End-edge-cloud multi-scale spatio-temporal semantic state representation method based on brain-like bionic mapping

Abstract

This invention discloses a multi-scale spatiotemporal semantic object state representation method based on brain-inspired biomimetic mapping, belonging to the field of brain-inspired computing technology. Addressing the problems of cross-level semantic abstraction fragmentation, low efficiency of topological association, and semantic redundancy accumulation in city-level video swarm intelligence perception, this invention constructs a hierarchical mapping architecture from edge to middle layer to cloud, sequentially performing progressive abstraction representations from pixel level to individual level, group level, and city level. A multi-dimensional weighted sensor network topology model is constructed, generating a full-cycle continuous spatiotemporal state vector through cross-sensor association matching and trajectory stitching, and completing global optimization and deduplication of cross-camera samples. Simultaneous parallel encoding of metadata and feature streams is performed for swarm intelligence semantic compactness, globally deduplicating and differentially encoding duplicate features and redundant samples. This invention achieves efficient multi-scale spatiotemporal semantic expression through edge-cloud collaboration, significantly reducing transmission and storage overhead and improving the accuracy of cross-regional long-term trajectory stitching.

Description

Technical Field

[0001] This invention relates to the field of neuromorphic computing technology, and in particular to a method for representing multi-scale spatiotemporal semantic states of edge-cloud based on neuromorphic biomimetic mapping. Background Technology

[0002] The rapid development of smart cities and digital twins has propelled city-level video swarm intelligence sensing systems to become a core support for scenarios such as traffic management and public security. In a city-wide video swarm intelligence sensing system that integrates edge, cloud, and mobile devices, it is necessary to represent the state of semantic objects (such as people and vehicles) in the video data collected by massive amounts of cameras across time, space, and devices to support upper-layer applications such as target retrieval, trajectory tracking, and situational awareness analysis.

[0003] Currently, mainstream semantic object state representation methods are based on deep learning video structuring techniques. Their basic characteristics are: frame-by-frame target detection and tracking of a single video stream, generating target bounding box sequences and local trajectories, and extracting deep features as object representation vectors; in cross-camera scenarios, re-identification techniques are introduced, and local trajectories are concatenated through feature matching and simple spatial constraints to form a behavior description with a longer time scale. The output of these methods is mainly a discrete set of trajectory points or a fixed-dimensional feature vector.

[0004] However, the above methods have significant limitations when adapting to a hierarchical architecture of edge-cloud: First, there is a lack of semantic abstraction across different levels: optimization is only performed for a single scenario or local area, and there is no unified spatiotemporal representation framework that spans the edge, cloud, and device. This makes it impossible to achieve seamless semantic transformation from micro-level target behavior to macro-level regional situation, and it is difficult to adapt to heterogeneous computing power and hierarchical task requirements.

[0005] Second, large-scale topology association is inefficient: relying on static rules or simple topology constraints to characterize camera relationships cannot accurately describe the complex physical space and logical relationships within the metropolitan area, resulting in poor accuracy of cross-regional and long-term target tracking.

[0006] Third, insufficient system-level redundancy suppression: data compression only targets the storage or transmission of single-channel video, and does not combine semantic understanding from the perspective of edge-cloud collaboration to perform global redundancy reduction across sensors and time and space, resulting in huge overall computation, storage and transmission overhead.

[0007] Neuroscience research shows that the human visual system possesses extremely high information processing and transmission efficiency. The ratio of retinal photoreceptor cells to optic nerve fibers is approximately 126:1. This system achieves efficient information transmission not by simply compressing the original visual signal at the pixel level, but by following a hierarchical perception and progressive abstraction mechanism of the retina-lateral geniculate nucleus-visual cortex: the retina is responsible for acquiring the original visual signal and completing primary feature extraction and signal filtering; the lateral geniculate nucleus undertakes intermediate feature transmission, noise suppression, and information relay functions; and the visual cortex integrates, performs high-level abstraction, and semantic interpretation of multi-scale and multi-level visual features. Through the progressive extraction of multi-level visual features, the elimination of redundant information, and compact expression, high-fidelity visual information cognition is achieved with extremely low transmission overhead.

[0008] Inspired by this hierarchical perception mechanism, we explore a multi-scale spatiotemporal representation strategy based on brain-inspired biomimetic mapping at the edge-cloud level. This strategy is expected to break through the energy efficiency bottleneck of the existing architecture and further improve the system efficiency and perception accuracy of city-level massive video swarm intelligence perception. Summary of the Invention

[0009] The purpose of this invention is to provide a multi-scale spatiotemporal semantic state representation method for edge-cloud based on brain-inspired biomimetic mapping. Through hierarchical abstraction, weighted topological association, and dual-path compact coding, it achieves progressive and efficient expression of semantic information and systematic reduction of resource overhead in city-level swarm intelligence sensing systems.

[0010] To achieve the above objectives, this invention provides a method for representing multi-scale spatiotemporal semantic states of edge-cloud based on brain-inspired biomimetic mapping, the steps of which are as follows: S1. Construct an edge-cloud hierarchical mapping architecture based on the hierarchical perception mechanism of the human visual system, mapping the edge to the retinal layer, the edge to the optic nerve layer, the middle layer to the lateral geniculate nucleus layer, and the cloud to the visual cortex layer; Under the edge-cloud hierarchical mapping architecture, perform multi-scale progressive abstraction representation of spatiotemporal semantic objects in the hierarchical order of edge, middle, and cloud, from pixel level to individual level, group level, and city level. S2. Construct a multi-dimensional weighted sensor network topology model, perform correlation matching and trajectory stitching on target feature data across sensors based on the topology model, generate a continuous spatiotemporal state vector of individual targets throughout the entire cycle, and complete the global optimization and redundancy removal of target samples across cameras based on semantic information maximization constraints. S3. Perform cross-level collective intelligent semantic compact coding, construct a dual-path parallel coding system of metadata stream and feature stream, perform global deduplication processing on cross-camera duplicate features and redundant samples of the same target, and perform differentiated coding according to the semantic value of the target and the priority of the task.

[0011] Preferably, in step S1, the specific processing procedure on the end side is as follows: target detection and tracking processing is performed on the input raw video stream to complete the spatiotemporal state detection of the target within a single sensor, outputting the target bounding box sequence, spatiotemporal coordinate data, multi-dimensional feature vectors and hierarchical semantic labels, and the target images within a single video stream are selected and optimized based on the semantic information maximization constraint.

[0012] Preferably, in step S1, the specific processing procedure of the edge side is as follows: receiving target feature data reported by multiple end sides within the jurisdiction, performing feature matching and trajectory association on cross-sensor targets within the local area, completing target association matching and trajectory splicing processing, generating continuous spatiotemporal state vectors of individual targets, and outputting standardized spatiotemporal representation data of individual targets.

[0013] Preferably, in step S1, the specific processing procedures for the intermediate layer and the cloud are as follows: The intermediate layer receives spatiotemporal representation data of individual targets reported from the edge side, performs spatiotemporal feature fusion and progressive abstraction processing, aggregates discrete individual target trajectories into regional group activity patterns, and generates regional-level situation summaries and high-dimensional feature vectors. The cloud receives regional-level situation summaries reported from each intermediate layer, and fuses multi-regional situation data into a city-level semantic spatiotemporal knowledge graph through a preset graph construction function, completing cross-regional, long-term global spatiotemporal association.

[0014] Preferably, in step S2, the multi-dimensional weighted sensor network topology model is constructed as follows: Each sensor is treated as a topology node, and metadata such as the sensor's physical location, orientation, and field of view is recorded. A weighted directed topology graph is constructed based on the sensor nodes, and physical distance weight, topological distance weight, and historical transition probability weight are calculated for the connecting edges between nodes. A topological adjacency matrix is ​​generated based on the three types of weights to quantify the spatiotemporal correlation strength between different sensors.

[0015] Preferably, step S2, cross-sensor target association matching and trajectory stitching, specifically includes: Based on the target's deep features and semantic labels, feature similarity is calculated to generate an initial matching candidate set. Temporal continuity constraints, spatial rationality constraints, and data quality constraints are introduced to filter the initial matching candidate set and filter out false matching results. The target's full-cycle trajectory sequence is linked through a multi-sensor association algorithm to update the target's continuous spatiotemporal state vector in real time.

[0016] Preferably, the dual-path parallel coding system in step S3 specifically includes: The metadata stream branch performs optimization filtering and prior encoding processing on the structured metadata of target attributes and spatiotemporal coordinates to generate a standardized metadata stream; the feature stream branch performs alignment fusion and lossless encoding processing on the target's deep features, structured features, manual features, and dynamic behavior features to generate a standardized feature stream.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0018] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) A multi-scale spatiotemporal semantic progressive representation architecture for edge-cloud collaboration was constructed. The edge side, cloud side, and cloud side were mapped to the retinal layer, optic nerve layer, and visual cortex layer, respectively, realizing a hierarchical abstraction from pixel-level data to city-level semantic situation. This architecture enables the intermediate features extracted from the edge side to be efficiently reused in the cloud, avoiding redundant calculations and transmissions, and significantly improving the collaborative efficiency of spatiotemporal semantic representation under the edge-cloud hierarchical architecture.

[0019] (2) A multi-dimensional weighted sensor network topology model was established, which improved the accuracy of cross-sensor spatiotemporal correlation. The dynamic correlation strength between sensors was quantified by combining three types of weights: physical distance, topological distance, and historical transition probability. False matches were filtered out by combining three types of constraints: time, space, and data quality. Compared with existing static rule-based methods, this scheme significantly improves the robustness of cross-regional long-term trajectory stitching under complex metropolitan-level camera layouts.

[0020] (3) A cross-level dual-path parallel swarm intelligence semantic compact coding method is proposed. By using dual-path coding of metadata stream and feature stream, duplicate features of the same semantic target generated by multiple cameras and multiple spatiotemporal locations are globally deduplicated, and differentiated coding is performed according to semantic value and task priority. This mechanism significantly reduces transmission and storage overhead while preserving core semantic information, effectively alleviating the systemic bottleneck of massive video data being "unable to be transmitted, stored, or computed".

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1This is a schematic diagram of the edge-cloud hierarchical mapping architecture according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the sensor network topology according to an embodiment of the present invention; Figure 3 is a schematic diagram of the multi-dimensional redundancy of the video group spatiotemporal sensor according to an embodiment of the present invention, wherein (a) is a preferred semantic target image of a single sensor; and (b) is a preferred semantic image across sensors. Figure 4 This is a schematic diagram illustrating the spatiotemporal convergence of semantic target information according to an embodiment of the present invention; Figure 5 This is a schematic diagram of compact encoding of video crowd intelligence semantic targets according to an embodiment of the present invention; Figure 6 This is a multi-sensor target association algorithm according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Example Based on typical application scenarios of the urban brain eye-brain perception model, this invention provides a detailed implementation description of the edge-cloud multi-scale spatiotemporal semantic state representation method based on brain-like biomimetic mapping.

[0027] S1. Construct an edge-cloud hierarchical mapping architecture based on the hierarchical perception mechanism of the human visual system, mapping the edge to the retinal layer, the edge to the optic nerve layer, the middle layer to the lateral geniculate nucleus layer, and the cloud to the visual cortex layer. Under the edge-cloud hierarchical mapping architecture, perform multi-scale progressive abstraction representation of spatiotemporal semantic objects in the order of edge, middle, and cloud, from pixel level to individual level, group level, and city level.

[0028] like Figure 1 As shown, the layered processing procedure is as follows: (1) On the end side ( The model performs initial signal-level compaction on the original multi-source data. (Including video, audio, RFID, environmental parameters, etc.) are encoded using a set of parallel standardized coding functions. Processing, such as H.26x encoding of video streams, adaptive sampling and encapsulation of environmental sensor data, and outputting a deduplicated set of source data. .

[0029] (2) Data arrives at the edge ( After that, cross-modal primary semantic extraction is performed to convert heterogeneous data into a unified semantic metadata format (visual data). →Target bounding box and identification sequence Audio data → Description of sound events Scalar data → Status and anomaly markers Here, a cross-modal fusion operator is introduced. It is responsible for deeply fusing the high-level features of each modality, represented as:

[0030] For visual encoding functions, For audio encoding functions; via Operators and models can uncover the intrinsic connections between different sensory information.

[0031] (3) In the intermediate layer ( The core task of the model is to use hierarchical feature fusion operators. Spatiotemporal feature fusion and progressive abstraction are represented as follows:

[0032] The output evolves from discrete cross-sensor target trajectories to overall group activity patterns, ultimately forming a regional situation summary, which is usually represented in the form of high-dimensional feature vectors, realizing the leap of information from individual to group and from micro to macro.

[0033] (4) Finally, in the city brain cloud ( All the collected regional situation summaries are constructed using a graph construction function. It was synthesized into a city-level semantic spatiotemporal knowledge graph. This map can be formally represented as:

[0034] In this triplet, It is a physical node. It is a relation edge. It is a dynamic set of attributes. It is no longer just a collection of data, but an insight into the dynamic operation patterns of the entire city, and the highest degree of condensation of raw petabyte-level observation data.

[0035] This establishes a continuous [system / mechanism]. This is a layered computing architecture for a progressive, multi-level, compact feature representation and abstraction mechanism for multi-source heterogeneous data. The core idea of ​​this model is that as data flows from bottom to top, the dimensionality and redundancy of its representation decrease progressively, while the semantic concentration and abstraction level of the information increase progressively. This process achieves a fundamental transformation from a "sea of ​​pixels" to a "forest of semantic objects," and then to a "single map of urban conditions," aiming to systematically solve the bottleneck problems of "inability to transmit, store, and compute" caused by massive amounts of data.

[0036] S2. Construct a multi-dimensional weighted sensor network topology model. Based on this topology model, perform correlation matching and trajectory stitching on target feature data across sensors to generate a continuous spatiotemporal state vector of individual targets throughout their entire lifecycle. Then, based on semantic information maximization constraints, complete the global optimization and redundancy removal of target samples across cameras. Specifically, this includes: (1) Sensor network topology modeling like Figure 2 As shown, in order to achieve cross-sensor continuous tracking and behavior prediction of semantic targets, it is necessary to construct an adjacent sensor network topology graph (weighted directed graph) to digitize the spatial and logical relationships of the physical world, so that the system can understand which adjacent sensor B is most likely to appear in the field of view after a target disappears from the field of view of sensor A.

[0037] Within the same cluster, each sensor is treated as a node, and each node records its own metadata: physical location (GPS coordinates, floor, building number, etc.), orientation, field of view, etc. Weights are used to quantify the "strength" or "probability" of a connection between two sensors, and are derived from three different weighting factors.

[0038] Physical distance weight ( ): The physical distance between the field-of-view boundaries of the two sensors; the closer the distance, the higher the weight. Topological distance weight ( In building structures, the fewer the channels and access points a sensor needs to pass through to get from one sensor to another, the higher its weight. Transition probability weights ( ): Empirical probability obtained through statistical analysis of historical trajectory data. Based on this, using... Represents the sensor network topology, where It is an edge set. It is an adjacency matrix (all elements are weights); V Let be the vertex set, represented as: ,in This represents the number of sensors.

[0039] (2) Target spatiotemporal state detection Single-sensor target status: Employing the industry's latest target detection and tracking algorithms, each data stream is processed in real time, outputting a sequence of semantic target bounding boxes. Track the time / location of the target in the current sensor. / ), disappearance time / location ( / For each target (unique) Extracting four core features: deep features ( ), structural features ( ), handcraft characteristics ( ), dynamic behavioral characteristics ( (e.g., key point sequences) constitute the target feature set. Then, hierarchical semantic labels are assigned to each target. Clearly define the target type and attributes. Simultaneously, consider the confidence level of semantic target detection. Semantic target ambiguity within a semantic rectangle Semantic integrity parameters are defined based on the degree of occlusion exposure and the integrity of skeleton key points. Average relevance of adjacent semantic target images Based on four parameters , , , Based on the constraints of maximizing semantic information preservation and maximizing average relevance, frontal, back, and side semantic target images are selected from the semantic target bounding box image sequence within each video stream. The output includes a preferred image set, an auxiliary image set (with different lighting and poses), and target feature vectors. These calculations are deployed on edge intelligent servers or on intelligent computing platforms built into front-end sensors.

[0040] Cross-sensor target state: Deploy a cross-sensor semantic target optimization algorithm module on the swarm intelligence computing server to optimize the target state within a certain range of physical space. Correlation analysis is performed on neighboring sensor data. Based on depth features ( ) and semantic tags ( ), calculate the feature correlation of the target under different sensors. Introducing time continuity ( ), spatial rationality ( ), quality constraints ), filtering out false matches. A multi-sensor target association algorithm links trajectory sequences, recording the sensor position at the previous moment. And the sensor at the next moment (Based on movement speed and trajectory trend prediction), specifically as follows: Figure 6Algorithm 1 is shown in the diagram. Simultaneously, based on the constraints of maximizing semantic information preservation and maximizing average relevance, frontal, back, and side semantic images (with optimal labeling) are selected from the cross-sensor tracking trajectory according to the constraints. Update The image selection process is shown in Figure 3.

[0041] (3) Target spatiotemporal trajectory link like Figure 4 As shown, the scale is defined. describe Range, time window Inner Semantic target events for time intervals , Corresponding timestamp ,in and These are the indices and total number corresponding to the discretized time window. Assume that in... There are a total of A semantic goal, the semantic goal is used express, Each discrete time point needs to record all The spatiotemporal state of a semantic target.

[0042] Let the first semantic goals In time The spatial location is (query) (Based on calculations using the BeiDou code), its spatiotemporal state vector is defined as: The spatiotemporal states of all semantic targets at all discrete time points constitute a two-dimensional set (target dimension × time dimension). This set fully describes the scale. Down, A semantic target in [ , The spacetime trajectory within ]

[0043] S3. Perform cross-level collective intelligent semantic compact encoding, construct a dual-path parallel encoding system for metadata stream and feature stream, perform global deduplication processing on duplicate features and redundant samples of the same target across cameras, and perform differentiated encoding based on the target semantic value and task priority. Specifically: Crowdsourced videos exhibit significant redundancy in semantic information across sensors, such as... Figure 5As shown, in the application of crowd intelligence in urban brains, maximizing the compression of cross-sensor semantic redundancy is key to further reducing computational, storage, and transmission costs. This study draws on the 126-fold bandwidth attenuation mechanism of the HVS visual neural channel to explore an effective method for compact encoding of semantic targets from the perspective of crowd intelligence perception. After compact encoding and compression, the data is aggregated in the cloud and used to realize crowd intelligence perception services in the cloud brain.

[0044] Considering the different needs of machine vision tasks, some applications may require target object behavior analysis or sentiment analysis. Drawing on the HVS eye-brain perception structure, this project plans to adopt optimized semantic metadata and deep features. ), structural features ( ), handcraft characteristics ( ), dynamic behavioral characteristics ( This model employs a multi-layered, compact feature representation method. Specifically, it starts with multi-modal inputs in parallel streams. Each input first passes through a feature extraction module to generate a compressed data stream, a condensed summary stream, and structured features. ), unstructured features ( , , Four types of intermediate features are processed. Data from the condensed summary stream enters the metadata stream processing branch, where effective metadata is filtered by the metadata optimization module and then a standardized metadata stream is generated through prior encoding. Structured and unstructured features enter the feature stream processing branch, where multi-source features undergo spatiotemporal / semantic alignment and fusion through the alignment and fusion module, and a standardized feature stream is generated through lossless encoding. The metadata stream and feature stream then flexibly call features of corresponding dimensions according to different level requirements, and then input them into the spatial-temporal-semantic target fusion analysis engine. Combining spatial geographic location, temporal series changes, and semantic target correlations such as crowds / vehicles, multi-dimensional fusion calculations are performed, and finally, scenario-based situational analysis results such as crowd gathering and traffic conditions are output.

[0045] The entire process forms a closed loop of "multimodal input → hierarchical feature extraction → dual-stream (metadata / feature) standardization processing → hierarchical feature reuse → spatiotemporal-semantic fusion → situational result output". The core logic is to achieve intelligent perception and decision output of complex scenarios through the refined processing of multi-source features and multi-dimensional correlation analysis.

[0046] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for representing the multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping, characterized in that, The steps are as follows: S1. Construct an edge-cloud hierarchical mapping architecture based on the hierarchical perception mechanism of the human visual system, mapping the edge to the retinal layer, the edge to the optic nerve layer, the middle layer to the lateral geniculate nucleus layer, and the cloud to the visual cortex layer; Under the edge-cloud hierarchical mapping architecture, perform multi-scale progressive abstraction representation of spatiotemporal semantic objects in the hierarchical order of edge, middle, and cloud, from pixel level to individual level, group level, and city level. S2. Construct a multi-dimensional weighted sensor network topology model, perform correlation matching and trajectory stitching on target feature data across sensors based on the topology model, generate a continuous spatiotemporal state vector of individual targets throughout the entire cycle, and complete the global optimization and redundancy removal of target samples across cameras based on semantic information maximization constraints. S3. Perform cross-level collective intelligent semantic compact coding, construct a dual-path parallel coding system of metadata stream and feature stream, perform global deduplication processing on cross-camera duplicate features and redundant samples of the same target, and perform differentiated coding according to the semantic value of the target and the priority of the task.

2. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 1, characterized in that: In step S1, the specific processing procedure on the edge is as follows: target detection and tracking processing is performed on the input raw video stream to complete the spatiotemporal state detection of the target within a single sensor, output the target bounding box sequence, spatiotemporal coordinate data, multi-dimensional feature vector and hierarchical semantic label, and the target images in a single video stream are selected and optimized based on the semantic information maximization constraint.

3. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 2, characterized in that: In step S1, the specific processing procedure on the edge side is as follows: receiving target feature data reported by multiple end sides within the jurisdiction, performing feature matching and trajectory association on cross-sensor targets within the local area, completing target association matching and trajectory stitching processing, generating continuous spatiotemporal state vectors of individual targets, and outputting standardized spatiotemporal representation data of individual targets.

4. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 1, characterized in that: In step S1, the specific processing procedures for the intermediate layer and the cloud are as follows: The intermediate layer receives spatiotemporal representation data of individual targets reported from the edge side, performs spatiotemporal feature fusion and progressive abstraction processing, aggregates discrete individual target trajectories into regional group activity patterns, and generates regional situation summaries and high-dimensional feature vectors. The cloud receives regional situation summaries reported by each intermediate layer and integrates multi-regional situation data into a city-level semantic spatiotemporal knowledge graph through a preset graph construction function, thus completing cross-regional, long-term global spatiotemporal association.

5. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 1, characterized in that: In step S2, the multi-dimensional weighted sensor network topology model is constructed as follows: Each sensor is treated as a topology node, and metadata such as the sensor's physical location, orientation, and field of view is recorded. A weighted directed topology graph is constructed based on sensor nodes. Physical distance weight, topological distance weight, and historical transition probability weight are calculated for the connecting edges between nodes. A topological adjacency matrix is ​​generated based on the three types of weights to quantify the spatiotemporal correlation strength between different sensors.

6. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 1, characterized in that: Step S2, cross-sensor target association matching and trajectory stitching, specifically includes: Based on the target's deep features and semantic labels, feature similarity is calculated to generate an initial matching candidate set. Temporal continuity constraints, spatial rationality constraints, and data quality constraints are introduced to filter the initial matching candidate set and filter out false matching results. The target's full-cycle trajectory sequence is linked through a multi-sensor association algorithm to update the target's continuous spatiotemporal state vector in real time.

7. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 1, characterized in that: The dual-path parallel coding system in step S3 specifically includes: The metadata stream branch performs optimization filtering and prior encoding processing on the structured metadata of target attributes and spatiotemporal coordinates to generate a standardized metadata stream; the feature stream branch performs alignment fusion and lossless encoding processing on the target's deep features, structured features, manual features, and dynamic behavior features to generate a standardized feature stream.

8. The method for representing multi-scale spatiotemporal semantic state of edge-cloud based on brain-inspired biomimetic mapping according to claim 2, characterized in that: The specific method for the terminal to optimize the target image within a single video stream is as follows: The algorithm integrates four metrics: semantic target detection confidence, semantic target ambiguity, semantic completeness, and average relevance of adjacent semantic target images. It selects frontal, back, and side semantic target images from the semantic target bounding box image sequence within each video stream and outputs a preferred image set, an auxiliary image set with different lighting and poses, and target feature vectors.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images