the crowd density in a certain channel exceeds the threshold (≥4 people / square meter) or the parking lot saturation exceeds 90%, the early warning information is pushed to the event security command center in real time through the gRPC Server Stream channel, with an end-to-end delay of ≤1.5 seconds.

[0081] Example 3 This embodiment demonstrates the application of the present invention in urban-level traffic congestion management and digital twin situation simulation scenarios, highlighting the system's technical capabilities and beneficial effects in large-scale (tens of thousands of video feeds) and complex scenarios (cross-administrative regions and multi-business department collaboration).

[0082] Scenario: A municipal transportation commission (central platform), in conjunction with its eight district traffic detachments (N=8 visual hub nodes) and 50 key traffic intersections and commercial complexes (M=50 social unit aggregation nodes), has constructed a unified "1+8+50" AI visual hub platform for the entire city. The system connects to approximately 12,000 fixed cameras and 200 mobile devices (drones + law enforcement recorders) throughout the city, deploying approximately 35 AI algorithms covering categories such as vehicle detection, traffic flow statistics, congestion detection, event recognition, license plate recognition, and face detection.

[0083] Step 1: Large-scale video data aggregation and adaptive governance The IaaS layer accesses 12,000 IPCs and NVRs citywide via the GB / T 28181 protocol standard, and accesses structured view data for motor vehicles, non-motor vehicles, and pedestrians via the GA / T 1400 protocol. The video asset center performs automatic hierarchical classification on the 12,000 devices. Automatic location-based classification: Based on the geographic coordinates (GIS) and POI data of the equipment, the equipment is automatically labeled and divided into four categories: main road intersections (approximately 3,000), secondary road intersections (approximately 4,000), branch roads / community roads (approximately 3,000), and commercial complexes / parking lots (approximately 2,000). Automatic labeling based on business dimensions: Automatically associate application tags (traffic monitoring, congestion detection, event recognition, bus priority, pedestrian safety, etc.) according to traffic management business needs; Automatic classification of critical equipment: Based on the average annual traffic flow data and historical accident rate data of the equipment, 3,000 main road intersection equipment and 800 accident-prone road section equipment are marked as "core level" (level=1) and given priority in computing resources.

[0084] Step 2: Typical Applications of Time-Sharing Multiplexing Scheduling Strategies in Transportation Scenarios

[0085] In response to the periodic characteristics of traffic scenarios—high load during morning / evening rush hours, medium to low load during off-peak hours, and extremely low load at night—the system's WPRR time-sharing multiplexing scheduling algorithm operates according to the following strategy:

[0086] Key scheduling decision examples: During the morning rush hour (07:00-09:00), when the system detects that the vehicle speed in the video stream of a main road (such as Yan'an Elevated Road) is consistently below 15km / h and exhibits a "tailing" phenomenon (vehicle queue length exceeds the intersection capacity), the system automatically performs the following scheduling actions: The priority of the congestion detection algorithm is temporarily increased from 10 (normal congestion monitoring) to 10+ (emergency preemption mode), obtaining a larger computing time slice (increased from 30 seconds / round to continuous operation mode); a "traffic event recognition algorithm container" (priority=9) is automatically deployed in the video stream of that road segment to detect whether there are abnormal events such as traffic accidents or vehicle breakdowns; if a traffic accident is confirmed, the system automatically deploys a "path replanning detection algorithm container" in the video stream of adjacent road segments to analyze the vehicle detour capacity of the surrounding road network; the priority of the license plate recognition task is temporarily reduced (from 7 to 5) in the video streams of the two intersections upstream and two intersections downstream of the congested road segment, and the freed computing resources are allocated to the detection tasks in the congested area.

[0087] Step 3: Urban traffic operation perception and situational simulation based on digital twins This invention combines digital twin technology, fusing structured video data with urban road network GIS data and traffic signal control data to construct a digital twin model of urban traffic. Specific implementation process: Data Fusion Layer: Through the Flink streaming engine, video AI structured data (traffic flow, vehicle speed, occupancy rate), traffic signal timing data, and floating car GPS trajectory data are spatiotemporally aligned in the data lake with a unified time axis (accuracy down to the second). Model building layer: Based on the SUMO (Simulation of Urban MObility) open-source traffic simulation engine, it uses real road network data as the skeleton and real-time video data as input to build a city-level digital twin model of traffic operation. The model parameters (road capacity, vehicle following model parameters, lane changing behavior parameters) are calibrated offline using historical video data. Situation simulation layer: In the digital twin environment, the following scenarios are simulated in a forward-looking manner to enhance the real world: simulation of the impact range of traffic accidents (simulation time: 30 minutes in the future, simulation time ≤ 3 minutes); simulation of traffic signal timing optimization (simulating the road network operation effect under different timing schemes); prediction of traffic pressure for large-scale events (sports events, concerts, etc.) (simulation 12 hours in advance).

[0088] Step 4: Cross-departmental collaborative management – ​​refined service sharing based on the enabling center In this embodiment, four types of business entities—the Municipal Transportation Commission, the Traffic Management Department of the Municipal Public Security Bureau, the Municipal Emergency Management Bureau, and the Public Transport Group—obtain video data services uniformly through the empowerment center. Municipal Transportation Commission: Obtain aggregated data services such as road network congestion index and traffic characteristics (call quota: 1 million times / month); Municipal Public Security Bureau Traffic Management Department: Obtain detailed data services such as license plate recognition results and traffic incident early warnings (call quota: 2 million times / month); Municipal Emergency Management Bureau: Obtain real-time video push service for emergencies (no quota limit, highest priority); Public Transport Group: Obtain bus lane occupancy detection data and bus priority passage suggestions (quota: 500,000 times / month).

[0089] Each service call is authenticated and logged through a unified empowerment bus, and the system generates monthly statistical reports on the empowerment effectiveness of each department.

[0090] Step 5: Data Asset Management – ​​Continuous Value Added from Traffic Video Data The core data assets generated in this embodiment include: Road network traffic feature dataset (approximately 2 million time-series records per day) Traffic incident database (approximately 500 incident records per day, including time, location, type, and impact assessment) Congestion Characteristics and Patterns Analysis Report (Updated Quarterly) The aforementioned data assets are comprehensively valued using a combination of market approach (referencing the transaction prices of similar urban traffic data, with a benchmark price of approximately 0.5 yuan / MB of structured data) and cost approach, taking into account a technical correction factor (1.2, based on data quality) and a value density correction factor (1.3, based on the breadth of business applications). This results in standardized data asset trading targets, which are then managed uniformly through the data asset trading module.

[0091] Summary of the beneficial effects of this embodiment:

[0092] The above description is merely a preferred embodiment of the present invention and does not limit the implementation and protection scope of the present invention. Those skilled in the art should realize that any equivalent substitutions and obvious changes made using the content of this specification and illustrations should be included within the protection scope of the present invention.

Claims

1. A containerized framework-based intelligent computing and empowerment system for urban traffic video data, characterized in that, include: The multi-source aggregation and access module is used to uniformly aggregate and access multi-source video data based on standard video transmission protocols and view library protocols; the video data full-element governance module is used to perform hierarchical and classified governance of aggregated videos and build a video data resource ledger containing video point resource attribute information. The containerized video data intelligent computing module includes: a computing power management unit, used to uniformly register heterogeneous computing power resources to a computing power management pool; an algorithm management unit, used to independently package various AI algorithms into container images; a task orchestration unit, used to query the video data resource ledger according to the received AI computing task, obtain the resource attribute information corresponding to the target video point pointed to by the AI ​​computing task, then determine the computing power scheduling priority and the matching target AI algorithm of the AI ​​computing task based on the resource attribute information, and then dynamically create or reuse container instances carrying the target AI algorithm based on the computing power scheduling priority and the real-time load status of the computing power management pool, and distribute the video stream of the target video point to the container instance for processing; and a multi-dimensional sharing and empowerment module, used to receive structured data output by the algorithm container instance, and execute access control policies based on the resource attribute information to publish data services externally through standardized interfaces.

2. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, The video data full-element governance module includes a classification and hierarchical governance unit, which is used to classify and govern the aggregated videos according to at least one of the following dimensions: location, business, application, and security. The corresponding video data resource ledger is constructed in a tree-like directory structure according to the dimensions of region, unit, attribute, or object.

3. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, The video data full-element governance module also includes a data quality audit unit, which is used to perform full-link data quality audit on the structured data reported from the front-end device and output from the algorithm container. The audit includes integrity check, consistency check, standardization check and uniqueness check.

4. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, The algorithm management unit registers AI algorithms from different algorithm providers to the algorithm management pool through a standardized algorithm interface, and configures corresponding container images for each AI algorithm. The standardized algorithm interface includes a management plane channel and a data plane channel. The management plane channel uses the REST over HTTP / 2 protocol for algorithm registration, computing power application, task status query, and configuration distribution. The data plane channel uses the gRPC bidirectional streaming protocol for transmitting video frame data to algorithm container instances and receiving structured analysis results returned by the algorithm container instances. The algorithm management unit also pre-caches frequently used algorithm container images to each computing node to reduce container cold start latency.

5. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, The resource attribute information includes location level, location dimension label, and business dimension label. Based on this resource attribute information, the priority of computing power scheduling and the matching of target AI algorithms are determined, including: when the location level is core level, exclusive computing power resources are allocated and given the highest scheduling priority; when the location level is important level, computing power resources are allocated using a time-sharing multiplexing mode; when the location level is general level, computing power resources are allocated only when they are idle; when the location dimension label is port, container recognition or truck flow statistics algorithms are matched; when the business dimension label is emergency command, the video stream priority is increased to the highest level.

6. The intelligent computing and empowerment system for urban traffic video data according to claim 5, characterized in that, In the time-sharing multiplexing mode, the task orchestration unit allocates the same computing resources to multiple AI computing tasks based on a weighted priority time slice round-robin strategy. This includes: calculating a normalized weight based on the priority of each AI computing task; allocating a basic time slice for each AI computing task based on the normalized weight; executing each AI computing task sequentially according to the scheduling queue order; and moving unfinished tasks to the end of the queue to await the next round of scheduling after execution. When a high-priority task arrives and there are no available computing resources, the lowest priority task currently running is selected to perform a checkpoint saving, releasing the computing resources it occupies and allocating them to the high-priority task. The preempted task is then re-enqueued and marked as pending recovery.

7. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, The access control policy executed by the multi-shared empowerment module includes ultra-fine-grained authorization management. The dimensions of the authorization management include at least one of video device resource authorization, service resource authorization, service call quota authorization, call data volume quota authorization, and service call time and space range authorization. Each interface call generates a traceable audit log.

8. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, It also includes a video data asset management module, used for asset identification and confirmation, asset valuation, asset transaction management, and asset lifecycle management of video data assets; wherein, the video data assets include video source data and structured data output by the containerized video data intelligent computing module; the asset valuation adopts a dual-track valuation model combining the market approach and the cost approach, and introduces technology correction coefficient, value density correction coefficient, date correction coefficient, and capacity correction coefficient for dynamic adjustment.

9. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, It also includes a full-link security module, which provides at least one of the following security measures in each stage of video data access, aggregation, storage, governance, application and sharing: access control, security encryption, data desensitization, technical anti-counterfeiting, application logging and audit logs.

10. The intelligent computing and empowerment system for urban traffic video data according to claim 1, characterized in that, The system logically adopts a three-layer architecture and physically employs a distributed architecture consisting of one central platform, N departmental nodes, and M social nodes. The three-layer architecture comprises: a bottom layer (infrastructure layer) deploying the multi-source aggregation access module and the video data full-element governance module; a middle layer (capability layer) deploying the containerized video data intelligent computing module; and a top layer (business application layer) deploying the multi-dimensional sharing and empowerment module. The distributed architecture includes: the central platform deploying the computing power management unit and the task orchestration unit for unified scheduling of computing resources across the entire domain; the N departmental nodes serving as regional computing power work nodes, deploying local computing versions of the containerized video data intelligent computing module to execute video data computing tasks within their respective regions and report structured data to the central platform; and the M social nodes serving as temporary authorized access nodes, deploying only the video stream forwarding proxy of the multi-source aggregation access module to connect the video stream to the central platform or the nearest departmental node within the defined video sharing time and space range.

Images