A smart garden safety management and control scheduling system based on multi-modal perception and GIS

Abstract

The present application relates to a kind of based on multi-modal perception and GIS's wisdom garden safety control scheduling system, belong to wisdom city technical field.Wherein, the system includes: geographic information space framework;Management garden vector element, real-time sensing data, resource point and dynamic analysis result are spatialized fusion superposition and hierarchical rendering;Intelligent monitoring and analysis module, based on the real-time video stream of each area equipment of garden, carries out multi-target detection and behavior identification;Electricity situation analysis module, access the real-time power consumption data flow of garden power distribution internet of things terminal, generate the power consumption trend prediction result of preset period by time series prediction model;Dynamic maintenance module, for the incremental update of structured data and the structured operation data generated by garden operation and maintenance task. Realized the spatialization integration and visual presentation of multi-source garden safety data, provides integrated intelligent technical support for garden safety management, emergency dispatch.

Description

Technical Field

[0001] This invention belongs to the field of smart city technology, specifically relating to a smart garden safety management and scheduling system based on multimodal perception and GIS. Background Technology

[0002] With the advancement of smart city construction, park management is gradually transforming from the traditional manual inspection model to a digital and intelligent one. Currently, smart park management systems mainly rely on Geographic Information Systems (GIS) to achieve visualized management of park spatial data, deploy video surveillance equipment to achieve park security monitoring, use power distribution IoT terminals to collect energy consumption data, and rely on operation and maintenance work order systems to complete daily management tasks such as vegetation maintenance and facility repair.

[0003] In the field of intelligent video analytics, deep learning-based target detection and behavior recognition technologies are relatively mature, capable of locating and classifying targets such as pedestrians and vehicles in surveillance footage, and identifying some abnormal behaviors. In the field of power load forecasting, time-series forecasting models (such as ARIMA and Prophet) are widely used in power grid load forecasting, capable of extracting periodic trends based on historical electricity consumption data. In the field of GIS applications, vector base map engines, dynamic layer overlay, and spatial indexing technologies have become the fundamental support for spatial data management.

[0004] However, the application of these technologies in smart garden scenarios remains relatively fragmented: GIS systems are merely used as static base map display tools; video surveillance systems operate independently of GIS and their analysis results remain at the pixel level; energy consumption monitoring systems lack correlation analysis with spatial location and pedestrian activity; and the linkage update between the operation and maintenance management system and the GIS layer relies on manual triggering. The inconsistent data formats and spatiotemporal benchmarks among the subsystems make it difficult to form a unified spatial situational awareness capability for garden safety management and operation and maintenance scheduling.

[0005] In current smart garden construction, GIS platforms, video surveillance systems, energy consumption monitoring systems, and operation and maintenance management systems are typically built independently by different vendors, using their own data formats, communication protocols, and coordinate systems. Video analysis results cannot be automatically mapped to the GIS spatial framework for positioning and overlay, electricity consumption data lacks spatial correlation with garden functional zoning, and operation and maintenance records are disconnected from GIS layer status updates. This forces managers to switch between multiple systems to view data, making it difficult to form a comprehensive situational awareness with a single map. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a smart garden safety management and scheduling system based on multimodal perception and GIS. The objective of this invention can be achieved through the following technical solutions: A smart garden safety management and scheduling system based on multimodal perception and GIS includes: The geographic information spatial framework consists of a vector base map engine, a dynamic layer overlay engine, and a spatiotemporal database. The vector base map engine is used to manage vector elements such as garden functional zoning, road and water system, vegetation distribution, facility locations, and pipeline routes. The dynamic layer overlay engine is used to spatially fuse, overlay, and hierarchically render real-time sensing data, resource locations, and dynamic analysis results. The spatiotemporal database is used to store structured sensing event streams. The intelligent monitoring and analysis module, embedded in the geographic information spatial framework, decodes real-time video streams from devices in various areas of the garden and inputs them into a multi-task spatiotemporal perception inference architecture for multi-target detection and behavior recognition. This generates a structured perception event stream containing event type, location coordinates, timestamp, and confidence level, which is then stored in the spatiotemporal database. Simultaneously, it drives the dynamic layer overlay engine on the geographic information spatial framework to overlay and visualize the structured perception event stream according to event type and spatial location, expanding it into a video monitoring function layer group. The electricity consumption situation analysis module accesses the real-time electricity consumption data stream from the garden power distribution IoT terminal, constructs a historical electricity consumption curve library divided by garden functional zones and time period granularity, and generates electricity consumption trend prediction results for a preset period by superimposing external regression variables through a time series prediction model; based on the electricity consumption trend prediction results, the dynamic layer overlay engine performs fusion situation rendering. The dynamic maintenance module has a built-in real-time update engine, which is used to incrementally update the structured data generated by the intelligent monitoring and analysis module and the power consumption status analysis module, as well as the structured operation and maintenance data generated by the garden operation and maintenance task operations initiated through the geographic information spatial framework.

[0007] Specifically, the vector basemap engine includes: The basic geographic data layer stores the baseline vector data of garden boundaries, elevation topography, road network and water system network, serving as the underlying reference for spatial positioning and path calculation; The garden-specific data layer overlays garden-specific vector elements, including vegetation distribution, facility locations, and pipeline routes. The vegetation distribution is coded by attribute according to species, crown width, tree age, and maintenance level; the facility locations are symbolically configured according to functional categories; and the pipeline routes are managed in layers according to media type and burial depth. The spatial indexing engine constructs a spatial index for the vector features, supporting the dynamic layer overlay engine in spatial positioning queries and neighborhood retrieval of real-time perceived events. The version management mechanism is used to record the change history of the vector elements and supports the dynamic maintenance module in detecting conflicts and merging and rolling back incremental update data.

[0008] Furthermore, the dynamic layer overlay engine includes: The layer manager is used to maintain the stacking order and visibility control of the base map layer, real-time perception layer, resource scheduling layer, and analysis decision layer. Each layer supports independent transparency adjustment and blending mode configuration. The spatiotemporal alignment submodule receives structured perception event streams from the intelligent monitoring and analysis module, electricity consumption trend prediction results from the electricity consumption situation analysis module, and multi-source heterogeneous data from external IoT sensors. Based on a unified time reference and geographic coordinate system, it performs spatiotemporal alignment and coordinate transformation, mapping each source data to the same spatial reference frame of the vector base map engine. The rendering pipeline is integrated and a hierarchical rendering strategy is adopted: real-time perceived events are mapped to differentiated visual symbols according to risk level, electricity consumption trend prediction results are mapped to thermal gradient surfaces according to load intensity, and resource scheduling status is mapped to dynamic path lines according to availability; transparency blending and depth sorting are supported when multiple layers are overlaid. The interactive response interface responds to user-side zooming, panning, selection, and timeline dragging operations, dynamically triggering the spatial index engine's view cropping query and the spatiotemporal database's time-series slice retrieval, thereby enabling real-time refreshing and historical backtracking of the situation layer.

[0009] Preferably, the spatial fusion and overlay uses the spatiotemporal alignment submodule to spatially associate and match the location coordinates in the structured perception event stream with the facility locations and pipeline directions in the garden thematic data layer. When the event coordinates fall within a preset buffer zone around the facility, the attribute identifier of the facility is automatically associated, forming an event-facility binding relationship and writing it into the fusion attribute table. The electricity consumption trend prediction results are spatially aggregated according to the garden functional zones to generate a zone load density raster, which is then overlaid onto the corresponding zone boundary.

[0010] Furthermore, the hierarchical rendering is a three-level rendering rule, specifically including: Feature-level rendering: For a single structured perception event, a preset symbol template is selected based on the event type, and the symbol opacity is adjusted based on the confidence level. Regional rendering generates an event aggregation heatmap for multiple events aggregated within the same spatiotemporal window, which is then overlaid on the boundary of the corresponding garden functional area. Situation-level rendering integrates the cross-modal outputs of the intelligent monitoring and analysis module and the power consumption situation analysis module. When visual anomalies and power consumption anomalies are simultaneously present in the same area, composite warning rendering is triggered, and a fused information card pops up to display the details of multimodal cross-verification.

[0011] Specifically, the multi-task spatiotemporal perception inference architecture extracts single-frame visual features based on a spatiotemporal feature encoder and models inter-frame motion correlations, and deploys multi-task decoding heads in parallel after the spatiotemporal feature encoder.

[0012] Furthermore, the multi-task decoding head includes: The target detection decoding head performs target localization and classification on the spatiotemporal feature map output by the spatiotemporal feature encoder, and outputs the category label, bounding box coordinates and detection confidence of the target in the frame; The behavior recognition decoding head performs temporal convolution and pooling operations on the same spatiotemporal feature map along the temporal dimension to generate the probability distribution of action categories and output the behavior classification results and corresponding confidence scores at the frame or segment level.

[0013] Preferably, the video monitoring function layer group includes: The pedestrian density heat map layer receives the pedestrian statistics output by the target detection decoding head in the multi-task spatiotemporal perception inference architecture, performs kernel density interpolation and grid aggregation according to the garden road and square areas, generates a pedestrian density heat map surface, and superimposes it on the basic geographic data layer of the vector base map engine. The target recognition annotation layer is connected to the real-time video stream of the equipment in various areas of the garden through the geographic information spatial framework. Using the video screen as the base map, the intra-frame target category label, bounding box coordinates and detection confidence output by the target detection decoding head in the multi-task spatiotemporal perception inference architecture are superimposed and drawn in real time on the spatial position of the corresponding video screen to form the target detection annotation box and category text label. The non-safe behavior warning layer receives the frame-level or segment-level behavior classification results and corresponding confidence levels output by the behavior recognition decoding head, and annotates the behavior classification results in real time with differentiated visual symbols according to behavior categories; it generates a list of behavior analysis events in reverse chronological order in the sidebar of the interface, and each record in the list includes the behavior category name and the frame-level or segment-level time interval.

[0014] Specifically, the time-series prediction model includes: The historical electricity consumption curve library construction sub-module performs hierarchical aggregation and standardized storage of the real-time electricity consumption data stream according to the garden functional zoning and time period granularity, forming a multi-dimensional historical electricity consumption curve library. The baseline prediction sub-model, based on the historical electricity consumption curve library, extracts the periodic trend component and seasonal component of electricity load to generate a basic electricity consumption trend prediction curve that does not consider external factors. The external regression variable fusion unit collects and preprocesses external regression variables affecting electricity load, including meteorological data, holiday markers, garden activity schedules, and the zonal pedestrian density index in the pedestrian density heat map output by the intelligent monitoring and analysis module; it performs lag order analysis and correlation screening on the external regression variables, retaining the variable set that is significantly related to electricity load; The fusion prediction sub-model inputs the trend component output by the baseline prediction sub-model and the variable set filtered by the external regression variable fusion unit into the gradient boosting tree. The baseline prediction bias is corrected through the residual learning mechanism, and the fused electricity consumption trend prediction result is output. The fusion prediction sub-model supports online incremental training and automatically triggers model parameter updates when new observation data accumulates to a preset batch. The prediction result evaluation and correction unit calculates the root mean square error and mean absolute percentage error between the predicted electricity consumption trend and the actual observed value. When the error index exceeds the preset threshold, it triggers automatic adjustment of model hyperparameters or re-screening of external regression variables, and pushes the corrected prediction result to the dynamic layer overlay engine for fusion rendering.

[0015] Specifically, the fusion situation rendering of the power consumption situation analysis module includes: The load density rendering of the zones involves spatially aggregating the electricity consumption trend prediction results according to the functional zones of the garden, calculating the predicted load density value of each zone, generating a zone load density raster surface using contour lines or hierarchical coloring, and superimposing it on the corresponding zone boundary of the vector base map engine. The time-series comparison rendering overlays a comparison layer of the historical actual load curve and the current predicted load curve within the same view, embedded in the side analysis panel in the form of a dual-axis time series graph.

[0016] Specifically, the garden operation and maintenance tasks include at least changes in vegetation maintenance status, updates to facility status, and corrections to pipeline attributes.

[0017] Specifically, the incremental update and status monitoring include: After the real-time update engine detects the completion of the operation and maintenance task, it drives the dynamic layer overlay engine to refresh the corresponding vector element layer in real time, so that the status of the garden elements displayed on the GIS interface is synchronized with the operation and maintenance results.

[0018] The beneficial effects of this invention are as follows: By collaborating with a vector base map engine, a dynamic layer overlay engine, and a spatiotemporal database, video analysis events, power consumption prediction results, and operation and maintenance data are mapped to the same geographic coordinate system and time base, achieving comprehensive situational awareness and avoiding multi-system switching. The detection boxes and behavior recognition results output by the multi-task spatiotemporal awareness inference architecture are mapped to geographic coordinates via spatial positioning, driving real-time alert rendering and situational mapping. This supports neighborhood retrieval and spatial correlation analysis, unlocking the spatial decision-making value of intelligent video analysis. The spatiotemporal feature encoder extracts shared features, and the target detection and behavior recognition decoding heads are deployed in parallel, avoiding redundant computation, significantly reducing GPU memory usage and inference latency, and meeting the real-time analysis needs of multiple video streams.

[0019] Element-level symbols annotate individual events, region-level heatmaps display clustering patterns, and situation-level composite warnings provide cross-modal anomalies. Combined with transparency blending and interactive responses, this achieves hierarchical information expression and rapid identification of core risks. A real-time update engine captures events from completed maintenance, facility, and pipeline operations, instantly refreshing thematic layers. A version management mechanism supports change tracking and conflict detection, ensuring scheduling decisions are based on the latest data. Attached Figure Description

[0020] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of the system structure of a smart garden safety management and scheduling system based on multimodal perception and GIS according to the present invention; Figure 2 This is a flowchart illustrating the lifecycle management process for temporary tasks according to the present invention. Figure 3 This is a flowchart illustrating the multi-level operation and maintenance task plan of the present invention. Figure 4 This is a schematic diagram of the work order process in the area of ​​one river and two banks according to an embodiment of the present invention; Figure 5 This is a flowchart illustrating the external complaint assessment process of this invention. Figure 6 This is a schematic diagram of the safety inspection process of the present invention; Figure 7 This is a schematic diagram of the cross-departmental external work order processing flow according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the facility and equipment maintenance business processing flow of the present invention; Figure 9 This is a schematic diagram of the greening inspection business processing flow of the present invention; Figure 10 A schematic diagram of the facility maintenance business process from the perspective of the maintenance management department; Figure 11A schematic diagram of the greening inspection process from the perspective of the maintenance and management department. Detailed Implementation

[0022] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of example embodiments to those skilled in the art. Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring aspects of this disclosure. The blocks shown in the drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices. The flowcharts shown in the drawings are merely illustrative and do not necessarily include all contents and operations / steps, nor do they necessarily have to be performed in the order described. For example, some operations / steps can be broken down, while others can be combined or partially combined. Therefore, the actual execution order may change depending on the actual situation.

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0024] Please see Figure 1 A smart garden safety management and scheduling system based on multimodal perception and GIS includes: The geographic information spatial framework consists of a vector base map engine, a dynamic layer overlay engine, and a spatiotemporal database. The vector base map engine is used to manage vector elements such as garden functional zoning, road and water system, vegetation distribution, facility locations, and pipeline routes. The dynamic layer overlay engine is used to spatially fuse, overlay, and hierarchically render real-time sensing data, resource locations, and dynamic analysis results. The spatiotemporal database is used to store structured sensing event streams. The intelligent monitoring and analysis module, embedded in the geographic information spatial framework, decodes real-time video streams from devices in various areas of the garden and inputs them into a multi-task spatiotemporal perception inference architecture for multi-target detection and behavior recognition. This generates a structured perception event stream containing event type, location coordinates, timestamp, and confidence level, which is then stored in the spatiotemporal database. Simultaneously, it drives the dynamic layer overlay engine on the geographic information spatial framework to overlay and visualize the structured perception event stream according to event type and spatial location, expanding it into a video monitoring function layer group. The electricity consumption situation analysis module accesses the real-time electricity consumption data stream from the garden power distribution IoT terminal, constructs a historical electricity consumption curve library divided by garden functional zones and time period granularity, and generates electricity consumption trend prediction results for a preset period by superimposing external regression variables through a time series prediction model; based on the electricity consumption trend prediction results, the dynamic layer overlay engine performs fusion situation rendering. The dynamic maintenance module has a built-in real-time update engine, which is used to incrementally update the structured data generated by the intelligent monitoring and analysis module and the power consumption status analysis module, as well as the structured operation and maintenance data generated by the garden operation and maintenance task operations initiated through the geographic information spatial framework.

[0025] Specifically, the vector basemap engine includes: The basic geographic data layer stores the baseline vector data of garden boundaries, elevation topography, road network and water system network, serving as the underlying reference for spatial positioning and path calculation; The garden-specific data layer overlays garden-specific vector elements, including vegetation distribution, facility locations, and pipeline routes. The vegetation distribution is coded by attribute according to species, crown width, tree age, and maintenance level; the facility locations are symbolically configured according to functional categories; and the pipeline routes are managed in layers according to media type and burial depth. The spatial indexing engine constructs a spatial index for the vector features, supporting the dynamic layer overlay engine in spatial positioning queries and neighborhood retrieval of real-time perceived events. The version management mechanism is used to record the change history of the vector elements and supports the dynamic maintenance module in detecting conflicts and merging and rolling back incremental update data.

[0026] In this embodiment, the basic geographic data layer uses a topographic map as the base map reference. Spatial data of park boundaries, elevation topography, road network, and water system network are acquired through UAV oblique photogrammetry, and after aerial triangulation and orthorectification, they are imported into the vector base map engine.

[0027] The park boundary is stored using polygonal elements, including the park's red line and the sub-boundaries of each functional zone within it. Attribute tables record the zone name, area, and responsible management personnel. Elevation and topography are stored in the form of an irregular triangular network for subsequent camera calibration parameter calculations during video spatial positioning and slope analysis in path planning. The road network is stored hierarchically as main park roads, secondary park roads, and walkways, with attributes including road surface material, traffic direction, height and weight limits.

[0028] This layer serves as the underlying reference for spatial positioning and path calculation, and the spatial coordinates of all upper-layer thematic data are aligned with it. Vegetation distribution is configured with species-specific symbols, and facility location symbols support dynamic status switching: online status displays a solid icon, offline status displays a gray hollow icon with a broken line marker, and alarm status icons are surrounded by a red pulsed halo. Clicking on a symbol will pop up an attribute card, displaying complete field information and a list of associated maintenance work orders.

[0029] Pipelines of different media types are stored independently in thematic layers, and vertical filtering queries are supported by the burial depth field. There is a topological relationship between pipelines and facility locations.

[0030] Furthermore, the dynamic layer overlay engine includes: The layer manager is used to maintain the stacking order and visibility control of the base map layer, real-time perception layer, resource scheduling layer, and analysis decision layer. Each layer supports independent transparency adjustment and blending mode configuration. The spatiotemporal alignment submodule receives structured perception event streams from the intelligent monitoring and analysis module, electricity consumption trend prediction results from the electricity consumption situation analysis module, and multi-source heterogeneous data from external IoT sensors. Based on a unified time reference and geographic coordinate system, it performs spatiotemporal alignment and coordinate transformation, mapping each source data to the same spatial reference frame of the vector base map engine. The rendering pipeline is integrated and a hierarchical rendering strategy is adopted: real-time perceived events are mapped to differentiated visual symbols according to risk level, electricity consumption trend prediction results are mapped to thermal gradient surfaces according to load intensity, and resource scheduling status is mapped to dynamic path lines according to availability; transparency blending and depth sorting are supported when multiple layers are overlaid. The interactive response interface responds to user-side zooming, panning, selection, and timeline dragging operations, dynamically triggering the spatial index engine's view cropping query and the spatiotemporal database's time-series slice retrieval, thereby enabling real-time refreshing and historical backtracking of the situation layer.

[0031] In this embodiment, the layer manager maintains the stacking order and visibility control of four layers, from bottom to top as follows: The base map layer is placed at the bottom layer and contains basic geographic data and garden-themed data provided by the vector base map engine. Currently, it displays garden boundaries, road networks, water systems, vegetation distribution, and facility locations.

[0032] The real-time perception layer is placed on the second layer, receiving the structured perception event stream output by the intelligent monitoring and analysis module. For example, if an event of exceeding the limit for crowd density is detected, or if a tourist illegally enters a restricted area in the lakeside area, a red flashing icon and an orange mid-frequency flashing icon are superimposed on the corresponding location, respectively.

[0033] The resource scheduling layer is placed on the third layer, displaying the real-time location of park security personnel and emergency supplies. These are labeled on the GIS interface using security personnel icons and equipment icons.

[0034] The analysis and decision layer is placed at the top level, receiving the prediction results output by the electricity consumption situation analysis module.

[0035] Preferably, the spatial fusion and overlay uses the spatiotemporal alignment submodule to spatially associate and match the location coordinates in the structured perception event stream with the facility locations and pipeline directions in the garden thematic data layer. When the event coordinates fall within a preset buffer zone around the facility, the attribute identifier of the facility is automatically associated, forming an event-facility binding relationship and writing it into the fusion attribute table. The electricity consumption trend prediction results are spatially aggregated according to the garden functional zones to generate a zone load density raster, which is then overlaid onto the corresponding zone boundary.

[0036] Furthermore, the hierarchical rendering is a three-level rendering rule, specifically including: Feature-level rendering: For a single structured perception event, a preset symbol template is selected based on the event type, and the symbol opacity is adjusted based on the confidence level. Regional rendering generates an event aggregation heatmap for multiple events aggregated within the same spatiotemporal window, which is then overlaid on the boundary of the corresponding garden functional area. Situation-level rendering integrates the cross-modal outputs of the intelligent monitoring and analysis module and the power consumption situation analysis module. When visual anomalies and power consumption anomalies are simultaneously present in the same area, composite warning rendering is triggered, and a fused information card pops up to display the details of multimodal cross-verification.

[0037] Specifically, the multi-task spatiotemporal perception inference architecture extracts single-frame visual features based on a spatiotemporal feature encoder and models inter-frame motion correlations, and deploys multi-task decoding heads in parallel after the spatiotemporal feature encoder.

[0038] Furthermore, the multi-task decoding head includes: The target detection decoding head performs target localization and classification on the spatiotemporal feature map output by the spatiotemporal feature encoder, and outputs the category label, bounding box coordinates and detection confidence of the target in the frame; The behavior recognition decoding head performs temporal convolution and pooling operations on the same spatiotemporal feature map along the temporal dimension to generate the probability distribution of action categories and output the behavior classification results and corresponding confidence scores at the frame or segment level.

[0039] In this embodiment, after the video stream is decoded by the intelligent monitoring and analysis module, it is input into the spatiotemporal feature encoder in a sliding window of 16 consecutive frames. The spatiotemporal feature encoder uses a CSPDarknet structure as the backbone for spatial feature extraction, performs 5 levels of downsampling on each frame, and outputs a multi-scale feature pyramid. Based on this, the encoder performs 3D convolution operations on the feature maps of the same spatial location in adjacent frames along the temporal dimension to model the motion correlation between frames, and finally outputs a spatiotemporal feature map. This spatiotemporal feature map is simultaneously input to the parallel-deployed target detection decoder and behavior recognition decoder.

[0040] After receiving the spatiotemporal feature map, the object detection decoding head extracts the spatial feature slice corresponding to the current keyframe, and then uses the decoupled detection head to predict three branches: The classification branch outputs a dimension of [20×36×3×(number of categories+1)], with each grid cell predicting the probability distribution of 3 anchor boxes across 5 target categories (pedestrians, vehicles, animals, floating objects, and others).

[0041] The regression branch outputs the bounding box offsets of each anchor frame (center point x, y offsets and width and height scaling factors), which are then decoded to obtain the bounding box coordinates of the target within the frame.

[0042] The confidence branch outputs the confidence level of the target in each anchor box.

[0043] The result is immediately transmitted to the target identification and annotation layer in the video monitoring function layer group, and a green rectangular annotation box and text label are superimposed on the corresponding video screen position on the GIS interface.

[0044] The behavior recognition decoder simultaneously receives the same spatiotemporal feature map. First, using the bounding box output by the object detection decoder as the region of interest, the feature tensor region corresponding to the pedestrian target is extracted along the spatial dimension from the spatiotemporal feature map and used as the behavior recognition input.

[0045] This decoder employs a SlowFast dual-channel approach to capture temporal features: the slow channel extracts spatial semantic features from 4 frames out of 16 frames at a low frame rate, performs spatial convolution using 3D convolution kernels, and captures static semantics such as pedestrian pose and position; the fast channel processes at the original frame rate and captures the pedestrian's motion change features. The two channels' features are laterally connected, aligned temporally, and then fused. Temporal pooling is then used to compress the temporal dimension, outputting an action category probability distribution.

[0046] Preferably, the video monitoring function layer group includes: The pedestrian density heat map layer receives the pedestrian statistics output by the target detection decoding head in the multi-task spatiotemporal perception inference architecture, performs kernel density interpolation and grid aggregation according to the garden road and square areas, generates a pedestrian density heat map surface, and superimposes it on the basic geographic data layer of the vector base map engine. The target recognition annotation layer is connected to the real-time video stream of the equipment in various areas of the garden through the geographic information spatial framework. Using the video screen as the base map, the intra-frame target category label, bounding box coordinates and detection confidence output by the target detection decoding head in the multi-task spatiotemporal perception inference architecture are superimposed and drawn in real time on the spatial position of the corresponding video screen to form the target detection annotation box and category text label. The non-safe behavior warning layer receives the frame-level or segment-level behavior classification results and corresponding confidence levels output by the behavior recognition decoding head, and annotates the behavior classification results in real time with differentiated visual symbols according to behavior categories; it generates a list of behavior analysis events in reverse chronological order in the sidebar of the interface, and each record in the list includes the behavior category name and the frame-level or segment-level time interval.

[0047] Specifically, the time-series prediction model includes: The historical electricity consumption curve library construction sub-module performs hierarchical aggregation and standardized storage of the real-time electricity consumption data stream according to the garden functional zoning and time period granularity, forming a multi-dimensional historical electricity consumption curve library. The baseline prediction sub-model, based on the historical electricity consumption curve library, extracts the periodic trend component and seasonal component of electricity load to generate a basic electricity consumption trend prediction curve that does not consider external factors. The external regression variable fusion unit collects and preprocesses external regression variables affecting electricity load, including meteorological data, holiday markers, garden activity schedules, and the zonal pedestrian density index in the pedestrian density heat map output by the intelligent monitoring and analysis module; it performs lag order analysis and correlation screening on the external regression variables, retaining the variable set that is significantly related to electricity load; The fusion prediction sub-model inputs the trend component output by the baseline prediction sub-model and the variable set filtered by the external regression variable fusion unit into the gradient boosting tree. The baseline prediction bias is corrected through the residual learning mechanism, and the fused electricity consumption trend prediction result is output. The fusion prediction sub-model supports online incremental training and automatically triggers model parameter updates when new observation data accumulates to a preset batch. The prediction result evaluation and correction unit calculates the root mean square error and mean absolute percentage error between the predicted electricity consumption trend and the actual observed value. When the error index exceeds the preset threshold, it triggers automatic adjustment of model hyperparameters or re-screening of external regression variables, and pushes the corrected prediction result to the dynamic layer overlay engine for fusion rendering.

[0048] In this embodiment, the historical electricity consumption curve library construction submodule aggregates the raw data collected from each power distribution zone on a daily basis, with a 15-minute time interval. The statistics, along with the zone identifier, date, date type label (weekday / weekend / holiday), time interval number, and time interval start timestamp, together constitute a structured feature vector.

[0049] The baseline prediction sub-model performs STL decomposition on the power mean time series of each region and time period in the historical electricity consumption curve library, decomposing the electricity load time series into three additive components: a trend component reflecting the long-term slow change direction of electricity load, a seasonal component capturing the intraday fluctuation pattern of a fixed period, and a residual component of random fluctuation. During baseline forecasting, the trend component and seasonal component corresponding to the forecast date are added together, and the residual component is not included in the forecast (because its mean is zero and it is not predictable), generating a basic electricity consumption trend forecast curve that does not consider external factors. The external regression variable fusion unit performs lag order analysis and correlation screening on variables such as meteorological data, holiday markers, park activity schedules, and zonal pedestrian density indices. It calculates the correlation coefficients between each external variable and electricity load. The fusion prediction sub-model constructs three types of features: The first category is the Prophet baseline component, which includes the trend term, annual seasonal term, and weekly seasonal term, enabling the residual model to understand the structure of the baseline prediction.

[0050] The second category is meteorological and environmental characteristics, including temperature deviation (actual temperature minus historical average for the same period), temperature deviation squared (capturing nonlinear effects), humidity lag value, and heavy rain indicator.

[0051] The third category is the characteristics of garden activities, including holiday signs, activity scale level, and peak lag value of pedestrian density.

[0052] For each prediction period, the baseline predictions and external variables are concatenated to form a feature vector, and a gradient boosting tree model is used to fuse the baseline predictions and external variables.

[0053] The trained model is stored as a file. Every morning, the system loads the model, inputs the baseline value and external variable value for the prediction day, and outputs the fused prediction value for each time period.

[0054] Specifically, the fusion situation rendering of the power consumption situation analysis module includes: The load density rendering of the zones involves spatially aggregating the electricity consumption trend prediction results according to the functional zones of the garden, calculating the predicted load density value of each zone, generating a zone load density raster surface using contour lines or hierarchical coloring, and superimposing it on the corresponding zone boundary of the vector base map engine. The time-series comparison rendering overlays a comparison layer of the historical actual load curve and the current predicted load curve within the same view, embedded in the side analysis panel in the form of a dual-axis time series graph.

[0055] Specifically, the garden operation and maintenance tasks include at least changes in vegetation maintenance status, updates to facility status, and corrections to pipeline attributes.

[0056] Specifically, the incremental update and status monitoring include: After the real-time update engine detects the completion of the operation and maintenance task, it drives the dynamic layer overlay engine to refresh the corresponding vector element layer in real time, so that the status of the garden elements displayed on the GIS interface is synchronized with the operation and maintenance results.

[0057] Example 1, such as Figure 2 As shown, the dynamic maintenance module's real-time update engine handles the entire lifecycle management of temporary tasks during garden operation and maintenance. Specific process: Task creation phase: Operations personnel select a target area (such as a vegetation maintenance area) on the GIS interface through the interactive response interface of the geographic information spatial framework, triggering the creation of a temporary task. The system automatically maps the task coordinates to the same spatial reference frame of the vector base map engine, generating structured operations data containing location coordinates, task type, and creation timestamp.

[0058] Status transition phase: The task status changes from "Processing pending review" → "Administrator review" → "Processed" or "Processing rejected". Each status change is captured by the real-time update engine, driving the dynamic layer overlay engine to refresh the corresponding vector feature layer in real time. For example, when the task status changes to "Processed", the maintenance level symbol for that vegetation area on the GIS interface switches from "Pending maintenance" to "Maintained" in real time.

[0059] Exception handling branch: If a task is invalidated, the version management mechanism records the change history, supporting conflict detection and merge rollback. The data of invalidated tasks is marked as deleted, but is retained in the time-series slices of the spatiotemporal database for historical backtracking queries.

[0060] Example 2, as Figure 3 The diagram illustrates how the spatiotemporal alignment submodule of the dynamic layer overlay engine processes multi-level operation and maintenance task plan data. The specific process is as follows: Task planning: Administrators create periodic maintenance task plans (such as monthly facility inspection plans). Task data includes the planned execution area (polygon coordinates), time period granularity, and responsible personnel. The spatiotemporal alignment submodule performs coordinate transformation based on a unified time base and geographic coordinate system.

[0061] Task execution monitoring: During task execution, operations and maintenance personnel report the progress via mobile devices. The real-time update engine receives structured operations and maintenance data streams and aggregates them in layers according to garden functional zones and time periods. The dynamic layer overlay engine displays the task execution trajectory on the GIS interface as dynamic path lines, mapped to differentiated colors based on completion rate.

[0062] Task feedback loop: After the task is completed, structured operation and maintenance data (including execution photo coordinates and maintenance status change records) are stored in the spatiotemporal database. The fusion rendering pipeline maps the task results into visual symbols according to risk level and overlays them onto the real-time perception layer.

[0063] Example 3, as Figure 4 The diagram illustrates how the spatial indexing engine supports the rapid location and facility association of work orders across the riverbanks area. The specific process is as follows: Work Order Reception: When a facility malfunction occurs in the area on both sides of a river (such as a street light breaking down), the system receives work order information. The spatiotemporal alignment submodule transforms the work order location coordinates to the same spatial reference frame of the vector base map engine.

[0064] Spatial association matching: The spatial indexing engine constructs an R-Tree spatial index for vector features and performs neighborhood retrieval. When the work order coordinates fall within the preset buffer zone around the street light facility, the attribute identifiers of the street light facility (facility number, installation date, maintenance record) are automatically associated, forming an "event-facility binding relationship" and written into the fused attribute table.

[0065] Work order processing and closure: After maintenance personnel complete the work, they confirm it via mobile device. A real-time update engine drives a dynamic layer overlay engine to instantly refresh the symbol status of the street light facility location (switching from a "warning" red pulse halo to an "online" icon). A version management mechanism records the facility status change history, supporting change tracking.

[0066] Example 4, as Figure 5 The diagram illustrates how the cross-modal outputs of the electricity consumption situation analysis module and the intelligent monitoring and analysis module assist in the assessment of external complaints. Specific process: Complaint access: External complainants submit complaints via mini-program / hotline (e.g., "Abnormal nighttime lighting in a certain area"). The system generates a structured event stream, including complaint type, location coordinates, and timestamp.

[0067] Multimodal cross-verification: The video analysis results of the area output by the dynamic layer overlay engine's integrated intelligent monitoring and analysis module (the thermal layer of nighttime pedestrian density shows that the pedestrian flow is normal during this period) and the power consumption trend prediction results output by the power consumption situation analysis module (the load of the lighting circuit in this area is abnormally low) trigger situation-level composite warning rendering.

[0068] Fusion Information Card: A fusion information card pops up on the GIS interface, displaying details of multimodal cross-verification - video monitoring shows that the lighting facility physically exists but has no light output, and power consumption data confirms that the current in the circuit is zero. Cross-verification indicates that it is a lamp fault rather than a power supply fault, accurately locating the maintenance plan.

[0069] Example 5, as Figure 6 The intelligent monitoring event-driven security patrol tasks are automatically dispatched. Specific process: Automatic monitoring trigger: The multi-task spatiotemporal perception inference architecture of the intelligent monitoring and analysis module detects tourists illegally entering a restricted area in the lakeside area. A structured perception event stream is generated.

[0070] Spatial positioning and layer rendering: Event coordinates are mapped to the vector base map engine via the spatiotemporal alignment submodule. The dynamic layer overlay engine overlays orange mid-frequency flashing icons at the corresponding positions. Simultaneously, if multiple out-of-bounds events aggregate within the same spatiotemporal window, an event aggregation heatmap is generated (region-level rendering).

[0071] Automatic patrol task assignment: After the real-time update engine detects a high-risk event, it automatically assigns a security patrol work order to the nearest security personnel's mobile device. The work order includes the event's precise coordinates, video screenshots, and records of similar historical events. After the security personnel complete the task handling, the task status changes, driving a GIS layer refresh, and the event icon switches from "Pending Handling" to "Handled".

[0072] Example 6, as Figure 7 The diagram illustrates how the spatiotemporal alignment submodule handles multi-source heterogeneous data from external work orders across departments. The specific process is as follows: Multi-source data access: Receives work order data from external units such as the Greening Center and the Urban Development Group, with varying data formats. The spatiotemporal alignment submodule extracts key fields: location description (text address), time information, and problem type.

[0073] Geocoding and Coordinate Transformation: Text addresses are geocoded into WGS84 latitude and longitude coordinates, then projected and transformed to the CGCS2000 coordinate system, aligned with the spatial reference frame of the vector base map engine. Time information is uniformly converted to UTC timestamps.

[0074] Construction of the fusion attribute table: Spatially match the transformed coordinates with the facility locations and pipeline routes in the landscape thematic data layer. For example, if the coordinates of the work order "damaged manhole cover at a certain intersection" fall within the buffer zone around the sewage pipeline, the pipeline attribute identifier will be automatically associated to form a complete fusion attribute record.

[0075] Visualized Collaborative Processing: The dynamic layer overlay engine marks external work orders with differentiated visual symbols on the GIS interface, and supports layer filtering and transparency adjustment by source unit.

[0076] Example 7, as Figure 8 The diagram illustrates how the vector base map engine's thematic data layer for gardens supports the park management department's facility and equipment maintenance operations. Specific workflow: Problem Reporting: Park management personnel discovered a malfunction in an amusement facility and reported it via mobile device. The problem's coordinates were obtained via GPS location, and the engine received structured operation and maintenance data in real time for updates.

[0077] Dynamic status switching for facilities: In the garden-themed data layer of the vector base map engine, the symbol for the amusement facility location is configured according to its functional category (amusement facility category), and its status dynamically switches from "online" (solid colored icon) to "alarm". This status change is rendered in real time by the dynamic layer overlay engine, and the GIS interface does not need to be manually refreshed.

[0078] Maintenance process version management: During the maintenance work order process, the facility status undergoes multiple changes from "alarm" to "under maintenance" to "repaired". The version management mechanism records the timestamp, operator, and content of each change. If problems are still found after maintenance, it supports rolling back to the previous version and triggering conflict detection (such as whether it matches the current power consumption data).

[0079] Closed-loop statistics: After maintenance is completed, structured operation and maintenance data (including before and after photos and a list of replaced parts) are stored in a spatiotemporal database and incorporated into the facility's full life cycle archive.

[0080] Example 8, as Figure 9 As shown, this is a viewpoint from the park management department, demonstrating the vector base map engine's detailed attribute encoding of vegetation distribution and real-time updates of maintenance status. Specific process: Vegetation attribute coding system: In the thematic data layer of gardens, vegetation distribution is coded according to species, crown width, tree age, and maintenance level. Each tree corresponds to a unique vector element ID.

[0081] Daily data reporting: After daily inspections, the park management department's greening patrol personnel report problems via mobile devices (e.g., "Camphor trees in a certain area have pests and diseases"). The reported data includes tree ID, problem type, and photo.

[0082] Maintenance Status Change: After the real-time update engine detects a "pest and disease" issue report, it drives the dynamic layer overlay engine to incrementally update the maintenance level attribute of the vegetation element. On the GIS interface, the tree symbol changes from green to red, and a pest and disease icon is added.

[0083] Tiered rendering display: The fusion rendering pipeline maps differentiated visual symbols according to maintenance level. Regional rendering aggregates trees with multiple diseases and pests in the same area, generating a heat map of disease and pest distribution.

[0084] Archived statistics: After the daily reports are confirmed by the park director, the records of changes in maintenance status are included in the statistics, and multi-dimensional analysis can be carried out by species, region and time period.

[0085] Example 9, as Figure 10 The diagram illustrates the perspective of the maintenance management department, demonstrating how the version control mechanism handles data consistency for the same facility maintenance tasks from different departmental viewpoints. Specific process: Multi-source operation access: The same facility malfunction may be reported by both the park management department and the maintenance management department at the same time. The park management department reports "amusement facility malfunction", while the maintenance management department reports "the facility has abnormal power supply".

[0086] Conflict Detection: The version management mechanism detected two parallel state change records for the same facility location. The system automatically triggers conflict detection. Park Management Record: Status = "Mechanical Failure", Time; Maintenance Management Department Record: Status = "Electrical Fault", Time; Cross-modal verification: The video analysis results from the integrated intelligent monitoring and analysis module of the dynamic layer overlay engine (showing no physical damage to the facility) and the real-time data from the power consumption status analysis module (showing zero current in the facility's circuit) indicate an electrical fault rather than a mechanical fault. The system prompts the park management department that there is a discrepancy in the records and suggests merging them into "electrical fault" and rolling back the "mechanical fault" record.

[0087] Merge and rollback: After manual confirmation, the version management mechanism performs the merge operation, retaining correct records and marking incorrect records as "rollbacked" but keeping them in the change history for auditing and traceability.

[0088] Example 10, as Figure 11 As shown, this demonstration, from the perspective of the maintenance management department, illustrates how the real-time update engine ensures real-time synchronization between maintenance team operations and the GIS interface status. Specific process: Problem Reporting: When the maintenance team leader / inspection personnel discover vegetation problems (such as "hedges in a certain area are not trimmed in time"), they can report them via mobile device. The coordinates of the problem are located by GPS and spatially matched with vegetation distribution elements in the landscape thematic data layer.

[0089] Team leader confirmation and task assignment: After the team leader confirms the issue on the mobile device, the system automatically generates a maintenance work order and assigns it to the responsible work team. At this time, the status of the symbol for the corresponding vegetation area on the GIS interface changes from "normal" to "pending maintenance".

[0090] The process is synchronized in real time: After maintenance personnel complete the pruning on-site, they upload photos of the process (including GPS coordinates and timestamps) via mobile device. Once the real-time update engine detects the "operation completion event," it immediately drives the dynamic layer overlay engine to refresh the corresponding vector feature layer. The vegetation maintenance level has been updated from "Needing Maintenance" to "Maintained"; The symbol color changed from flashing yellow to solid green; The maintenance date field is updated to the current date; Timeliness Guarantee: The system requires daily report submissions. Real-time engine updates ensure that all daily maintenance operations have been updated in the GIS layer, guaranteeing that the GIS interface viewed by managers before leaving get off work is completely consistent with the actual site conditions.

[0091] Multi-dimensional statistics: The head of the maintenance management department / supervisor can select and query through the interactive response interface to count indicators such as the monthly vegetation maintenance completion rate and average response time in a certain area. The results are displayed in the form of an analysis and decision layer.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A smart garden safety management and scheduling system based on multimodal perception and GIS, characterized in that, include: The geographic information spatial framework consists of a vector base map engine, a dynamic layer overlay engine, and a spatiotemporal database. The vector base map engine is used to manage vector elements such as garden functional zoning, road and water system, vegetation distribution, facility locations, and pipeline routes. The dynamic layer overlay engine is used to spatially fuse, overlay, and hierarchically render real-time sensing data, resource locations, and dynamic analysis results. The spatiotemporal database is used to store structured sensing event streams. The intelligent monitoring and analysis module, embedded in the geographic information spatial framework, decodes real-time video streams from devices in various areas of the garden and inputs them into a multi-task spatiotemporal perception inference architecture for multi-target detection and behavior recognition. This generates a structured perception event stream containing event type, location coordinates, timestamp, and confidence level, which is then stored in the spatiotemporal database. Simultaneously, it drives the dynamic layer overlay engine on the geographic information spatial framework to overlay and visualize the structured perception event stream according to event type and spatial location, expanding it into a video monitoring function layer group. The electricity consumption situation analysis module connects to the real-time electricity consumption data stream of the garden power distribution IoT terminal, constructs a historical electricity consumption curve library divided by garden functional zoning and time period granularity, and generates electricity consumption trend prediction results for a preset period by superimposing external regression variables through a time series prediction model. Based on the electricity consumption trend prediction results, the dynamic layer overlay engine is used to perform fusion rendering. The dynamic maintenance module has a built-in real-time update engine, which is used to incrementally update the structured data generated by the intelligent monitoring and analysis module and the power consumption status analysis module, as well as the structured operation and maintenance data generated by the garden operation and maintenance task operations initiated through the geographic information spatial framework.

2. The system according to claim 1, characterized in that, The vector base map engine includes: The basic geographic data layer stores the baseline vector data of garden boundaries, elevation topography, road network and water system network, serving as the underlying reference for spatial positioning and path calculation; The garden-specific data layer overlays garden-specific vector elements, including vegetation distribution, facility locations, and pipeline routes. The vegetation distribution is coded by attribute according to species, crown width, tree age, and maintenance level; the facility locations are symbolically configured according to functional categories; and the pipeline routes are managed in layers according to media type and burial depth. The spatial indexing engine constructs a spatial index for the vector features, supporting the dynamic layer overlay engine in spatial positioning queries and neighborhood retrieval of real-time perceived events. The version management mechanism is used to record the change history of the vector elements and supports the dynamic maintenance module in detecting conflicts and merging and rolling back incremental update data.

3. The system according to claim 2, characterized in that, The dynamic layer overlay engine includes: The layer manager is used to maintain the stacking order and visibility control of the base map layer, real-time perception layer, resource scheduling layer, and analysis decision layer. Each layer supports independent transparency adjustment and blending mode configuration. The spatiotemporal alignment submodule receives structured perception event streams from the intelligent monitoring and analysis module, electricity consumption trend prediction results from the electricity consumption situation analysis module, and multi-source heterogeneous data from external IoT sensors. Based on a unified time reference and geographic coordinate system, it performs spatiotemporal alignment and coordinate transformation, mapping each source data to the same spatial reference frame of the vector base map engine. The rendering pipeline is integrated and a hierarchical rendering strategy is adopted: real-time perceived events are mapped to differentiated visual symbols according to risk level, electricity consumption trend prediction results are mapped to thermal gradient surfaces according to load intensity, and resource scheduling status is mapped to dynamic path lines according to availability; transparency blending and depth sorting are supported when multiple layers are overlaid. The interactive response interface responds to user-side zooming, panning, selection, and timeline dragging operations, dynamically triggering the spatial index engine's view cropping query and the spatiotemporal database's time-series slice retrieval, thereby enabling real-time refreshing and historical backtracking of the situation layer.

4. The system according to claim 3, characterized in that, The spatial fusion and overlay uses the spatiotemporal alignment submodule to spatially associate and match the location coordinates in the structured perception event stream with the facility locations and pipeline directions in the garden thematic data layer. When the event coordinates fall within the preset buffer zone around the facility, the attribute identifier of the facility is automatically associated, forming an event-facility binding relationship and writing it into the fusion attribute table. The electricity consumption trend prediction results are spatially aggregated according to the garden functional zones to generate a zone load density grid, which is then superimposed onto the corresponding zone boundaries.

5. The system according to claim 3, characterized in that, The hierarchical rendering refers to a three-level rendering rule, specifically including: Feature-level rendering: For a single structured perception event, a preset symbol template is selected based on the event type, and the symbol opacity is adjusted based on the confidence level. Regional rendering generates an event aggregation heatmap for multiple events aggregated within the same spatiotemporal window, which is then overlaid on the boundary of the corresponding garden functional area. Situation-level rendering integrates the cross-modal outputs of the intelligent monitoring and analysis module and the power consumption situation analysis module. When visual anomalies and power consumption anomalies are simultaneously present in the same area, composite warning rendering is triggered, and a fused information card pops up to display the details of multimodal cross-verification.

6. The system according to claim 1, characterized in that, The multi-task spatiotemporal perception inference architecture extracts single-frame visual features based on a spatiotemporal feature encoder and models inter-frame motion correlations, and deploys multi-task decoding heads in parallel after the spatiotemporal feature encoder.

7. The system according to claim 6, characterized in that, The multi-task decoding head includes: The target detection decoding head performs target localization and classification on the spatiotemporal feature map output by the spatiotemporal feature encoder, and outputs the category label, bounding box coordinates and detection confidence of the target in the frame; The behavior recognition decoding head performs temporal convolution and pooling operations on the same spatiotemporal feature map along the temporal dimension to generate the probability distribution of action categories and output the behavior classification results and corresponding confidence scores at the frame or segment level.

8. The system according to claim 7, characterized in that, The video monitoring function layer group includes: The pedestrian density heat map layer receives the pedestrian statistics output by the target detection decoding head in the multi-task spatiotemporal perception inference architecture, performs kernel density interpolation and grid aggregation according to the garden road and square areas, generates a pedestrian density heat map surface, and superimposes it on the basic geographic data layer of the vector base map engine. The target recognition annotation layer is connected to the real-time video stream of the equipment in various areas of the garden through the geographic information spatial framework. Using the video screen as the base map, the intra-frame target category label, bounding box coordinates and detection confidence output by the target detection decoding head in the multi-task spatiotemporal perception inference architecture are superimposed and drawn in real time on the spatial position of the corresponding video screen to form the target detection annotation box and category text label. The non-safe behavior warning layer receives the frame-level or segment-level behavior classification results and corresponding confidence levels output by the behavior recognition decoding head, and annotates the behavior classification results in real time with differentiated visual symbols according to behavior categories; it generates a list of behavior analysis events in reverse chronological order in the sidebar of the interface, and each record in the list includes the behavior category name and the frame-level or segment-level time interval.

9. The system according to claim 1, characterized in that, The time-series prediction model includes: The historical electricity consumption curve library construction sub-module performs hierarchical aggregation and standardized storage of the real-time electricity consumption data stream according to the garden functional zoning and time period granularity, forming a multi-dimensional historical electricity consumption curve library. The baseline prediction sub-model, based on the historical electricity consumption curve library, extracts the periodic trend component and seasonal component of electricity load to generate a basic electricity consumption trend prediction curve that does not consider external factors. The external regression variable fusion unit collects and preprocesses external regression variables affecting electricity load, including meteorological data, holiday markers, garden activity schedules, and the zonal pedestrian density index in the pedestrian density heat map output by the intelligent monitoring and analysis module; it performs lag order analysis and correlation screening on the external regression variables, retaining the variable set that is significantly related to electricity load; The fusion prediction sub-model inputs the trend component output by the baseline prediction sub-model and the variable set filtered by the external regression variable fusion unit into the gradient boosting tree. The baseline prediction bias is corrected through the residual learning mechanism, and the fused electricity consumption trend prediction result is output. The fusion prediction sub-model supports online incremental training and automatically triggers model parameter updates when new observation data accumulates to a preset batch. The prediction result evaluation and correction unit calculates the root mean square error and mean absolute percentage error between the predicted electricity consumption trend and the actual observed value. When the error index exceeds the preset threshold, it triggers automatic adjustment of model hyperparameters or re-screening of external regression variables, and pushes the corrected prediction result to the dynamic layer overlay engine for fusion rendering.

10. The system according to claim 1, characterized in that, The fusion situation rendering of the electricity consumption situation analysis module specifically includes: The load density rendering of the zones involves spatially aggregating the electricity consumption trend prediction results according to the functional zones of the garden, calculating the predicted load density value of each zone, generating a zone load density raster surface using contour lines or hierarchical coloring, and superimposing it on the corresponding zone boundary of the vector base map engine. The time-series comparison rendering overlays a comparison layer of the historical actual load curve and the current predicted load curve within the same view, embedded in the side analysis panel in the form of a dual-axis time series graph.

11. The system according to claim 1, characterized in that, The garden operation and maintenance tasks include at least the following: changes to vegetation maintenance status, updates to facility status, and corrections to pipeline attributes.

12. The system according to claim 1, characterized in that, The incremental update and status monitoring include: After the real-time update engine detects the completion of the operation and maintenance task, it drives the dynamic layer overlay engine to refresh the corresponding vector element layer in real time, so that the status of the garden elements displayed on the GIS interface is synchronized with the operation and maintenance results.

Images