Community parking collaborative navigation method and system based on digital twinning and federated learning

Abstract

The application provides a community parking collaborative navigation method and system based on digital twinning and federated learning, belongs to the technical field of intelligent parking, and specifically comprises the following steps: constructing a digital twin of community parking resources based on real-time monitoring data of each contracted community; receiving model parameters of a local prediction model uploaded by each contracted community; generating a global prediction model corresponding to the community parking resources; pushing a global prediction result corresponding to a real-time position uploaded by a user terminal; generating a navigation path from the digital twin according to a path planning method based on a target parking space fed back by the user terminal; and generating a contract use time period through a sensor signal uploaded by the target parking space when a user vehicle corresponding to the user terminal arrives at the target parking space, so as to count parking fee information of the user terminal. Through the processing scheme, precise prediction of cross-regional resource supply and demand is realized under the premise of protecting the data privacy of each community, and the overall matching efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent parking technology, and in particular to a community parking collaborative navigation method and system based on digital twins and federated learning. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of private car ownership, the "last mile" problem for urban residents' travel is becoming increasingly prominent. Among these challenges, parking difficulties and disorderly parking in older residential areas have become a major pain point in community governance. Survey data shows that drivers spend an average of 4 to 8 minutes blindly wandering around looking for parking spaces after entering a residential area. This not only wastes fuel and increases carbon emissions but also poses serious safety hazards due to narrow roads and mixed pedestrian and vehicle traffic within the community. Furthermore, there is a significant structural mismatch in parking resources, with surface parking spaces in high demand while underground or private parking spaces are largely idle. While shared parking models have emerged, allowing residents to share idle parking spaces during certain periods, existing shared parking solutions mostly focus on supply and demand matching, failing to integrate shared parking space status and real-time navigation. Users still need to find shared parking spaces themselves after entering the community, and there is a lack of assessment of the surrounding safety environment of shared parking spaces. Therefore, existing solutions still have significant shortcomings in terms of refined guidance, safety avoidance, dynamic allocation, trust mechanisms, and privacy protection. Summary of the Invention

[0003] Therefore, in order to overcome the shortcomings of the prior art, the present invention provides a community parking collaborative navigation method and system based on digital twins and federated learning. This method can perceive the parking space status in real time and integrate the community road topology. It provides precise point-to-point guidance from the entrance to the target parking space through digital twins. Furthermore, by utilizing federated learning, it achieves accurate prediction of cross-regional resource supply and demand while protecting the data privacy of each community, thereby improving the overall matching efficiency.

[0004] To achieve the above objectives, this invention provides a community parking collaborative navigation method based on digital twins and federated learning, comprising: performing high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and constructing a digital twin database of community parking resources, wherein the digital twin database corresponds to multiple contracted communities; receiving model parameters of local prediction models uploaded by the local end of each contracted community, wherein the local prediction models are trained by the local end using a federated learning framework on the historical parking space occupancy data of each contracted community; and generating a global prediction model corresponding to the community parking resources based on the model parameters of the local prediction models and the digital twins. This allows for the generation of global predictions of demand and availability distribution for parking spaces in each contracted community within a real-time estimated time period. Based on account information and real-time location uploaded by the user terminal, the global prediction results corresponding to that real-time location are pushed to the user terminal for displaying contracted community parking spaces. The system receives target parking spaces from the user terminal, calls a path planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal. When the user's vehicle arrives at the target parking space, the system generates a contract usage period based on sensor signals uploaded by the target parking space, in order to collect parking fee information for that user terminal.

[0005] In some embodiments, the method for collecting real-time monitoring data includes: acquiring video images of the contracted community in real time through the video acquisition unit of the contracted community; acquiring parking space information of the contracted community parking spaces in real time through the status sensors of the parking spaces in the contracted community; and summarizing the video images and the parking space information based on timestamps to obtain the real-time monitoring data.

[0006] In some embodiments, the local prediction model is trained on the local end using a federated learning framework on historical parking space occupancy data of each contracted community. This includes: acquiring historical parking space occupancy data for each contracted community, which at least includes vehicle entry / exit records, owner parking space sharing records, holiday data, and parking space usage time periods; training a local shared parking space prediction model on the local end of each contracted community using a federated learning framework based on the historical parking space occupancy data; uploading the model parameters of the local shared parking space prediction model to a cloud server and receiving and adjusting the parameters from the cloud server; and optimizing the local shared parking space prediction model based on the adjusted parameters to obtain the local prediction model.

[0007] In some embodiments, the step of using a federated learning framework to train a local shared parking space prediction model based on the historical parking space occupancy data includes: receiving an initialized shared parking space prediction model architecture from a cloud server, wherein the shared parking space prediction model architecture uses a temporal convolutional network based on an attention mechanism as the basic backbone network; converting the vehicle entry and exit records, owner parking space sharing records, holiday data, and parking space usage time periods in the historical parking space occupancy data into a local training dataset with multi-dimensional features; inputting the local training dataset into the shared parking space prediction model architecture for local learning training, and continuously iterating the model parameters of the local shared parking space prediction model.

[0008] In some embodiments, the step of generating a global prediction model corresponding to the community parking resources based on the model parameters of the local prediction model and the digital twin, so as to generate a global prediction result of the demand heat and idle time distribution of parking spaces in each contracted community within a real-time estimated time period, includes: inputting the model parameters of the received multiple local prediction models as prior knowledge into a preset basic neural network framework, which has the same input and output dimensions as each local prediction model, forming a model copy corresponding to each local prediction model; fusing the model parameters of each digital twin to obtain fusion parameters, and using the fusion parameters as the global prediction result. The initialization parameters of the global prediction model are set, and the dynamic attribute layer of the digital twin is encoded as an additional static covariate input and embedded into the input layer of the global prediction model to generate a global prediction model corresponding to the community parking resources; real-time monitoring data from each of the contracted communities is received in real time at a preset time frequency; the real-time monitoring data is input into the digital twin of the corresponding community to correct the estimated vehicle arrival rate at the current moment; the real-time monitoring data sequence of the most recent time window is input into the global prediction model to generate a global prediction result of the demand heat and idle time distribution of parking spaces in each contracted community within the expected time period from the beginning of the time window.

[0009] In some embodiments, receiving the target parking space fed back by the user terminal, invoking a route planning method to generate a navigation path from the digital twin, and pushing the navigation path to the user terminal includes: receiving and determining the target contracted community where the target parking space is located; acquiring a global video image of the target contracted community in real time; performing personnel detection analysis on the global video image, and marking the specific area as an avoidance area when the personnel density in a specific area exceeds a predetermined threshold; acquiring a road planning map from the digital twin, confirming the road congestion situation within the contracted community based on the global video image, and generating a navigation path based on the avoidance area.

[0010] In some embodiments, the method further includes: receiving evaluation information from the user terminal after the vehicle corresponding to the user terminal leaves the target parking space, and storing the sharing behavior of the owner corresponding to the target parking space; and statistically analyzing the sharing behavior of the owner and the compliance behavior of the car owner corresponding to the user terminal to generate a credit score list to provide point rewards.

[0011] A community parking collaborative navigation system based on digital twins and federated learning includes: a cloud server for performing high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and constructing a digital twin database of community parking resources, which corresponds to multiple contracted communities; generating a global prediction model corresponding to the community parking resources based on the model parameters of the local prediction model uploaded by the local terminal and the digital twin, so as to generate global prediction results of the demand heat and idle time distribution of parking spaces in each contracted community within a real-time estimated time period; receiving the target parking space feedback from the user terminal, calling the path planning method to generate a navigation path from the digital twin, and pushing the navigation path to the user terminal; a local terminal, a server set in each contracted community, for learning and training a local prediction model using the historical parking space occupancy data of each contracted community using a federated learning framework; and a user terminal for uploading account information and real-time location, and receiving the global prediction results corresponding to the real-time location pushed by the cloud server, so as to select the target parking space from the displayed parking spaces in the contracted community.

[0012] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described method.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0014] Compared with existing technologies, the advantages of this invention are as follows: By using digital twins, navigation accuracy is extended from the community entrance to specific parking spaces, solving the "last 100 meters" blind spot problem; moreover, by utilizing federated learning, accurate prediction of cross-regional resource supply and demand is achieved while protecting the data privacy of each community, improving overall matching efficiency. Furthermore, the entire solution makes full use of existing monitoring facilities, adopts lightweight edge computing, reduces system deployment and transformation costs, and creates additional revenue for property owners by revitalizing idle parking space resources, providing car owners with convenient and low-cost parking options, achieving optimized resource allocation and a win-win situation for all parties. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the community parking collaborative navigation method based on digital twins and federated learning in an embodiment of the present invention. Figure 2 This is a global topological knowledge graph of a community parking collaborative navigation method based on digital twins and federated learning in an embodiment of the present invention; Figure 3 This is a structural block diagram of a community parking collaborative navigation system based on digital twins and federated learning, as described in an embodiment of the present invention. Figure 4 This is an internal structural diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0017] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0018] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] It should be noted that the following description covers various aspects of embodiments within the scope of protection of this invention. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0020] It should also be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0021] Furthermore, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0022] This application provides a community parking collaborative navigation method based on digital twins and federated learning, applied on a cloud server or server. The server can be a standalone server or a server cluster composed of multiple servers. Figure 1 As shown, taking the application of this community parking cooperative navigation method based on digital twins and federated learning on a cloud server as an example, the method includes the following steps: Step 101: Perform high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and construct a digital twin database of community parking resources, which corresponds to multiple contracted communities.

[0023] The cloud server performs high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and constructs a digital twin database of community parking resources, which corresponds to multiple contracted communities.

[0024] The cloud server can be an IoT aggregation gateway or a cloud-based digital twin engine deployed at the edge of the community. The execution entity receives monitoring data uploaded by various IoT sensing devices deployed in each contracted community in real time through heterogeneous network interfaces (including but not limited to LoRa, ZigBee, Wi-Fi 6, and wired Ethernet) at a preset sampling frequency (e.g., once every 100 milliseconds).

[0025] Real-time monitoring data can include the following three types of source data: 1. Spatial geometric data: Video analysis images collected by community cameras deployed at key nodes in the community (such as entrances and exits, crossroads, and elevator positions in multi-level parking garages) are used to capture changes in the physical structure of the parking garage. 2. Status semantic data: vehicle occupancy status (occupied / free) reported by sensors installed above each parking space and vehicle entry and exit event streams recorded in real time by the gate system.

[0026] 3. Environmental context data: including current light intensity and weather conditions within the community (provided by a small weather station), as well as carbon monoxide / temperature and humidity sensor data from the underground parking garage, used to correct noise in visual mapping during subsequent modeling.

[0027] After receiving the above data, the cloud server first performs multi-source data alignment based on a unified timestamp (synchronized to the millisecond level using the NTP network time protocol) to form structured real-time data packets in frames, and caches them in the memory database for later use.

[0028] The cloud server performs high-precision 3D modeling of each contracted community, generating a digital twin that includes roads, parking spaces, buildings, and high-risk areas. Real-time monitoring data can include real-time parking space usage status data from parking space sensors and video analysis data collected by community cameras, facilitating dynamic updates of parking space status, vehicle trajectories, and pedestrian density within the digital twin. The digital twin can include static attribute layers, dynamic attribute layers, and rule layers. The static attribute layer contains fixed information within the community, such as the community's geographical coordinates, total number of parking spaces, physical dimensions of parking spaces, entrance / exit gate locations, and channel topology. The dynamic attribute layer can include information on parking space changes within the community, such as the vehicle arrival rate distribution function for different time periods, statistical characteristics of average parking duration, and historical parking space turnover rate curves for the same period (e.g., holidays / weekdays). The rule layer can include billing-related information such as parking fee strategies for each community, monthly / temporary parking ratios, and locking rules for special periods (e.g., overnight parking).

[0029] The cloud server utilizes a pre-trained deep learning semantic segmentation model (such as Mask R-CNN) to segment video images captured by community cameras, accurately identifying the bounding box, center point coordinates, and current occupancy status label of each individual parking space. The execution unit then binds each segmented parking space instance to its corresponding spatial coordinates in a dynamic geometry layer—specifically, it calculates the homography matrix between image pixel coordinates and 3D world coordinates using a perspective-n-Perspective (PnP) transformation algorithm, precisely projecting and pasting the parking space's occupancy status and vehicle color as attribute labels onto the corresponding surface of the 3D model, forming a semantic texture layer.

[0030] The cloud server can also perform real-time lighting compensation rendering on the surface of the 3D model based on changes in ambient light intensity. For example, when the outdoor light dims or the underground parking garage lights are turned on, the cloud server automatically adjusts the diffuse reflection coefficient and shadow generation parameters of the model surface, so that the final generated 3D model visually approximates the real-time scene of a parking lot.

[0031] A cloud server can generate a high-precision 3D digital image for a single contracted community, characterized by real-time synchronization, geometric accuracy, and rich semantics. This image can evolve synchronously with changes in the physical parking space status. During continuous operation, the cloud server opens an independent data processing channel for each constructed digital twin. For each channel, the cloud server dynamically adjusts the synchronization strategy based on the drastic changes in the community's data.

[0032] In one embodiment, the real-time monitoring data may further include an auxiliary sensing unit for assisting in sensing the real-time location information of the vehicle waiting to be parked, thereby further increasing the credibility of the digital twin. The auxiliary sensing unit may be a Bluetooth locator installed on the vehicle waiting to be parked.

[0033] Step 102: Receive the model parameters of the local prediction model uploaded by the local end of each contracted community. The local prediction model is trained by the local end using a federated learning framework on the historical parking space occupancy data of each contracted community.

[0034] The cloud server sends model parameter upload requests to the local terminals of each contracted community. In response to the request, each contracted community uploads the model parameters of its pre-trained local prediction model (including but not limited to: the weight matrix of the network structure, bias vector, learning rate decay coefficient, and the mean and variance of the normalization layer, etc.) to the cloud server after encryption (such as using homomorphic encryption, differential privacy technology, or gradient encryption).

[0035] The cloud server receives model parameters from the local prediction models uploaded by the local clients of each contracted community. These local prediction models are trained on the local clients using a federated learning framework based on the historical parking space occupancy data of each contracted community.

[0036] For example, for a contracted community A, its local prediction model is built on a Long Short-Term Memory (LSTM) network to predict the occupancy status of each parking space (e.g., spaces A-001 to A-200) within the community over the next two hours based on historical parking data (entry and exit records every 15 minutes over the past 7 days). The model parameters received by the cloud server may include the weight parameters W of each gate of the LSTM. f W i W o W c and bias term b f , b i , b o , b c .

[0037] Step 103: Generate a global prediction model corresponding to community parking resources based on the model parameters of the local prediction model and the digital twin, so as to generate global prediction results of the demand heat and idle time distribution of parking spaces in each contracted community within the real-time expected time period.

[0038] The cloud server retrieves the digital twins corresponding to each contracted community from its built-in or cloud-based digital twin database. The cloud server then inputs the model parameters from multiple local prediction models as prior knowledge into a pre-defined basic neural network framework. This basic neural network framework has the same input dimension (time step) and output dimension (parking space demand probability) as each local prediction model.

[0039] Cloud servers can employ an attention-based weighted aggregation algorithm to fuse the weight parameters of various local prediction models. The fusion weights can be determined by the similarity of the dynamic attribute layers in the digital twins of each community: for example, if the historical parking behavior patterns of several contracted communities are similar, their model parameters will be assigned greater weights during fusion; if the historical parking behavior patterns of a certain contracted community have a low repetition rate, its model parameters may be assigned smaller weights during fusion.

[0040] Let θ be the LSTM weights for communities A, B, and C. A θ B θ C The corresponding digital twin similarity coefficient is α. A α B α C (α) A +α B +α C If =1), then the fused global initial model parameter θ global = α A ·θ A + α B ·θ B + α C ·θ C .

[0041] Then, the cloud server will merge the θ global As initialization parameters for the global prediction model, the dynamic attribute layer of the digital twin (such as the arrival rate distribution function of each community) is encoded as an additional static covariate input and embedded into the input layer of the global prediction model, thereby generating a global prediction model corresponding to community parking resources. The architecture of this global prediction model is configured to support multi-task learning and can simultaneously output parking space status prediction results for multiple contracted communities.

[0042] The global forecast results can be displayed using a multi-dimensional forecast matrix, which can include forecast values ​​for demand intensity and idle time distribution.

[0043] The demand heat forecast value can be the probability value (a continuous value between 0 and 1) of each contracted parking space in each contracted community being requested to be occupied within the expected time period, as well as the overall demand heat level (high, medium, low) of the area where the parking space is located (such as the ground of area A and the underground of area B).

[0044] The predicted vacancy time distribution can be a prediction of the time when a currently occupied parking space is expected to become vacant (e.g., parking space A-001 is expected to become vacant within 14:35±5 minutes). Based on the predicted vacancy time of all parking spaces, a distribution curve of the number of vacant parking spaces in the community over a future expected period can be generated.

[0045] Step 104: Based on the account information and real-time location uploaded by the user terminal, push the global prediction result corresponding to the real-time location so that the user terminal can display the signed community parking space.

[0046] The cloud server obtains a user profile based on the account information uploaded by the user terminal. Then, based on the user profile and real-time location, it determines the user's potential demand for shared parking spaces. Based on the potential demand for shared parking spaces, it obtains the local global prediction results corresponding to the real-time location for the expected time period (e.g., the next 30 minutes, 1 hour, and 2 hours) from the global prediction results, so that the user terminal can display the available contracted community parking spaces.

[0047] Step 105: Receive the target parking space from the user terminal, call the route planning method to generate a navigation path from the digital twin, and push the navigation path to the user terminal.

[0048] The cloud server receives the target parking space from the user terminal, calls a route planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal. Route planning methods can include common grid-based methods, A* algorithms, dynamic programming, etc.

[0049] The cloud server obtains road topology from the map in the digital twin, real-time parking status from the parking management service in the digital twin, and population density in high-risk areas from the edge nodes in the digital twin, calculating a navigation route that balances efficiency and safety. The navigation route is pushed to the user's terminal in real time via a mini-program or 5G message, and the cloud server also supports replanning if the target parking space is occupied during navigation. If the target parking space is occupied during navigation, the cloud server automatically pushes a replanning instruction based on available parking spaces in the contracted community where the selected target parking space is located.

[0050] Step 106: When the user vehicle corresponding to the user terminal arrives at the target parking space, the contract usage period is generated through the sensor signal uploaded by the target parking space in order to collect the parking space fee information of the user terminal.

[0051] When a user's vehicle arrives at the target parking space, the local terminal uploads the sensor signal from the target parking space sensor. The cloud server receives this sensor signal and records the start time of use. After the vehicle leaves, the local terminal uploads the sensor signal from the target parking space sensor again, and the cloud server receives this sensor signal and records the end time of use. The cloud server generates a contract usage period based on the start and end times of use to calculate the parking space fee information for that user terminal.

[0052] The cloud server stores the billing rules corresponding to the contracted community and generates parking space fee information for the user's terminal based on the contract usage period. The cloud server can then deduct the corresponding parking space fee from the user's account.

[0053] The aforementioned method extends navigation accuracy from the community entrance to specific parking spaces through digital twins, solving the "last 100 meters" blind spot problem. Furthermore, by utilizing federated learning, it achieves accurate prediction of cross-regional resource supply and demand while protecting the data privacy of each community, improving overall matching efficiency. Moreover, the entire solution makes full use of existing monitoring facilities and employs lightweight edge computing, reducing system deployment and upgrade costs. By revitalizing idle parking space resources, it generates additional revenue for property owners and provides drivers with convenient and low-cost parking options, achieving optimized resource allocation and a win-win situation for all parties.

[0054] In one embodiment, the method for collecting real-time monitoring data includes: acquiring video images of the contracted community in real time through a video acquisition unit of the contracted community; acquiring parking space information of the contracted community parking spaces in real time through a status sensor of the parking spaces in the contracted community; and summarizing the video images and parking space information based on timestamps to obtain real-time monitoring data.

[0055] The local end acquires real-time video images of the contracted communities through video capture units. These video capture units can include high-definition cameras deployed above the gates at community vehicle entrances and exits, cameras deployed in internal community passageways, and fisheye cameras located above parking spaces. High-definition cameras capture full frontal images and close-ups of license plate areas of vehicles entering / exiting in real time. Cameras deployed in internal community passageways capture wide-angle video images of various areas of the parking lot in real time, facilitating the determination of crowd density within the camera's coverage area. Fisheye cameras capture real-time images of the occupancy status of multiple parking spaces and vehicle exteriors within their coverage area. These video capture units are existing monitoring equipment in each contracted community, allowing for lightweight edge computing on the local end, reducing the overall system deployment and upgrade costs.

[0056] At each physical parking space within the contracted community, a pre-installed status sensor is synchronously connected to a local terminal. These status sensors can be geomagnetic sensors for open-air parking spaces or ultrasonic sensors for underground parking garages. Each status sensor actively reports parking space information at a preset sampling period (e.g., once every 500 milliseconds). The parking space information includes at least: a unique identifier for the parking space (e.g., "A-05"), an occupancy status code (e.g., "0" indicates vacant, "1" indicates occupied), a status confidence score (a floating-point number between 0 and 1, provided by the sensor's internal algorithm), and a timestamp indicating the first occurrence of that status.

[0057] The local end extracts the hardware-level timestamp for each frame of video image and each parking space information (synchronized with a standard clock source via Network Time Protocol (NTP), with millisecond precision). Based on this, the local end performs a rigorous timestamp aggregation and alignment process: 1. Using the system master clock as a reference, the continuous time axis is divided into aligned time windows of fixed length. The length of the aligned time window can be preset to 100 milliseconds, 500 milliseconds, or 1 second, depending on the monitoring accuracy requirements. In this embodiment, 100 milliseconds is preferred to achieve higher real-time resolution.

[0058] 2. For each alignment time window, the local end acquires the earliest timestamped video frame within that time window as the representative image frame for that window; and acquires parking space information data reported by all status sensors within that time window. If multiple records exist for the same parking space within a time window, the record with the timestamp closest to the representative image frame's timestamp is taken as the alignment data for that parking space. If no sensor reports data for a parking space within that time window, a linear interpolation method is used to estimate and fill in the gaps based on the status values ​​of the previous and next windows.

[0059] 3. Using the aligned time window as a unit, the hardware timestamp of the representative image frame is used as the unified time label for the time window. All parking space information gathered within the time window is associated and bound with the unified time label to form a structured data tuple of "one video image + multiple parking space status records" to generate the final real-time monitoring data.

[0060] The above method combines video acquisition with multiple types of state sensors (geomagnetic and ultrasonic) to form a dual-redundant acquisition system of "visual semantics + physical field perception". False alarms from a single sensor under adverse weather or obstruction conditions can be verified and corrected by the other type of data, improving the overall accuracy of parking space status detection. Moreover, the alignment window division and interpolation mechanism of synchronization time effectively solves the timing mismatch problem caused by the inconsistency between the video frame rate (15~25fps) and the sensor reporting frequency (2Hz), keeping the time deviation of the fused real-time monitoring data within ±50 milliseconds, providing a reliable data foundation for subsequent high-precision dynamic modeling.

[0061] In one embodiment, the local prediction model is trained on the local end using a federated learning framework on historical parking space occupancy data of each contracted community. This includes: acquiring historical parking space occupancy data for the contracted community, which at least includes vehicle entry and exit records, owner parking space sharing records, holiday data, and parking space usage time periods; training a local shared parking space prediction model on the local end of each contracted community using a federated learning framework based on the historical parking space occupancy data; uploading the model parameters of the local shared parking space prediction model to a cloud server, and receiving and adjusting the parameters based on feedback from the cloud server; and optimizing the local shared parking space prediction model based on the adjusted parameters to obtain the local prediction model.

[0062] The cloud server generates an initial global shared parking space prediction model architecture, which uses a common attention-based neural network as its backbone. The server encrypts the initial network parameters of the global model (including convolutional kernel weights, bias terms, and attention score matrix) and distributes them to the local servers of each participating community. Upon receiving these initial parameters, each local server creates an identical copy of the model and awaits local training.

[0063] The local terminal obtains historical parking space occupancy data of the contracted community. This historical parking space occupancy data includes at least the entry and exit records of vehicles in the contracted community, the records of owners sharing parking spaces, holiday data, and the time period of parking space usage.

[0064] Vehicle entry and exit records can include a timestamp (accurate to the second), a unique vehicle identifier (anonymized license plate number), entry / exit direction (entry / exit), and gate channel number. This record is used to deduce the total capacity and turnover rate of vehicles within the community for each time period.

[0065] Owner-owned parking space sharing records can include the time periods during which owners set their own parking spaces to "shareable," the start and end times of actual temporary parking, and the settlement of sharing revenue. This record is a core source of information for predicting future available shared parking space capacity.

[0066] Holiday data can include national statutory holiday dates, work schedule adjustments, and dates of specific folk activities in the community. This data is used to mark periodic anomalies in the time series.

[0067] The parking space usage time period can include the precise time sequence of each physical parking space (including shared parking spaces and fixed parking spaces) changing from "idle" to "occupied" and from "occupied" to "idle" in each historical day.

[0068] Each participating community's local endpoint utilizes its unique historical parking space occupancy data to train a local copy of the model. After completing local training, each local endpoint updates the local model parameters W. local new Differential privacy noise is applied (the noise level is dynamically adjusted according to the number of local data entries to ensure that the privacy budget ε≤2), and then the noise-added parameters are uploaded to the cloud server.

[0069] After receiving local model parameters from all participating communities, the cloud server executes the federated aggregation algorithm. The cloud server then uses the sample size (denoted as n) of the historical parking space occupancy data uploaded by each community. i ) and data distribution variance (denoted as σ) i Calculate the dynamic weighting coefficient β i The aggregation formula is: W global new = Σ (β i · W local newi ), where β i =n i / Σ n i And when σ i When the threshold is exceeded, the contribution weight of that community is automatically reduced to prevent abnormal data from polluting the global model.

[0070] The cloud server will aggregate the new global model parameters W global new The broadcast is sent to the local terminals of all contracted communities. Each local terminal updates its local model parameters to W. global new Based on this, the system prepares to enter the next round of federated training until the preset termination conditions are met (e.g., the global loss function value decreases by less than 0.001% within three consecutive rounds, or the preset maximum number of communication rounds of 100 is reached). At this point, the models held by the local terminals of each contracted community are trained into mature local shared parking space prediction models. This model has fully learned the unique parking behavior patterns of its own community, and at the same time, it has borrowed the general temporal characteristics of other communities through the federated framework. Even if the data distribution of a community drifts due to sudden activities, the model can quickly adapt and correct itself in subsequent rounds of federated iterations.

[0071] The above method strictly adheres to data privacy regulations by storing historical parking space occupancy data locally and transmitting only model parameters (after differential privacy protection). This prevents the leakage of sensitive information such as owner parking space sharing records. Furthermore, by incorporating multi-source features, including holiday data and owner parking space sharing records, and aggregating them through federated averaging and dynamic weights, the trained local shared parking space prediction model can accurately capture both periodic patterns and sudden demand.

[0072] In one embodiment, a federated learning framework is used to train a local shared parking space prediction model based on historical parking space occupancy data. This includes: receiving an initialized shared parking space prediction model architecture from a cloud server, wherein the shared parking space prediction model architecture uses a temporal convolutional network based on an attention mechanism as the basic backbone network; converting vehicle entry and exit records, owner parking space sharing records, holiday data, and parking space usage time periods from the historical parking space occupancy data into a local training dataset with multi-dimensional features; inputting the local training dataset into the shared parking space prediction model architecture for local learning training, and continuously iterating the model parameters of the local shared parking space prediction model.

[0073] The local end receives the initial shared parking space prediction model architecture sent by the cloud server. This architecture uses a temporal convolutional network (TCN) based on an attention mechanism as the basic backbone network.

[0074] Before training, local data cleaning and feature extraction are performed: vehicle entry and exit records are timestamped with parking space usage time periods to generate time-series sample slices with a granularity of 15 minutes (the time granularity can be set as needed); holiday data is one-hot encoded; and the sharing duration in the owner parking space sharing records is converted into numerical features, thus forming a local training dataset D with multi-dimensional features (the number of feature dimensions is denoted as F) and a time step of T (e.g., 672 time steps in the past 7 days). local .

[0075] The local machine inputs processed time-series slices (e.g., 72 consecutive hours of historical data) into a temporal convolutional network. The network extracts long-range dependencies of the time series through dilated convolutional layers and adaptively assigns higher weights to different historical moments (especially data points before and after holidays) through an attention mechanism. The network output layer provides a predicted occupancy probability for each parking space in each time slice for the next preset time period (e.g., the next 24 hours).

[0076] The local unit employs a composite loss function combining Focal Loss and mean squared error to compare the predicted value with the true historical label (i.e., the actual occupancy status in the next time period) and calculate the local loss value L. localThe Focal Loss component is used to alleviate the imbalance between the number of "vacant parking spaces" and "occupied parking spaces" samples within the community.

[0077] Local end based on loss value L local An adaptive moment estimator (Adam) is used for backpropagation to calculate the gradients of the parameters in each layer of the network, and the model parameters are iteratively updated using local data to form the updated local model parameters W. local new To reduce communication overhead, in this embodiment, the local device only sends a parameter reporting request to the cloud server after completing five full epochs of local training.

[0078] After completing their local training, each local endpoint applies differential privacy noise (the noise level is dynamically adjusted based on the number of local data points to ensure a privacy budget ε≤2) to the updated local model parameters W_local_new, and then uploads the noise-added parameters to the cloud server. Upon receiving the local model parameters from all participating communities, the cloud server executes the federated aggregation algorithm.

[0079] The above method introduces multi-source features, including holiday data and owner parking space sharing records, and aggregates them through federated averaging and dynamic weights, enabling the trained local shared parking space prediction model to accurately capture periodic patterns and sudden demands.

[0080] like Figure 2 As shown, in one embodiment, a global prediction model corresponding to community parking resources is generated based on the model parameters of the local prediction model and the digital twin, in order to generate global prediction results of the demand and idle time distribution of parking spaces in each contracted community within a real-time estimated time period, including: Step 201: The model parameters of the received multiple local prediction models are input as prior knowledge into a preset basic neural network framework. The basic neural network framework has the same input and output dimensions as each local prediction model, forming a model copy corresponding to each local prediction model.

[0081] The cloud server sends a model parameter upload request to each participating community. Each participating community responds to this request by packaging and encrypting the model parameters of its pre-trained local prediction model before uploading it. Model parameters may include structural parameters, weight parameters, and statistical parameters.

[0082] The structural parameters include architectural description information such as the number of network layers, the number of neurons per layer, the type of activation function, and the number of attention heads in the local prediction model.

[0083] The weight parameters include the weight matrix and bias vector of each network layer, including convolutional kernel weights, gate weights of recurrent neural networks, and connection weights of fully connected layers.

[0084] Statistical parameters include normalization parameters such as the mean, variance, maximum and minimum values ​​of the local training dataset, as well as the loss value and accuracy metrics of the model on the local validation set.

[0085] After receiving the above model parameters, the server first verifies the data source of the model parameters to ensure data integrity and source credibility, and then stores the model parameters of each community in the parameter cache library according to the community identifier.

[0086] The cloud server receives model parameters from multiple local prediction models and inputs them as prior knowledge into a preset basic neural network framework. This basic neural network framework has the same input and output dimensions as each local prediction model, forming a model copy corresponding to each local prediction model.

[0087] Step 202: The model parameters of each digital twin are fused to obtain fusion parameters. The fusion parameters are used as the initialization parameters of the global prediction model. The dynamic attribute layer of the digital twin is encoded as an additional static covariate input and embedded into the input layer of the global prediction model, thereby generating a global prediction model corresponding to community parking resources.

[0088] The cloud server merges the model parameters of each digital twin to obtain fusion parameters, uses the fusion parameters as the initialization parameters of the global prediction model, and encodes the dynamic attribute layer of the digital twin as an additional static covariate input, embedding it into the input layer of the global prediction model, thereby generating a global prediction model corresponding to community parking resources.

[0089] For example, the cloud server denotes the LSTM weights of communities A, B, and C as θ. A θ B θ C The corresponding digital twin similarity coefficient is α. A α B α C (α) A +α B +α C If =1), then the fused global initial model parameter θ global = α A ·θ A + α B ·θ B +α C ·θ C .

[0090] The cloud server uses the merged θglobal as the initialization parameter of the global prediction model, and encodes the dynamic attribute layer of the digital twin (such as the arrival rate distribution function of each community) as an additional static covariate input, embedding it into the input layer of the global prediction model, thereby generating a global prediction model corresponding to community parking resources.

[0091] The cloud server extracts the digital twin features of each digital twin, such as the dynamic attribute layer of the digital twin, and encodes them as additional static covariate inputs. These features are then mapped to fixed-length embedding vectors via an independent feature embedding network (composed of multiple fully connected layers). These embedding vectors are then embedded into the input layer of the global prediction model, thereby generating a global prediction model corresponding to community parking resources.

[0092] Step 203: Receive real-time monitoring data from each contracted community at a preset time frequency.

[0093] The cloud server receives real-time monitoring data from each contracted community at a preset time frequency. For example, the cloud server receives real-time monitoring data from the gate systems, geomagnetic sensors, or parking space cameras of each contracted community every 5 minutes. The real-time monitoring data includes: the current occupancy / vacancy status of each parking space, the license plate recognition information and entry time of entering vehicles, and the exit time of exiting vehicles. This real-time monitoring data is anonymized.

[0094] Step 204: Input the real-time monitoring data into the digital twin of the corresponding community to correct the estimated vehicle arrival rate at the current moment.

[0095] The cloud server synchronizes and updates the real-time monitoring data to the dynamic attribute layer of the corresponding community's digital twin to correct the current vehicle arrival rate estimate. The cloud server can also concatenate and fuse the anonymized real-time monitoring data with the embedding vector from step 202 at the input layer to form an enhanced input layer with increased dimensions.

[0096] Specifically, the original input is the time-series data of parking space occupancy over the past T time steps (with a dimension of T×C, where C is the number of feature channels), and the dimension of the twin feature embedding vector is E. Then, the dimension of the enhanced input layer becomes T×(C+E). This enhanced input layer enables the global prediction model to simultaneously integrate the time-series information of the current real-time monitoring and the static / quasi-static prior knowledge provided by the digital twin when receiving real-time data.

[0097] Step 205: Input the real-time monitoring data sequence of the most recent time window into the global prediction model to generate a global prediction result of the demand for parking spaces in each contracted community and the distribution of idle time within the expected time period starting from the time window.

[0098] The cloud server normalizes the monitoring data sequence from the most recent time window (e.g., the past hour) to form an input tensor, which is then fed into the global prediction model. The global prediction model performs forward propagation calculations based on the current input tensor, initialized model parameters, and updated digital twin rule layers to generate real-time global prediction results for expected future time periods (e.g., the next 30 minutes, 1 hour, and 2 hours). The cloud server normalizes the real-time monitoring data using the same preprocessing method as during local model training, extracts the historical sequence from the most recent T time steps, and combines this with the current time label to extract the corresponding twin feature vector from the digital twin. These vectors are then concatenated to construct the prediction input tensor for the current moment.

[0099] The cloud server inputs the prediction tensor into the global prediction model and performs forward propagation calculations. For each contracted parking space in each contracted community, the global prediction model outputs the prediction results for the expected time period. The prediction results can include several display units: header metadata, demand popularity data area, and idle time distribution data area.

[0100] The header metadata can include the prediction generation time, the start and end times of the predicted time period, and the list of communities covered.

[0101] The demand heat data area can be the demand heat level of each parking space, the heat value matrix of each grid area, and the global demand trend curve.

[0102] The idle time distribution data area can be the predicted release time of each occupied parking space, the sequence of changes in the number of idle parking spaces in the community as a whole, and peak / valley markers.

[0103] The cloud server can push the global prediction results in real time to the management terminals of each contracted community and the user terminals of each user in the form of a JSON data stream. Simultaneously, the cloud server can display the global prediction results on the graphical user interface of the management terminal as a heatmap overlaid with a timeline line chart. Specifically, it is displayed in a visual chart format on the comprehensive management screen of the management terminal. Demand intensity is rendered as a heatmap overlaid on a digital twin 3D model, and idle time distribution is displayed as a dynamic line chart. This allows managers to intuitively grasp the supply and demand situation of parking space resources across communities in future periods, thereby enabling them to implement scheduling strategies in advance (such as guiding vehicles to low-demand communities and issuing temporary parking discounts), facilitating cross-community parking space scheduling for managers and allowing users to pre-select parking spaces before their trips.

[0104] The above method incorporates cross-community common knowledge into the global prediction model during the initialization phase by weighted aggregation of local model parameters from multiple communities based on digital twin similarity, thus improving prediction accuracy during the cold start phase. By using the static / quasi-static features of the digital twin as embedding vectors and concatenating them with real-time time-series inputs, the model simultaneously possesses prior structural knowledge and real-time dynamic response capabilities. In the event of sudden traffic events (such as temporary road closures leading to a surge of vehicles), the model can quickly adjust its prediction output within two time steps, reducing response latency. By simultaneously outputting prediction results in two dimensions—demand intensity and idle time distribution—it provides upper-layer applications with joint decision-making information on "space (which parking space is in high demand) — time (when it will be released)," thereby improving the overall efficiency of cross-community parking resource scheduling.

[0105] In one embodiment, receiving a target parking space from a user terminal, invoking a route planning method to generate a navigation path from a digital twin, and pushing the navigation path to the user terminal includes: receiving and determining the target contracted community where the target parking space is located; acquiring a global video image of the target contracted community in real time; performing personnel detection analysis on the global video image, and marking the specific area as an avoidance area when the personnel density in a specific area exceeds a predetermined threshold; acquiring a road planning map from the digital twin, confirming the road congestion situation within the contracted community based on the global video image, and generating a navigation path based on the avoidance area.

[0106] The cloud server receives a request from a user terminal containing a target parking space. The user terminal includes, but is not limited to, in-vehicle navigation screens, smartphone apps, or in-vehicle infotainment systems. The request must include at least a unique identifier for the target parking space the user wishes to park in. Upon receiving the request, the cloud server parses the target parking space identifier and extracts the community code segment from its prefix field. Based on a pre-stored community code mapping table (which records the correspondence between each community code and its corresponding physical community entity, digital twin identifier (Twin-ID), and communication address), the cloud server determines the target contracted community where the target parking space is located and obtains the digital twin identifier (Twin-ID) and corresponding data access interface address of that target contracted community.

[0107] For example, if the target parking space identifier is "Twin-Community-B-Carport-087", the navigation server parses the community code segment "Twin-Community-B", then determines that the target contracted community is "Second Contracted Community B", and obtains its corresponding digital twin access address.

[0108] The cloud server acquires real-time global video images of the target contracted community. Based on the acquired digital twin access address, the cloud server sends a real-time video stream acquisition request to the monitoring and management system of the target contracted community. In response to this request, multiple video acquisition units deployed within the target contracted community (including but not limited to high-definition capture cameras at entrances and exits, panoramic dome cameras inside the venue, fisheye cameras in passageways, and PTZ tracking cameras deployed in key areas) begin continuously uploading their real-time video streams at a preset frame rate (e.g., 25 frames per second).

[0109] The cloud server performs hardware decoding on each video stream using its video decoding module, and then stitches and merges the decoded video frames according to their actual spatial locations within the community. Specifically, the cloud server extracts the camera intrinsic and extrinsic parameters (including focal length, principal point coordinates, rotation matrix, and translation vector) from each video frame, and uses image stitching algorithms (such as an adaptive stitching method based on grid optimization) to synthesize the multiple video frames into a global video image covering the main roads, passageways, and parking areas of the target contracted community. The global video image can be a panoramic 2D or 3D stitched image, capable of completely displaying the distribution of people and vehicles within the target contracted community.

[0110] The cloud server caches global video images in memory and marks them with their corresponding acquisition timestamps (accurate to the millisecond level) for subsequent synchronization and analysis.

[0111] The cloud server invokes a pre-deployed local or cloud-based neural network model for people detection (e.g., an object detection model based on YOLOv8 or Faster R-CNN architecture) to perform frame-by-frame people detection analysis on the global video images. Specifically, the cloud server inputs the global video images into the people detection model, and the model outputs the detection bounding box and its center pixel coordinates for each person instance in the image. The cloud server maps each detected person instance to a pre-established community spatial coordinate system, forming a set of person location points P = {p1, p2, p3, ..., p...} in physical coordinates. m}, where m is the total number of people detected in the current frame.

[0112] The cloud server extracts a two-dimensional road plan of the target contracted community from its digital twin and divides the entire community's ground area into multiple grid units according to a preset grid granularity (e.g., a square grid of 5 meters x 5 meters). For each grid unit, the cloud server counts the number of people falling within that unit and calculates the ratio of the number of people in that unit to the grid area, which is used as the instantaneous population density value (unit: people / square meter) for that grid unit.

[0113] To avoid misjudgments caused by random fluctuations in single-frame detection, the cloud server performs cumulative smoothing processing on the personnel density value of each grid cell over a time window. Specifically, the cloud server caches the density calculation results of the most recent N consecutive frames (e.g., N=15 frames, corresponding to a 0.6-second time window), and calculates a sliding weighted average ρ of the density values ​​of the same grid cell within that time window. temporal .

[0114] Cloud servers will ρ temporal Compared with the preset warning threshold ρ for crowd gathering density threshold (e.g., ρ) threshold = 0.8 people / square meter) for comparison. When the ρ of a certain grid cell temporal When the predetermined threshold is exceeded, the cloud server determines that the spatial area where the grid cell is located is an area with abnormal population density and marks it as a high-risk gathering area.

[0115] The cloud server performs connectivity analysis on all grid cells identified as anomalously clustered, merging adjacent anomalous grid cells into a single connected region. For each merged connected region, the navigation server calculates its geometric center coordinates and coverage boundary (composed of the outermost grid's outer boundary), marking this connected region as an avoidance area on a 2D planar graph. The cloud server also assigns a unique avoidance area identifier (Avoid-ID) and a generation timestamp to this avoidance area.

[0116] The cloud server classifies and marks the attributes of the avoidance area: if ρ in the area temporal If the congestion exceeds 1.5 times the threshold, it is marked as a Level 1 avoidance zone (severe congestion); if it is between 1.0 and 1.5 times the threshold, it is marked as a Level 2 avoidance zone (moderate congestion), so that subsequent navigation strategies can take differentiated avoidance measures.

[0117] The cloud server extracts the road planning map data layer from the digital twin of the target signed community. The road planning map includes at least the centerline vector coordinates of all accessible roads in the community, road width, number of lanes, road attributes (main road / local road / fire lane), and the topological connection relationship between each road segment (represented by a directed graph G=(V,E), where V is a road node and E is a road segment).

[0118] The cloud server uses a pre-trained vehicle detection model (such as a YOLO- or SSD-based vehicle detector) to detect vehicles in the global video images. Based on the position of the vehicle detection bounding box in the image and its corresponding physical coordinate mapping, the cloud server calculates the number of vehicles on each road segment at the current time, and calculates the instantaneous vehicle density (unit: vehicles / meter) of each road segment in combination with the road segment length.

[0119] The cloud server calculates the average speed of vehicles on each road segment by analyzing the positional changes of the same vehicle in multiple consecutive images. Specifically, the cloud server uses optical flow or target tracking algorithms (such as SORT or DeepSORT) to track the displacement of vehicles in consecutive frames, calculates the instantaneous speed by combining the inter-frame time intervals, and then averages the speeds of all vehicles on the same road segment to obtain the average traffic speed v for that road segment. avg .

[0120] The cloud server comprehensively determines the congestion status of each road segment based on the instantaneous vehicle density and average traffic speed. The determination logic is as follows: when the instantaneous vehicle density exceeds a preset density threshold (e.g., 0.1 vehicles / meter) and the average traffic speed is lower than a preset speed threshold (e.g., 5 km / h), the road segment is determined to be "severely congested"; when the instantaneous vehicle density is between 0.5 and 1.0 times the preset density threshold and the average traffic speed is between 5 and 15 km / h, it is determined to be "moderately congested"; other situations are determined to be "unimpeded". The cloud server marks the road segments determined to be "severely congested" and "moderately congested" on the road planning map and records their congestion level and spatial location.

[0121] After obtaining the generated avoidance areas and confirmed road congestion information, the cloud server generates a navigation route. Based on the road planning map extracted from the digital twin, the cloud server constructs a rasterized cost map. The base cost value of each raster cell in the cost map is determined by its own attributes (low cost value for road areas, infinite cost value for green belts / building areas, indicating impassable). Based on this, the navigation server performs the following dynamic cost overlay updates: mapping the generated avoidance areas onto the cost map, and applying an additional passage penalty cost ΔC to the raster cells within the avoidance areas. avoid The penalty value for Level 1 avoidance zones is higher than that for Level 2 avoidance zones (for example, the penalty value for Level 1 avoidance zones is set to 10 times the base road cost value, while that for Level 2 avoidance zones is set to 5 times). This causes the navigation path planning algorithm to tend to actively avoid densely populated areas when finding a path. Confirmed congested road segments are mapped onto a cost map, and a time penalty value ΔC is applied to the grid cells of the congested road segments. congestion Furthermore, the penalty value increases with the congestion level (the penalty value for severe congestion is twice that for moderate congestion), causing the algorithm to tend to avoid congested road sections.

[0122] The cloud server performs path search on the updated cost map, starting from the user's current geographical location and ending at the target parking space. In this embodiment, the navigation server uses an improved A* search algorithm, combined with the previously calculated road traffic speed, to estimate the time cost. The search cost function is set as: f(n) = g(n) + h(n) + λ·P(n), where g(n) is the actual cumulative cost from the starting point to the current node n (including basic traffic cost and dynamic penalty cost), h(n) is the heuristically estimated cost from the current node n to the target node (based on Euclidean distance), P(n) is the dynamic penalty value of the area where the current node n is located (determined by the superposition of avoidance area and congestion status), and λ is the adjustment coefficient of the dynamic penalty (λ ranges from 1.0 to 3.0 and can be adjusted in real time according to user preferences).

[0123] The cloud server generates a navigation path from the starting point to the target parking space using the aforementioned search algorithm. The navigation path is an ordered sequence of road nodes connected by road segments. This path has the following characteristics: spatially, it avoids all primary avoidance zones and, as far as possible, secondary avoidance zones; temporally, it avoids severely congested road sections and prioritizes alternative roads that are either uncongested or moderately congested but have higher traffic speeds.

[0124] In one embodiment, the cloud server can further generate at least two alternative navigation routes (e.g., the second route prioritizes avoiding congestion but does not deliberately bypass avoidance areas, and the third route is the shortest path but may pass through avoidance areas and congested sections). The cloud server performs a comprehensive scoring and ranking of each route based on indicators such as the estimated travel time, the level and distance of avoidance areas, and the length of congested sections, and outputs the ranked list of routes.

[0125] The cloud server encapsulates the generated navigation path list (including path geometric coordinate sequence, turn prompt instructions, estimated driving distance and estimated arrival time) in JSON format and pushes it to the user terminal that initiated the request via the mobile communication network.

[0126] After receiving the navigation route, the user terminal displays it on the map overlaid on the terminal's graphical user interface and guides the user to drive along the recommended route using voice and visual means.

[0127] The cloud server continuously analyzes real-time video images of the target contracted community during the user's journey. When the cloud server detects a significant change in the avoidance area or congestion status (e.g., the population density in the original avoidance area drops below a threshold, or a previously unobstructed road suddenly becomes congested), the cloud server automatically triggers a route replanning process, re-executes the aforementioned planning steps, pushes the updated navigation route to the user's terminal in real time, and simultaneously issues a "route change" voice prompt, ensuring that the navigation route is always dynamically optimized based on the latest population distribution and traffic conditions within the community.

[0128] The aforementioned method, by fusing the density analysis of people detection in global video images with the spatial grid of digital twins, achieves accurate identification and quantitative classification of densely populated areas within the community, fundamentally solving the safety hazard problem that traditional static map navigation cannot perceive densely populated areas. By simultaneously mapping avoidance areas and road congestion status to dynamic penalty values ​​in the cost map, the navigation path achieves synergistic optimization between safely avoiding densely populated areas and efficiently avoiding traffic congestion, improving overall traffic efficiency compared to single strategies that only consider the shortest distance or only consider traffic congestion. Through a dynamic triggering path replanning mechanism, the navigation path can adaptively adjust to changes in the real-time distribution of people and vehicles within the community, making it particularly suitable for scenarios where people gather rapidly during morning and evening rush hours, community activities, or emergencies, significantly improving the safety and user experience of end-point navigation.

[0129] In one embodiment, the community parking collaborative navigation method based on digital twins and federated learning further includes: receiving evaluation information from the user terminal after the vehicle corresponding to the user terminal leaves the target parking space, and storing the sharing behavior of the owner corresponding to the target parking space; statistically analyzing the sharing behavior of the owner and the compliance behavior of the vehicle owner corresponding to the user terminal to generate a credit score list to provide point rewards.

[0130] The cloud server also provides user management and data analysis services, specifically, such as managing user profiles, parking history, and credit scores, and providing property management with reports on parking space utilization and high-risk area event statistics.

[0131] The cloud server also features a blockchain smart contract layer deployed on a consortium blockchain or public blockchain, enabling fully automated and tamper-proof processes for the release, reservation, use, settlement, and default handling of shared parking spaces. After the owner sets the sharing time period and price, a unique digital asset is generated. The owner's reservation triggers a contract to lock the parking space and deduct the deposit. Vehicle entry and exit are automatically triggered by gate data, and the cloud server calculates and executes fee settlement and revenue distribution. All transaction records are recorded on the blockchain, ensuring transparency and trustworthiness.

[0132] When a vehicle arrives at the community entrance, the gate automatically opens after recognizing the license plate. Once the vehicle enters a shared parking space, sensors detect changes in status, and the smart contract on the cloud server records the start time of use. When the vehicle leaves, the cloud server's contract records the end time, automatically calculates the fee and deducts it from the deposit, transferring the proceeds to the resident's digital wallet. If overstaying occurs, the smart contract automatically deducts a penalty and records the breach, affecting the user's credit score.

[0133] Once the vehicle associated with the user's terminal leaves the target parking space, the cloud server receives the user's feedback and stores the corresponding sharing behavior of the owner of that parking space. The cloud server then analyzes the owners' sharing behavior and the vehicle owner's adherence to parking rules to generate a credit score ranking and provide point rewards. After parking, users provide feedback, and this data is fed back to the cloud server for algorithm optimization. The cloud server records the owners' sharing behavior and the vehicle owner's adherence to parking rules, generating credit scores and a community ranking, providing point rewards, thus forming a positive cycle of sharing, benefiting, and further sharing.

[0134] The above method, through blockchain smart contracts, constructs an automated, transparent, and tamper-proof shared parking space trading mechanism, solving the trust issue between property owners and car owners; it also promotes neighborly assistance and community harmony through shared incentives and a credit system, improving resident satisfaction; by revitalizing idle parking space resources, it creates additional income for property owners and provides car owners with convenient and low-cost parking options, achieving optimal resource allocation and a win-win situation for all parties.

[0135] like Figure 3 As shown, a community parking collaborative navigation system based on digital twins and federated learning is presented. The system includes a cloud server, a local terminal, and a user terminal.

[0136] The cloud server is used to perform high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and to build a digital twin database of community parking resources. This digital twin database corresponds to multiple contracted communities. Based on the model parameters of the local prediction model uploaded by the local terminal and the digital twin, a global prediction model corresponding to the community parking resources is generated to provide global prediction results of the demand for parking spaces and the distribution of idle time in each contracted community within the real-time expected time period. The server receives the target parking space feedback from the user terminal, calls the path planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal.

[0137] On the local end, servers are set up in each contracted community to learn and train a local prediction model using the historical parking space occupancy data of each contracted community through a federated learning framework.

[0138] The user terminal is used to upload account information and real-time location, and to receive global prediction results corresponding to the real-time location pushed by the cloud server, so as to select a target parking space from the displayed contracted community parking spaces.

[0139] In one embodiment, the community parking collaborative navigation system based on digital twins and federated learning also includes a management terminal installed at the property management office. The management terminal receives real-time global prediction results for its contracted community and displays them on the management page, providing the property management with data-driven governance tools such as parking space utilization analysis and high-risk area event statistics to assist in scientific decision-making.

[0140] Specific limitations regarding the community parking collaborative navigation system based on digital twins and federated learning can be found in the limitations of the community parking collaborative navigation method based on digital twins and federated learning mentioned above, and will not be repeated here. Each module in the aforementioned community parking collaborative navigation system device based on digital twins and federated learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in the computer device in hardware form, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0141] In one embodiment, a computer device is provided, which may be a cloud server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data such as digital twins. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a community parking cooperative navigation method based on digital twins and federated learning.

[0142] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0143] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: performing high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and constructing a digital twin database of community parking resources, the digital twin database corresponding to multiple contracted communities; receiving model parameters of a local prediction model uploaded by the local end of each contracted community, the local prediction model being trained on the historical parking space occupancy data of each contracted community using a federated learning framework on the local end; and generating a digital twin database based on the model parameters of the local prediction model and the community parking resources. A global prediction model for district parking resources is used to generate real-time predictions of demand and availability distribution for parking spaces in each contracted community within a given time period. Based on account information and real-time location uploaded by the user terminal, the global prediction results corresponding to that location are pushed to the user terminal for displaying contracted community parking spaces. The system receives target parking spaces from the user terminal, calls a path planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal. When the user's vehicle arrives at the target parking space, the system generates a contract usage period based on sensor signals uploaded by the target parking space to collect parking fee information for that user terminal.

[0144] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program performs the following steps: high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and constructs a digital twin database of community parking resources, the digital twin database corresponding to multiple contracted communities; receives model parameters of local prediction models uploaded by local terminals of each contracted community, the local prediction models being trained on historical parking space occupancy data of each contracted community using a federated learning framework; generates a global prediction model corresponding to community parking resources based on the model parameters of the local prediction models and the digital twins, so as to generate global prediction results of the demand heat and idle time distribution of parking spaces in each contracted community within a real-time estimated time period; pushes the global prediction results corresponding to the real-time location based on the account information and real-time location uploaded by the user terminal, so that the user terminal can display the parking spaces of the contracted community; receives the target parking space feedback from the user terminal, calls a path planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal; when the user vehicle corresponding to the user terminal arrives at the target parking space, generates a contract usage time period based on the sensor signals uploaded by the target parking space, so as to collect parking space fee information for the user terminal.

[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A community parking cooperative navigation method based on digital twins and federated learning, characterized in that, include: High-precision 3D modeling and mapping are performed based on real-time monitoring data of each contracted community, and a digital twin database of community parking resources is constructed, which corresponds to multiple contracted communities. Receive model parameters of the local prediction model uploaded by the local end of each of the contracted communities. The local prediction model is learned and trained by the local end using a federated learning framework on the historical parking space occupancy data of each contracted community. Based on the model parameters of the local prediction model and the digital twin, a global prediction model corresponding to the community parking resources is generated so as to generate global prediction results of the demand heat and idle time distribution of parking spaces in each contracted community within the real-time expected time period. Based on the account information and real-time location uploaded by the user terminal, the global prediction result corresponding to the real-time location is pushed so that the user terminal can display the signed community parking space. The system receives the target parking space from the user terminal, calls the path planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal. When the user vehicle corresponding to the user terminal arrives at the target parking space, the contract usage period is generated through the sensor signals uploaded by the target parking space, so as to collect the parking space fee information of the user terminal.

2. The community parking cooperative navigation method based on digital twins and federated learning according to claim 1, characterized in that, The method for collecting real-time monitoring data includes: The video capture unit of the contracted community acquires video images of the contracted community in real time. The parking information of the contracted community parking spaces is obtained in real time through the status sensors of the contracted community parking spaces; The real-time monitoring data is obtained by summarizing the video images and parking space information based on timestamps.

3. The community parking cooperative navigation method based on digital twins and federated learning according to claim 1, characterized in that, The local prediction model is trained on the local end using a federated learning framework on the historical parking space occupancy data of each contracted community, including: Obtain historical parking space occupancy data for the contracted community. This historical parking space occupancy data includes at least the entry and exit records of vehicles in the contracted community, owner parking space sharing records, holiday data, and parking space usage time periods. A federated learning framework is used to train a local shared parking space prediction model on the local end of each contracted community based on the historical parking space occupancy data. Upload the model parameters of the local shared parking space prediction model to the cloud server, and receive and adjust the parameters after feedback from the cloud server; The local shared parking space prediction model is optimized based on the adjustment parameters to obtain a local prediction model.

4. The community parking cooperative navigation method based on digital twins and federated learning according to claim 3, characterized in that, The step of using a federated learning framework to train a local shared parking space prediction model based on the historical parking space occupancy data includes: The system receives the initial shared parking space prediction model architecture from the cloud server. This shared parking space prediction model architecture uses a temporal convolutional network based on an attention mechanism as the basic backbone network. The historical parking space occupancy data, including vehicle entry and exit records, owner parking space sharing records, holiday data, and parking space usage time periods, is converted into a local training dataset with multi-dimensional features. The local training dataset is input into the shared parking space prediction model architecture for local learning and training, and the model parameters of the local shared parking space prediction model are continuously iterated.

5. The community parking cooperative navigation method based on digital twins and federated learning according to claim 1, characterized in that, The step of generating a global prediction model corresponding to the community parking resources based on the model parameters of the local prediction model and the digital twin, in order to generate global prediction results of the demand and idle time distribution of parking spaces in each contracted community within the real-time estimated time period, includes: The received model parameters of multiple local prediction models are input as prior knowledge into a preset basic neural network framework. This basic neural network framework has the same input and output dimensions as each local prediction model, forming a model copy corresponding to each local prediction model. The model parameters of each digital twin are fused to obtain fusion parameters, which are used as initialization parameters for the global prediction model. The dynamic attribute layer of the digital twin is encoded as an additional static covariate input and embedded into the input layer of the global prediction model, thereby generating a global prediction model corresponding to the community parking resources. Real-time monitoring data from each of the aforementioned contracted communities is received at a preset time frequency. The real-time monitoring data is input into the digital twin of the corresponding community to correct the estimated vehicle arrival rate at the current moment; The real-time monitoring data sequence of the most recent time window is input into the global prediction model to generate a global prediction result in real time of the demand for parking spaces and the distribution of idle time in each contracted community within the expected time period starting from that time window.

6. The community parking cooperative navigation method based on digital twins and federated learning according to claim 1, characterized in that, The process of receiving the target parking space feedback from the user terminal, invoking a route planning method to generate a navigation path from the digital twin, and pushing the navigation path to the user terminal includes: Receive and determine the target contracted community where the target parking space is located; Real-time acquisition of global video images of the target signed community; The global video image is analyzed for personnel detection. When the personnel density in a specific area exceeds a predetermined threshold, the specific area is marked as an avoidance area. The road planning map is obtained from the digital twin, and the road congestion situation within the contracted community is confirmed based on the global video image. A navigation path is generated based on the avoidance area.

7. The community parking cooperative navigation method based on digital twins and federated learning according to claim 1, characterized in that, Also includes: When the vehicle corresponding to the user terminal leaves the target parking space, the system receives the evaluation information fed back by the user terminal and stores the sharing behavior of the owner corresponding to the target parking space. The system tracks the sharing behavior of the property owners and the compliance behavior of the car owners corresponding to the user terminals to generate a credit score list and provide point rewards.

8. A community parking cooperative navigation system based on digital twins and federated learning, characterized in that, The system includes: A cloud server is used to perform high-precision 3D modeling and mapping based on real-time monitoring data of each contracted community, and to build a digital twin database of community parking resources. This digital twin database corresponds to multiple contracted communities. Based on the model parameters of the local prediction model uploaded by the local terminal and the digital twin, a global prediction model corresponding to the community parking resources is generated to provide a global prediction result of the demand for parking spaces and the distribution of idle time in each contracted community within a real-time estimated time period. The server receives the target parking space feedback from the user terminal, calls the path planning method to generate a navigation path from the digital twin, and pushes the navigation path to the user terminal. On the local end, servers are set up in each contracted community to learn and train a local prediction model from the historical parking space occupancy data of each contracted community using a federated learning framework. The user terminal is used to upload account information and real-time location, and to receive the global prediction result corresponding to the real-time location pushed by the cloud server, so as to select a target parking space from the displayed contracted community parking spaces.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. 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 steps of the method according to any one of claims 1 to 7.

Images