Space coordinate tag driven control ball data stream scheduling method and system

Abstract

This invention discloses a method and system for scheduling deployment ball data streams based on spatial coordinate labels, belonging to the field of deployment ball data stream scheduling technology. The method includes: constructing an asymmetric address mapping architecture in the near-memory processing memory; performing target detection through a binocular vision module to generate a spatial coordinate label stream, and sending the spatial coordinate label stream to the near-memory processing memory; obtaining a first processing result and a second processing result; and dynamically adjusting the mapping level of the spatial coordinate label stream of the target personnel in the asymmetric address mapping architecture. This invention solves the technical problems of serial processing of multi-dimensional feature data and fixed, rigid storage mapping in existing technologies, which lead to lagging data scheduling and poor real-time risk identification. It achieves the technical effect of realizing the parallel processing of the deployment ball spatial coordinate label stream and the dynamic adaptive adjustment of the storage mapping level, thereby improving the real-time processing efficiency of deployment ball data scheduling and risk identification.

Description

Technical Field

[0001] This invention relates to the field of data stream scheduling technology for deployed satellites, and specifically to a data stream scheduling method and system for deployed satellites driven by spatial coordinate labels. Background Technology

[0002] Currently, most live-line work sites use ordinary monitoring PTZ cameras for personnel safety monitoring. Traditional PTZ cameras lack binocular vision depth perception capabilities and built-in high-efficiency near-memory processing architecture. Multi-dimensional personnel spatial location, action semantics, and facial identity feature data are mostly processed serially and centrally. The storage address mapping relationship is fixed and cannot be adjusted, making it impossible to achieve split-stream parallel computing of feature tag streams and dynamic adaptation of storage resources. This results in problems such as large data scheduling delays, low feature processing efficiency, and delayed response to high-risk personnel identification, making it difficult to meet the application requirements for real-time personnel monitoring, rapid risk assessment, and immediate early warning in high-risk live-line work scenarios.

[0003] Existing technologies suffer from problems such as serial processing and fixed, rigid storage mapping of multi-dimensional feature data in the deployment ball, leading to data scheduling delays and poor real-time risk identification. Summary of the Invention

[0004] This application provides a data stream scheduling method and system for deployment balls based on spatial coordinate labels, which is used to address the technical problems of serial processing of multi-dimensional feature data of deployment balls and fixed and rigid storage mapping in existing technologies, resulting in data scheduling lag and poor real-time performance of risk identification.

[0005] In view of the above problems, this application provides a method and system for scheduling data streams of deployed spheres based on spatial coordinate labels.

[0006] The first aspect of this application provides a method for scheduling deployment ball data streams based on spatial coordinate labels, the method comprising: An asymmetric address mapping architecture is constructed in the near-memory processing memory, wherein the asymmetric address mapping architecture includes a first type of storage cluster and a second type of storage cluster; target detection is performed through a binocular vision module to generate a spatial coordinate label stream, and the spatial coordinate label stream is sent to the near-memory processing memory, wherein the spatial coordinate label stream includes the stereoscopic spatial coordinate data of the target person, continuous action semantic identifiers, and facial identity identifiers; the spatial coordinate label stream is processed in parallel through the first type of storage cluster and the second type of storage cluster in the asymmetric address mapping architecture within the near-memory processing memory to obtain a first processing result and a second processing result; based on the first processing result and the second processing result, the mapping level of the target person's spatial coordinate label stream in the asymmetric address mapping architecture is dynamically adjusted.

[0007] A second aspect of this application provides a data stream scheduling system for a controlled sphere based on spatial coordinate labels, the system comprising: A mapping architecture construction module is used to construct an asymmetric address mapping architecture in the near-memory processing memory, wherein the asymmetric address mapping architecture includes a first type of storage cluster and a second type of storage cluster; a spatial coordinate label stream generation module is used to perform target detection through a binocular vision module, generate a spatial coordinate label stream, and send the spatial coordinate label stream to the near-memory processing memory, wherein the spatial coordinate label stream includes the stereoscopic spatial coordinate data of the target person, continuous action semantic identifiers, and facial identity identifiers; a parallel processing execution module is used to perform parallel processing on the spatial coordinate label stream through the first type of storage cluster and the second type of storage cluster of the asymmetric address mapping architecture in the near-memory processing memory to obtain a first processing result and a second processing result; a mapping level adjustment module is used to dynamically adjust the mapping level of the spatial coordinate label stream of the target person in the asymmetric address mapping architecture according to the first processing result and the second processing result.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: An asymmetric address mapping architecture is constructed in the near-memory processing memory. Target detection is performed using a binocular vision module to generate a spatial coordinate label stream, which is then sent to the near-memory processing memory. The spatial coordinate label stream is processed in parallel using a first type of storage cluster and a second type of storage cluster within the asymmetric address mapping architecture in the near-memory processing memory, yielding a first processing result and a second processing result. Based on the first and second processing results, the mapping level of the spatial coordinate label stream of the target personnel in the asymmetric address mapping architecture is dynamically adjusted. This achieves the technical effect of realizing the splitting and parallel processing of the spatial coordinate label stream of the surveillance projectile and the dynamic adaptive adjustment of the storage mapping level, thereby improving the real-time processing efficiency of surveillance projectile data scheduling and risk identification. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic diagram of the data stream scheduling method for a deployment ball based on spatial coordinate labels provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a deployment ball data stream scheduling system driven by spatial coordinate labels, provided in an embodiment of this application.

[0011] Figure 3 Line graph comparing the performance of the parallel scheduling scheme provided in this application embodiment with the traditional serial scheduling scheme in four types of live-line working scenarios.

[0012] Figure labeling: Mapping architecture construction module 10, spatial coordinate label stream generation module 20, parallel processing execution module 30, mapping level adjustment module 40. Detailed Implementation

[0013] This application provides a data stream scheduling method and system for deployment balls based on spatial coordinate labels, which addresses the technical problems of serial processing and fixed and rigid storage mapping of multi-dimensional feature data of deployment balls in existing technologies, resulting in data scheduling lag and poor real-time performance of risk identification.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] Example 1, as Figure 1 As shown, this application provides a data flow scheduling method for a deployment ball driven by spatial coordinate labels. The deployment ball integrates a binocular vision module and near-memory processing memory. The method includes: Step S100: Construct an asymmetric address mapping architecture in the near-memory processing memory, wherein the asymmetric address mapping architecture includes a first type of storage cluster and a second type of storage cluster.

[0016] Specifically, the surveillance sphere, as a dedicated monitoring device for live-line work sites, integrates two core hardware components: a binocular vision module and a near-field processing memory. These two components work together to meet the real-time requirements of on-site personnel monitoring, data processing, and risk assessment. The binocular vision module consists of two synchronously operating imaging units that can simultaneously acquire binocular images from the site. Utilizing the principle of binocular parallax, it performs functions such as 3D spatial coordinate calculation, extraction of key points of personnel skeletons, and face capture, providing raw image and feature data support for subsequent spatial distance calculation, action intent recognition, and identity verification. The near-field processing memory is a dedicated storage component integrated close to the computing unit, breaking the traditional architecture of separating storage and computation. It enables local storage, rapid retrieval, and parallel computation of data, efficiently caching 3D spatial coordinate data, tag information, and facial feature databases, significantly reducing data transmission latency. This ensures high-speed and efficient execution of core steps such as tag stream parsing, spatial grid matching, and face comparison, guaranteeing that the surveillance sphere can perform real-time personnel behavior monitoring and rapid response in high-risk scenarios in complex environments at live-line work sites. Within the near-memory processing memory integrated into the control ball, an asymmetric address mapping architecture suitable for differentiated data stream processing is built. This architecture consists of two types of storage clusters with distinct functions and physical structures. The first type of storage cluster consists of physical storage blocks with parallel computing capabilities within subarrays, while the second type of storage cluster consists of physical storage blocks with parallel retrieval capabilities across storage banks. The two types of storage clusters correspond to different physical address ranges, providing hardware support for the subsequent classification, parallel processing, and dynamic scheduling of spatial coordinate tag streams.

[0017] Step S200: Perform target detection through the binocular vision module, generate a spatial coordinate label stream, and send the spatial coordinate label stream to the near-memory processing memory. The spatial coordinate label stream includes the stereoscopic spatial coordinate data of the target person, continuous action semantic identifiers, and facial identity identifiers.

[0018] Specifically, the binocular vision module mounted on the surveillance sphere performs real-time detection of personnel within the monitored area. Based on the principle of binocular parallax, it registers the left and right eye flows, calculates the three-dimensional bounding box distance between the personnel and the charged object, and generates a three-dimensional protective spatial label containing three-dimensional spatial coordinates. Simultaneously, it performs temporal convolutional encoding on the coordinate sequence of key points of the human skeleton in consecutive frames to obtain continuous action semantic labels such as climbing, falling, or arc discharge associated postures. Then, through the face capture module in the binocular vision module, it extracts and hashes facial features of personnel entering the monitored area to generate facial identity labels. The above three-dimensional spatial coordinate data, continuous action semantic labels, and facial identity labels are summarized to form a complete spatial coordinate label stream, which is sent to the built-in near-memory processing memory of the surveillance sphere to provide a data foundation for subsequent parallel processing.

[0019] Step S300: Perform parallel processing on the spatial coordinate label stream using the first type of storage cluster and the second type of storage cluster of the near-memory processing in-memory asymmetric address mapping architecture to obtain the first processing result and the second processing result.

[0020] Specifically, the spatial coordinate label stream is first parsed using hot-trigger semantic identifiers. A smart processing module performs temporal convolutional encoding on the coordinate sequences of key points on the human skeleton in consecutive frames to obtain continuous action semantic identifiers. The cumulative trigger frequency of various action semantics within a preset sliding time window is counted in the near-memory processing memory. When the frequency exceeds a preset threshold, a hot-trigger semantic identifier is generated. After identifying the hot-trigger semantic identifier, the corresponding coordinate label stream is scheduled to the first type of storage cluster. Subarray-level near-memory pre-merging logic is activated, and bit-domain differential processing is performed on the continuous frame stereoscopic spatial coordinate data to obtain an action residual feature map as the first processing result. Simultaneously, the spatial coordinate label stream containing facial identity identifiers is stored in the second type of storage cluster. The hardware scheduling engine distributes the data to multiple independent storage banks, and parallel facial feature matching is performed using a multi-storage bank simultaneous broadcast query method. The resulting facial feature comparison is used as the second processing result, thus achieving parallel processing of the spatial coordinate label stream by the two types of storage clusters.

[0021] Step S400: Based on the first processing result and the second processing result, dynamically adjust the mapping level of the spatial coordinate label stream of the target personnel in the asymmetric address mapping architecture.

[0022] Specifically, based on the action residual feature map output by the first type of storage cluster and the facial feature comparison result output by the second type of storage cluster, a comprehensive judgment is made. When the facial feature comparison result hits the preset high-risk label, the system immediately generates a hardware preemption signal. In response to the preemption signal, the three-dimensional spatial coordinate data and continuous action semantic identifier associated with the target person are hot-migrated from the second type of storage cluster to the first type of storage cluster, and exclusive subarray computing resources are allocated to it. In this way, the mapping level of the spatial coordinate label stream of the target person in the asymmetric address mapping architecture is dynamically adjusted, so as to realize the adaptive optimization scheduling of data stream and hardware computing resources.

[0023] In one possible implementation, step S100 further includes: Step S110: The first type of storage cluster consists of physical storage blocks with the ability to perform parallel operations within a subarray.

[0024] Step S120: The second type of storage cluster consists of physical storage blocks that have the ability to perform parallel retrieval across storage volumes.

[0025] The first type of storage cluster and the second type of storage cluster correspond to different physical address ranges.

[0026] Specifically, the first type of storage cluster is constructed using physical storage blocks with parallel computing capabilities within subarrays. These physical storage blocks use subarrays as basic computing units and support direct activation of parallel computing logic within the storage array. They can perform near-memory calculations such as vector difference and feature merging on continuous frame spatial coordinate data stored in different rows within the same subarray through local bit lines in the analog domain, achieving high-speed parallel processing without migrating the data to external computing units.

[0027] The second type of storage cluster is constructed using physical storage blocks with cross-storage parallel retrieval capabilities. These physical storage blocks use multiple independent storage units as basic units, and support the hardware scheduling engine to distribute facial identity identifiers to each storage unit. Through a multi-storage synchronous broadcast query mechanism, feature comparison and database matching are performed in parallel across different storage units, enabling high-speed parallel retrieval of large-scale data without serial access to each storage unit.

[0028] In this system, the first type of storage cluster and the second type of storage cluster are allocated independent and non-overlapping physical address ranges in the system's physical address space. The address domains are strictly isolated through hardware address decoding logic, ensuring that data access, storage operations and computation scheduling of the two types of storage clusters do not interfere with each other, providing a stable address partitioning basis for parallel processing under the asymmetric address mapping architecture.

[0029] In one possible implementation, step S300 further includes: Step S310: Perform hot-trigger semantic identifier parsing on the spatial coordinate label stream, wherein the hot-trigger semantic identifier is used to characterize whether the cumulative trigger frequency of the continuous action semantic identifier within the preset sliding time window exceeds the preset cumulative action trigger frequency threshold.

[0030] Step S320: When the hot-triggered semantic identifier is identified, the coordinate label stream containing the continuous action semantic identifier is scheduled to the first type of storage cluster, the sub-array level near-memory pre-merge logic is activated, bit-domain differential processing is performed on the three-dimensional spatial coordinate data of the continuous frames, and the obtained action residual feature map is used as the first processing result.

[0031] Step S330: Store the spatial coordinate label stream containing the face identity identifier into the second type of storage cluster, and schedule the parallel retrieval logic across storage volumes to perform parallel matching on the face identity identifier, and use the obtained face feature comparison result as the second processing result.

[0032] Specifically, the input spatial coordinate label stream is parsed for hot-trigger semantic markup. First, the intelligent processing module of the deployment ball performs temporal convolutional encoding on the coordinate sequence of key points of human skeleton in consecutive frames to generate continuous action semantic marks such as climbing, falling, and arc discharge associated postures. Then, in the near-memory processing memory, the trigger frequency counter bound to the preset three-dimensional protective space grid is used to count the cumulative number of triggers of various continuous action semantic marks in each grid cell within the preset sliding time window. It is then determined whether the cumulative trigger frequency exceeds the preset action trigger cumulative frequency threshold, thereby generating hot-trigger semantic marks for marking highly sensitive action areas.

[0033] Upon identifying a thermal trigger semantic identifier, the coordinate label stream containing continuous action semantic identifiers is first scheduled to the first type of storage cluster. Based on the real-time three-dimensional trajectory point sequence of the target personnel, the matching and positioning with the preset three-dimensional protective space grid is completed. Subsequently, the subarray-level near-memory pre-merging logic and the subarray parallel sensitive amplifier are activated to store the three-dimensional spatial coordinate data of consecutive frames within the same grid cell into different rows of the same subarray. The displacement difference calculation of the coordinate vectors of adjacent frames is directly completed in the analog domain through local bit lines. Bit-domain differential processing is performed on the three-dimensional spatial coordinate data of consecutive frames, and finally, an action residual feature map is generated to characterize the instantaneous change rate of human posture. This action residual feature map is then output as the first processing result.

[0034] The hardware scheduling engine distributes the spatial coordinate label stream containing facial identity identifiers to multiple independent storage banks in the second type of storage cluster. Then, the cross-storage parallel retrieval logic is scheduled and activated. A multi-storage synchronous broadcast query mechanism is adopted to perform parallel feature comparison and global matching between the facial identity identifier to be identified and the high-risk personnel face database data pre-stored in each storage bank, quickly complete the identity verification, and finally output the obtained facial feature comparison result as the second processing result.

[0035] In one possible implementation, step S310 further includes: Step S311: The intelligent processing module of the control ball performs temporal convolutional encoding on the human skeleton key point coordinate sequence of the continuous frames corresponding to the spatial coordinate label stream to obtain continuous action semantic identifiers, wherein the continuous action semantic identifiers include climbing semantic encoding, falling semantic encoding or arc discharge associated posture semantic encoding.

[0036] Step S312: Set a trigger frequency counter associated with the preset three-dimensional protective space grid in the near-memory processing memory, and use the trigger frequency counter to count the cumulative number of occurrences of various continuous action semantic identifiers in each space grid unit within the preset sliding time window.

[0037] Step S313: When the cumulative triggering frequency of a specific continuous action semantic identifier within a certain spatial grid cell exceeds the preset cumulative triggering frequency threshold, a hot trigger semantic identifier carrying the coordinate information of that grid cell is generated.

[0038] Specifically, the intelligent processing module of the control ball first extracts the coordinate sequence of key points of the human skeleton from the target person in the continuous video frames corresponding to the spatial coordinate label stream, and then performs temporal convolutional coding processing on the temporal coordinate data to extract continuous action semantic identifiers that can represent the person's behavioral intentions. The continuous action semantic identifiers are digitally encoded action feature information, specifically including climbing semantic codes for identifying climbing behavior, falling semantic codes for identifying falling behavior, and arc discharge associated posture semantic codes for identifying dangerous postures related to arc discharge, providing accurate action semantic basis for subsequent thermal triggering judgment.

[0039] In the near-memory processing memory of the control ball, a trigger frequency counter is set up and bound to the preset three-dimensional protective space grid. The counter uses a preset sliding time window as the statistical period to accumulate the semantic identifiers of various continuous actions such as climbing, falling, and arc discharge associated postures in each space grid unit in real time, thereby obtaining the cumulative number of occurrences of various continuous action semantic identifiers in each grid unit within the current time window.

[0040] The cumulative trigger frequency of specific continuous action semantic identifiers within each spatial grid cell is compared and determined in real time with a pre-set cumulative trigger frequency threshold. When the cumulative trigger frequency of the corresponding type of continuous action semantic identifier in any spatial grid cell exceeds the set standard, a hot trigger semantic identifier is immediately generated. At the same time, the location coordinate information of the corresponding spatial grid cell is loaded into the identifier to accurately mark the spatial area where the high-frequency dangerous behavior is located, providing a reliable basis for subsequent storage-level scheduling and parallel computing processing.

[0041] In one possible implementation, step S320 further includes: Step S321: The first type of storage cluster includes multiple logical partitions, wherein each logical partition corresponds to a preset three-dimensional protective space grid within the monitoring field of view of the deployment ball.

[0042] Step S322: Parse the stereoscopic spatial coordinate data of the spatial coordinate label stream to obtain the real-time three-dimensional trajectory point sequence of the target person in the binocular parallax coordinate system.

[0043] Step S323: Perform region matching in the preset three-dimensional protective space grid according to the real-time three-dimensional trajectory point sequence to obtain the matching hot spot area identifier.

[0044] Step S324: Load the memory scheduling and operator control strategy corresponding to the matching hotspot region identifier, activate the subarray parallel sensitive amplifier, store the three-dimensional spatial coordinate data of consecutive frames in the same grid cell in different rows of the same subarray, complete the displacement difference calculation of adjacent frame coordinate vectors in the analog domain through local bit lines, and output the action residual feature map.

[0045] Specifically, the first type of storage cluster is divided into several independent logical partitions. Each logical partition is mapped to a pre-defined three-dimensional protective space grid within the monitoring field of view of the surveillance ball, realizing a one-to-one binding relationship between the space grid and the storage logical partition. This facilitates subsequent targeted data scheduling and nearby computation processing based on the spatial location of personnel.

[0046] The system analyzes the stereo spatial coordinate data embedded in the spatial coordinate label stream frame by frame. Based on the parallax ranging principle of binocular imaging, it calculates the parallax and solves the spatial coordinates by combining the pixel correspondence of the left and right binocular images. The system converts the two-dimensional pixel positions into real three-dimensional positions through a coordinate transformation model. It continuously samples the spatial positions of the target personnel at continuous time points, thereby fitting and generating a complete and coherent real-time three-dimensional trajectory point sequence of the target personnel in the coordinate system constructed by binocular parallax.

[0047] Based on the real-time three-dimensional trajectory point sequence of the target personnel, each spatial point in the trajectory sequence is sequentially compared with the grid unit of the preset three-dimensional protective space grid for boundary determination and attribution. The spatial grid position of each trajectory point is verified one by one, and grid units with densely distributed trajectory points and high-frequency behavior are selected to complete the spatial area hit matching. At the same time, exclusive marking information is generated for the matched key grid units to obtain the matching hot spot area identification.

[0048] Based on the matching hotspot area identifier, the preset matching memory scheduling rules and operation operator control configuration strategies are retrieved and then triggered to activate the parallel sensitive amplifier of the subarray to enter the working ready state. According to the spatial grid affiliation principle, the stereo spatial coordinate data corresponding to the continuous video frames in the same protection grid unit are centrally arranged and stored in different physical storage rows of the same storage subarray. Using the local bit line path deployed inside the storage array, the coordinate vectors of adjacent frames are directly calculated in parallel displacement difference in the analog signal processing domain. Based on the difference calculation results, the instantaneous change law of human posture is characterized, and then a standardized motion residual feature map is generated and output.

[0049] In one possible implementation, step S320 further includes: The pre-set three-dimensional protective space grid includes the space grid at the top of the poles in the live working area, the space grid near the transformer, and the space grid for high-voltage cable routing.

[0050] Specifically, the pre-set three-dimensional protective space grid is a three-dimensional grid-based spatial division system formed by uniformly dividing the entire live-line work area within the monitoring field of view of the deployment ball according to a three-dimensional spatial scale. It is used to accurately delineate dangerous control areas and locate the location of personnel behavior. Among them, the pole top space grid specifically refers to the three-dimensional grid area demarcated around the top of the power pole, covering the equipment at the top of the pole and the surrounding activity range, and is used to monitor dangerous behaviors such as personnel climbing to the top of the pole; the transformer proximity space grid specifically refers to the three-dimensional grid area demarcated around the power transformer equipment, covering the entire space within the transformer body and safety protection distance, and is used to monitor risky behaviors such as personnel approaching the transformer and lingering, and violating regulations; the high-voltage cable routing space grid specifically refers to the long strip-shaped three-dimensional grid area demarcated along the route of the high-voltage transmission line, covering the high-voltage cable body, the space below the line and on both sides of the line, and is used to monitor safety hazards such as personnel walking close to the cable, staying, and making dangerous postures.

[0051] In one possible implementation, step S330 further includes: Step S331: The hardware scheduling engine distributes the facial identity identifier to multiple independent storage blocks in the second type of storage cluster.

[0052] Step S332: The second type of storage cluster uses a multi-storage-bank simultaneous broadcast query method to perform parallel matching between the facial identity identifier to be identified and the pre-stored high-risk personnel facial database, and outputs the facial feature comparison result.

[0053] Specifically, the hardware scheduling engine is a built-in hardware scheduling module on the surveillance sphere, possessing the capabilities of task scheduling, data distribution, and storage resource management. Facial identification identifiers are standardized feature data extracted from surveillance footage that characterizes a person's identity. The second type of storage cluster is a dedicated storage cluster specifically for storing facial feature data and personnel file data, internally divided into several independent storage units with parallel access capabilities. Through the hardware scheduling engine, according to preset resource allocation rules and data routing logic, the collected facial identification identifiers are precisely and directionally distributed to the various independent storage units within the second type of storage cluster for cache writing, preparing data for subsequent parallel identity matching and comparison across multiple storage units.

[0054] Relying on multiple independent storage units within the second type of storage cluster, a multi-storage unit synchronous broadcast query mechanism is enabled. The unified facial identity identifier to be verified is simultaneously distributed to each storage unit. Each storage unit retrieves the pre-stored high-risk personnel facial feature data locally and simultaneously performs facial feature similarity comparison and parallel matching operations. After each storage unit independently completes feature retrieval and matching judgment, the operation results are summarized, integrated and standardized, and the final facial feature comparison result is output.

[0055] In one possible implementation, step S400 further includes: Step S410: When the facial feature comparison result in the second processing result matches the preset high-risk label, a hardware preemption signal is generated.

[0056] Step S420: In response to the hardware preemption signal, the three-dimensional spatial coordinate data and continuous action semantic identifier associated with the target person are hot-migrated from the second type of storage cluster to the first type of storage cluster, and dedicated subarray computing resources are allocated.

[0057] Specifically, the system continuously performs full analysis and label verification on the facial feature comparison results included in the second processing result. Various high-risk labels representing dangerous individuals are pre-set within the system. The matching information of the facial feature comparison results is compared with the preset high-risk labels one by one according to rules and attribute matching. Once it is determined that the facial feature comparison result successfully matches any preset high-risk label, the system immediately triggers the internal hardware interrupt logic and resource preemption trigger mechanism. The main control logic unit locks the current system operating state and, according to the preset hardware signal generation timing and level control logic, drives the signal generation circuit step by step to complete level transitions and signal encoding encapsulation. A standard and valid hardware preemption signal is generated at the hardware level to forcibly preempt system storage and processing resources, providing hardware control basis for subsequent high-risk behavior handling and priority scheduling.

[0058] After receiving a hardware preemption signal and making a high-risk judgment, the system executes a scheduling response operation. The hardware preemption signal is a priority control instruction triggered by hardware logic after identifying a high-risk personnel scenario, which can forcibly elevate the system scheduling authority of the current task. The three-dimensional spatial coordinate data associated with the target personnel is complete location data that depicts the three-dimensional spatial position and movement trajectory of the target personnel in the monitoring scenario. The continuous action semantic identifier is standardized behavior label information used to distinguish personnel climbing, falling, dangerous postures, and other behavior types after feature extraction. The second type of storage cluster is a dedicated storage architecture for storing facial identity feature information and high-risk personnel file data. The first type of storage cluster is a dedicated storage architecture for behavior data that adapts to the three-dimensional protection space grid mapping relationship and supports nearby parallel computing processing. Hot migration is a lossless scheduling method that realizes real-time and complete data transfer across storage architectures without interrupting business processes while the system is running normally. Dedicated subarray computing resources are dedicated storage subarrays and hardware computing resources that are independently allocated for a single high-priority task, not shared, and not occupied by other businesses. This step responds immediately to the priority scheduling command issued by the hardware preemption signal, and completely migrates the three-dimensional spatial coordinate data and continuous action semantic identifiers of the target personnel from the second type of storage cluster to the first type of storage cluster through lossless hot migration. At the same time, it allocates and distributes dedicated subarray computing resources for the current high-risk monitoring task to ensure that the spatial location analysis and behavioral feature calculation of high-risk personnel can be completed without interference from other tasks, and the subsequent judgment and processing can be completed with the highest priority.

[0059] In one possible implementation, step S200 further includes: Step S210: Register the left and right eye streams using the binocular parallax principle, calculate the three-dimensional bounding box distance between the personnel and the charged body, and generate a three-dimensional protective space label.

[0060] Step S220: Obtain action intent semantic labels by performing temporal convolutional encoding on the sequence of key point coordinates of human skeleton in consecutive frames.

[0061] Step S230: Using the face capture module in the binocular vision module, extract facial features and hash encode the faces of people entering the monitoring area to generate facial identity identifiers.

[0062] Step S240: Summarize the three-dimensional protective space tags, action intent semantic tags, and facial identity identifiers to obtain a spatial coordinate tag stream.

[0063] Specifically, binocular vision imaging technology is employed, based on the principle of binocular parallax, to accurately register the optical flows of the left and right eyes and calculate the distance between the person and the three-dimensional bounding box of the charged object. First, the binocular vision module simultaneously acquires two monitoring images from the left and right eyes, extracting optical flow information from both images. This optical flow information is used to characterize the motion trajectory and velocity of pixels in the image. Subsequently, a feature point matching algorithm is used to register the optical flows of the left and right eyes. By extracting key feature points such as corners and edges from the optical flows, a one-to-one correspondence is established between the feature points of the left and right optical flows, correcting the positional deviation of the two optical flows, and achieving accurate alignment of the left and right optical flows, laying the foundation for subsequent three-dimensional distance calculations. After registration, based on the principle of binocular parallax ranging, combined with the pre-calibrated intrinsic parameters of the binocular camera, as well as focal length, pixel size, and extrinsic parameters (i.e., the distance between the two cameras), the three-dimensional spatial coordinates of the personnel and the charged object in the image are calculated using triangulation. Next, three-dimensional bounding boxes are constructed for the personnel and the charged object respectively. The three-dimensional bounding boxes cover the overall outline of the personnel and the charged object based on their three-dimensional coordinates. Then, through spatial geometric calculations, the shortest straight-line distance between the two three-dimensional bounding boxes is calculated, which is the actual safe distance between the personnel and the charged object. Finally, based on the calculated three-dimensional bounding box distance and the preset safe distance threshold, a corresponding three-dimensional protective space label is generated. The label contains information on the distance between the personnel and the charged object, their spatial location, and a hazard level association identifier, which is used for subsequent spatial grid matching and risk assessment.

[0064] The system detects key skeletal points of a target person in the surveillance footage frame by frame, accurately capturing the 3D coordinates of various joints in the head, neck, torso, and limbs. These coordinates are then organized chronologically according to the video frames to form a continuous and complete temporal sequence of skeletal key point coordinates. A temporal convolutional encoding process is then initiated. First, the coordinate sequence undergoes standardization preprocessing to eliminate coordinate scale differences between different frames. The preprocessed sequence is then input into a temporal convolutional network. This network uses multiple concatenated convolutional kernels along the time dimension, employing a fixed-length sliding window to capture skeletal key point coordinate data from several consecutive frames in a segmented manner. Each convolutional kernel performs convolution operations on the captured local temporal segments. This process extracts spatial distribution features of skeletal key points within a single frame, such as the relative positions of limb joints and the posture angles of the torso. It also captures motion changes in skeletal joints between adjacent frames, such as joint displacement amplitude and direction of motion. Meanwhile, by setting skip connections and batch normalization, gradient vanishing during temporal feature extraction is avoided, enhancing the transmission and aggregation of deep temporal features. Through layer-by-layer refinement via multi-layer convolutional operations, the original coordinate sequence is transformed into a deep temporal feature vector capable of representing the laws of limb movement. Finally, this feature vector is input into a fully connected classification layer, where a softmax classifier performs semantic mapping and category discrimination to accurately identify the type of human action, such as climbing, crossing, stopping, waving, etc., and converts it into standardized semantic labels for action intent, completing the encoding conversion from skeletal coordinate temporal sequence to action intent semantics.

[0065] The face capture module, integrated into the binocular vision module, performs high-risk individual identification and feature extraction. This module is specifically designed to capture facial features and generate identity identifiers, accurately extracting facial information and associating it with the target person. First, upon activation, the face capture module captures the face of the person entering the monitored area. A clear facial image is acquired using an image sensor. Then, a feature extraction algorithm precisely extracts detailed features such as facial contours, the position of facial features, skin tone, and texture—the core criteria for distinguishing different individuals. Next, the extracted facial features are standardized to remove redundant information. Finally, a hash encoding algorithm transforms the facial features into a unique, standardized numerical identifier—the facial identity identifier. This identifier uniquely identifies the target individual, facilitating subsequent association with their behavioral data. Then, the system determines whether the identity corresponding to the facial feature belongs to the high-risk category according to the preset high-risk judgment rules. If it is determined to be high-risk, the hardware preemption mechanism is immediately triggered to integrate the facial feature data, related behavioral data and identity identifier of the target person, transfer them to a dedicated storage area through hot migration, and allocate independent computing resources to ensure that the identification and behavior monitoring of high-risk persons are not interfered with by other tasks, and finally generate complete high-risk person identity association data.

[0066] The three-dimensional protective spatial tags, action intent semantic tags, and facial identity tags generated above are time-series aligned and structured for integration. Following a unified data encapsulation standard, the various tag information is synchronized and corresponds in the time dimension. The spatial tags representing the spatial location of personnel and their relationship with dangerous areas, the action semantic tags depicting the tendencies and states of personnel behavior, and the facial tags that can uniquely identify personnel are systematically collected and summarized, and then encapsulated and arranged in a continuous temporal sequence to form a spatial coordinate tag stream that is temporally coherent, structurally regular, and informationally complete.

[0067] This application is adapted to three types of preset three-dimensional protective space grid monitoring scenarios for live-line working in the power industry. At the same time, it constructs a complex working condition with multiple grids superimposed on high-risk personnel to complete comparative tests. The four types of test scenarios are as follows: Scenario S1: The space grid at the top of a 10kV distribution tower poses a risk of falls from height due to frequent climbing activities by personnel. Scenario S2: The box-type transformer is located near the space grid. Unauthorized personnel or unqualified workers may approach the equipment at close range, posing a risk of electric shock. Scenario S3: High-voltage cable routing space grid, people walking under the cables may accidentally fall and come into contact with live wires; Scenario S4: A complex multi-grid mixed scenario where high-risk individuals on the blacklist simultaneously exhibit dangerous behaviors such as climbing power poles and getting close to high-voltage cables. This scenario is used to verify the dynamic hierarchical scheduling capabilities of hardware preemption and data hot migration in this application.

[0068] The test used a standardized binocular control sphere and memory hardware with equivalent parameters for near-memory processing. Performance metrics were collected synchronously for one hour using both the existing traditional serial scheduling scheme and the parallel scheduling scheme of the asymmetric memory cluster proposed in this application. The measured comparison data is shown in Table 1. Table 1. Performance Comparison between Traditional Serial Scheduling Scheme and Parallel Scheduling Scheme of this Application

[0069] Based on the quantitative data in Table 1, the improvement effect of this application compared with the existing serial solution is verified under two types of working conditions: 1. In typical hazardous scenarios S1, S2, and S3 within a single grid, traditional solutions sequentially process multi-dimensional features such as spatial coordinates, actions, and faces, resulting in single-frame processing latency exceeding 40ms, high-risk warning response latency exceeding 170ms, hardware parallel resource utilization below 35%, and 6-9 missed warnings per hour. This application utilizes an asymmetric address mapping architecture to split dual storage clusters for parallel processing, simultaneously executing action residual calculation and face identity matching. Single-frame processing latency is controlled to within 10ms, warning response latency is reduced to within 30ms, hardware resource utilization is increased to over 90%, and there are no missed warnings throughout the process, fundamentally solving the problems of data congestion and recognition lag in sequential processing.

[0070] 2. In the harsh scenario S4, which involves a mix of high-risk personnel in multiple grids and multiple targets and features, traditional solutions suffer from severe computational backlog, with single-frame latency approaching 70ms, warning response latency exceeding 300ms, and the number of missed warnings surging to 14.

[0071] This application triggers a hardware preemption signal after a high-risk label is matched in face comparison, and hot-migrates all coordinates and action data of the target person to the first type of storage cluster and allocates exclusive sub-array resources. Even in complex mixed scenarios, the single-frame latency is only 11.4ms and the warning latency is 35.1ms, which still achieves zero-miss warning and verifies the adaptive scheduling advantages of dynamically adjusting the storage mapping level.

[0072] Meanwhile, the peak memory usage of this application is reduced by 35% to 52% compared with traditional solutions across all scenarios. By relying on spatial grid partitioning for nearby storage and near-memory domain differential lightweight operation, global image cache redundancy is reduced and memory overhead is lowered.

[0073] Further integration Figure 3 As shown: In the four scenarios, the single-frame processing latency and high-risk identification response latency of the traditional serial solution are significantly higher than those of the solution in this application, and the latency deteriorates sharply with the increase of scenario complexity (S4 mixed high-risk scenario); the hardware parallel resource utilization rate of the traditional solution is only 22.8%~33.5%, and the resources are seriously idle; the parallel scheduling solution of this application maintains the latency in an extremely low range in all scenarios, the hardware resource utilization rate is stable at over 90%, and the performance degradation is minimal in complex mixed scenarios, which fully verifies the dual improvement effect of the asymmetric address mapping parallel processing and dynamic mapping hierarchical hot migration mechanism of this application on real-time performance and hardware utilization.

[0074] Example 2, based on the same inventive concept as the spatial coordinate label-driven deployment ball data stream scheduling method in the aforementioned examples, such as... Figure 2 As shown, this application provides a data stream scheduling system for a controlled sphere based on spatial coordinate labels. The system and method embodiments in this application are based on the same inventive concept. The system includes: The mapping architecture construction module 10 is used to construct an asymmetric address mapping architecture in the near-memory processing memory, wherein the asymmetric address mapping architecture includes a first type of storage cluster and a second type of storage cluster.

[0075] The spatial coordinate label stream generation module 20 is used to perform target detection through a binocular vision module, generate a spatial coordinate label stream, and send the spatial coordinate label stream to the near-memory processing memory. The spatial coordinate label stream includes the stereoscopic spatial coordinate data of the target person, continuous action semantic identifiers, and facial identity identifiers.

[0076] The parallel processing execution module 30 is used to perform parallel processing on the spatial coordinate label stream through the first type of storage cluster and the second type of storage cluster of the near-memory processing in-memory asymmetric address mapping architecture to obtain the first processing result and the second processing result.

[0077] The mapping level adjustment module 40 is used to dynamically adjust the mapping level of the spatial coordinate label stream of the target person in the asymmetric address mapping architecture based on the first processing result and the second processing result.

[0078] Furthermore, the system is also used to implement the following functions: The first type of storage cluster consists of physical storage blocks capable of parallel operation within a subarray; the second type of storage cluster consists of physical storage blocks capable of parallel retrieval across storage banks; wherein the first type of storage cluster and the second type of storage cluster correspond to different physical address ranges.

[0079] Furthermore, the system is also used to implement the following functions: The spatial coordinate label stream is parsed using hot-trigger semantic identifiers, whereby the hot-trigger semantic identifiers characterize whether the cumulative trigger frequency of continuous action semantic identifiers within a preset sliding time window exceeds a preset cumulative action trigger frequency threshold. When the hot-trigger semantic identifier is identified, the coordinate label stream containing the continuous action semantic identifiers is scheduled to the first type of storage cluster, the subarray-level near-memory pre-merging logic is activated, bit-domain differential processing is performed on the stereoscopic spatial coordinate data of continuous frames, and the obtained action residual feature map is used as the first processing result. The spatial coordinate label stream containing the face identity identifier is stored to the second type of storage cluster, and cross-storage parallel retrieval logic is scheduled to perform parallel matching on the face identity identifiers, and the obtained face feature comparison result is used as the second processing result.

[0080] Furthermore, the system is also used to implement the following functions: The intelligent processing module of the control ball performs temporal convolutional encoding on the sequence of human skeleton key point coordinates of continuous frames corresponding to the spatial coordinate label stream to obtain continuous action semantic identifiers. The continuous action semantic identifiers include climbing semantic encoding, falling semantic encoding, or arc discharge associated posture semantic encoding. A trigger frequency counter associated with a preset three-dimensional protective spatial grid is set in the near-memory processing memory. The trigger frequency counter counts the cumulative occurrence of various continuous action semantic identifiers in each spatial grid cell within a preset sliding time window. When the cumulative trigger frequency of a specific continuous action semantic identifier in a spatial grid cell exceeds a preset cumulative trigger frequency threshold, a hot trigger semantic identifier carrying the coordinate information of that grid cell is generated.

[0081] Furthermore, the system is also used to implement the following functions: The first type of storage cluster includes multiple logical partitions, each of which corresponds to a preset three-dimensional protective space grid within the monitoring field of view of the surveillance ball; the three-dimensional spatial coordinate data of the spatial coordinate tag stream is parsed to obtain the real-time three-dimensional trajectory point sequence of the target personnel in the binocular parallax coordinate system; regional hit matching is performed in the preset three-dimensional protective space grid according to the real-time three-dimensional trajectory point sequence to obtain the matching hot spot area identifier; the memory scheduling and operator control strategy corresponding to the matching hot spot area identifier is loaded, the parallel sensitive amplifier of the sub-array is activated, the three-dimensional spatial coordinate data of consecutive frames in the same grid cell are stored in different rows of the same sub-array, the displacement difference calculation of the coordinate vectors of adjacent frames is completed in the analog domain through the local bit line, and the action residual feature map is output.

[0082] Furthermore, the system is also used to implement the following functions: The pre-set three-dimensional protective space grid includes the space grid at the top of the poles in the live working area, the space grid near the transformer, and the space grid for high-voltage cable routing.

[0083] Furthermore, the system is also used to implement the following functions: The hardware scheduling engine distributes the facial identity identifiers to multiple independent storage units in the second type of storage cluster. The second type of storage cluster performs parallel matching between the facial identity identifiers to be identified and the pre-stored high-risk personnel facial database by broadcasting queries to multiple storage units simultaneously, and outputs the facial feature comparison results.

[0084] Furthermore, the system is also used to implement the following functions: When the facial feature comparison result in the second processing result hits the preset high-risk label, a hardware preemption signal is generated; in response to the hardware preemption signal, the three-dimensional spatial coordinate data and continuous action semantic identifier associated with the target person are hot-migrated from the second type of storage cluster to the first type of storage cluster, and exclusive subarray computing resources are allocated.

[0085] Furthermore, the system is also used to implement the following functions: The left and right eye streams are registered using the binocular parallax principle, and the three-dimensional bounding box distance between the person and the charged object is calculated to generate a three-dimensional protective space label. Action intent semantic labels are obtained by performing temporal convolutional encoding on the coordinate sequence of key points of the human skeleton in consecutive frames. Facial features of the person entering the monitoring area are extracted and hashed using the face capture module in the binocular vision module to generate a facial identity identifier. The three-dimensional protective space label, action intent semantic label, and facial identity identifier are summarized to obtain a spatial coordinate label stream.

[0086] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Specific embodiments of this specification have been described above. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0087] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0088] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A data stream scheduling method for controlled spheres based on spatial coordinate labels, characterized in that, The control ball integrates a binocular vision module and a near-field processing memory, and the method includes: An asymmetric address mapping architecture is constructed in the near-memory processing memory, wherein the asymmetric address mapping architecture includes a first type of storage cluster and a second type of storage cluster; Target detection is performed by a binocular vision module, a spatial coordinate label stream is generated, and the spatial coordinate label stream is sent to the near-memory processing memory. The spatial coordinate label stream includes the stereo spatial coordinate data of the target person, continuous action semantic identifiers, and facial identity identifiers. The spatial coordinate label stream is processed in parallel by the first type of storage cluster and the second type of storage cluster using a near-memory processing in-memory asymmetric address mapping architecture to obtain the first processing result and the second processing result. Based on the first and second processing results, the mapping level of the spatial coordinate label stream of the target personnel in the asymmetric address mapping architecture is dynamically adjusted.

2. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 1, characterized in that, The first type of storage cluster consists of physical storage blocks capable of parallel computing within subarrays; The second type of storage cluster consists of physical storage blocks that have the ability to perform parallel retrieval across storage banks; The first type of storage cluster and the second type of storage cluster correspond to different physical address ranges.

3. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 1, characterized in that, Parallel processing of the spatial coordinate label stream is performed on the first and second type of storage clusters using a near-memory processing in-memory asymmetric address mapping architecture to obtain a first processing result and a second processing result, including: The spatial coordinate label stream is parsed using hot-triggered semantic identifiers, wherein the hot-triggered semantic identifiers are used to characterize whether the cumulative trigger frequency of continuous action semantic identifiers within a preset sliding time window exceeds a preset cumulative action trigger frequency threshold. When the hot-triggered semantic identifier is identified, the coordinate label stream containing the continuous action semantic identifier is scheduled to the first type of storage cluster, the sub-array level near-memory pre-merging logic is activated, bit-domain differential processing is performed on the three-dimensional spatial coordinate data of the continuous frames, and the obtained action residual feature map is used as the first processing result. The spatial coordinate label stream containing the facial identity identifier is stored in the second type of storage cluster, and the parallel retrieval logic across storage is scheduled to perform parallel matching on the facial identity identifier. The obtained facial feature comparison result is used as the second processing result.

4. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 3, characterized in that, Hot-triggered semantic tagging parsing of the spatial coordinate tag stream includes: The intelligent processing module of the control ball performs temporal convolutional encoding on the human skeleton key point coordinate sequence of the continuous frames corresponding to the spatial coordinate label stream to obtain continuous action semantic identifiers, wherein the continuous action semantic identifiers include climbing semantic encoding, falling semantic encoding or arc discharge associated posture semantic encoding. A trigger frequency counter associated with a preset three-dimensional protective space grid is set in the near-memory processing memory. The trigger frequency counter is used to count the cumulative number of occurrences of various continuous action semantic identifiers in each space grid unit within a preset sliding time window. When the cumulative triggering frequency of a specific continuous action semantic identifier within a spatial grid cell exceeds a preset cumulative triggering frequency threshold, a hot-trigger semantic identifier carrying the coordinate information of that grid cell is generated.

5. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 3, characterized in that, When the hot-triggered semantic identifier is identified, the coordinate label stream containing the continuous action semantic identifier is scheduled to the first type of storage cluster, the subarray-level near-memory pre-merge logic is activated, bit-domain differential processing is performed on the three-dimensional spatial coordinate data of the continuous frames, and the obtained action residual feature map is used as the first processing result, including: The first type of storage cluster includes multiple logical partitions, wherein each logical partition corresponds to a preset three-dimensional protective space grid within the monitoring field of view of the surveillance ball; The three-dimensional spatial coordinate data of the spatial coordinate label stream is parsed to obtain the real-time three-dimensional trajectory point sequence of the target person in the binocular parallax coordinate system; Based on the real-time three-dimensional trajectory point sequence, regional hit matching is performed in the preset three-dimensional protective space grid to obtain the matching hot spot area identifier; Load the memory scheduling and operator control strategy corresponding to the matching hotspot region identifier, activate the subarray parallel sensitive amplifier, store the three-dimensional spatial coordinate data of consecutive frames in the same grid cell in different rows of the same subarray, complete the displacement difference calculation of the coordinate vectors of adjacent frames in the analog domain through the local bit line, and output the action residual feature map.

6. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 5, characterized in that, The pre-set three-dimensional protective space grid includes the space grid at the top of the poles in the live working area, the space grid near the transformer, and the space grid for high-voltage cable routing.

7. The data stream scheduling method for deployment balls based on spatial coordinate labels as described in claim 3, characterized in that, The spatial coordinate label stream containing the facial identity identifier is stored in the second type of storage cluster, and parallel retrieval logic across storage volumes is scheduled to perform parallel matching on the facial identity identifier. The obtained facial feature comparison result is used as the second processing result, including: The hardware scheduling engine distributes the facial identification identifiers to multiple independent storage blocks in the second type of storage cluster. The second type of storage cluster uses a multi-storage bank simultaneous broadcast query method to perform parallel matching between the facial identity identifier to be identified and the pre-stored high-risk personnel facial database, and outputs the facial feature comparison result.

8. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 1, characterized in that, Based on the first and second processing results, the mapping level of the target personnel's spatial coordinate label stream in the asymmetric address mapping architecture is dynamically adjusted, including: When the facial feature comparison result in the second processing result matches the preset high-risk label, a hardware preemption signal is generated. In response to the hardware preemption signal, the three-dimensional spatial coordinate data and continuous action semantic identifier associated with the target person are hot-migrated from the second type of storage cluster to the first type of storage cluster, and dedicated subarray computing resources are allocated.

9. The data stream scheduling method for a controlled sphere based on spatial coordinate labels as described in claim 1, characterized in that, Object detection is performed using a binocular vision module, generating a spatial coordinate label stream, including: By registering the left and right eye streams using the principle of binocular parallax, calculating the three-dimensional bounding box distance between the personnel and the live conductor, and generating a three-dimensional protective space label; Action intent semantic tags are obtained by performing temporal convolutional encoding on the sequence of key point coordinates of human skeleton in consecutive frames. The face capture module in the binocular vision module extracts and hashes the facial features of people entering the monitored area to generate facial identity identifiers. The spatial coordinate label stream is obtained by summarizing the three-dimensional protective space label, action intent semantic label, and facial identity identifier.

10. A data stream scheduling system for a controlled sphere based on spatial coordinate labels, characterized in that, The system is used to implement the deployment ball data stream scheduling method based on spatial coordinate label driving according to any one of claims 1-9, and the system includes: A mapping architecture construction module is used to construct an asymmetric address mapping architecture in the near-memory processing memory, wherein the asymmetric address mapping architecture includes a first type of storage cluster and a second type of storage cluster; The spatial coordinate label stream generation module is used to perform target detection through a binocular vision module, generate a spatial coordinate label stream, and send the spatial coordinate label stream to the near-memory processing memory. The spatial coordinate label stream includes the stereo spatial coordinate data of the target person, continuous action semantic identifiers, and facial identity identifiers. The parallel processing execution module is used to perform parallel processing on the spatial coordinate label stream through a first type of storage cluster and a second type of storage cluster in a near-memory processing in-memory asymmetric address mapping architecture to obtain a first processing result and a second processing result. The mapping level adjustment module is used to dynamically adjust the mapping level of the spatial coordinate label stream of the target person in the asymmetric address mapping architecture based on the first processing result and the second processing result.

Images