Parking space intelligent guiding system based on edge calculation
By leveraging edge computing technology, combined with multi-source data acquisition and intelligent decision-making modules, the problem of low efficiency in traditional parking lot management has been solved, enabling efficient parking space guidance and resource optimization, thereby improving parking lot operation efficiency and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AEROSPACE SCI & IND INTELLIGENT OPERATION RES & INFORMATION SECURITY RES INST (WUHAN) CO LTD
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-05
AI Technical Summary
Traditional parking lot management is inefficient, making it difficult to find parking spaces quickly, leading to congestion and waste of resources. It also lacks intelligent guidance methods and cannot make efficient use of limited parking spaces.
The intelligent parking guidance system, based on edge computing, includes multi-source real-time data acquisition, a distributed edge computing architecture, and an intelligent decision-making core computing module. It achieves high-precision parking space status perception, dynamic path planning, and resource allocation optimization. Combined with IoT sensors, computer vision, and cloud-based collaborative computing, it provides refined management.
This has brought about a qualitative leap in parking lot operation efficiency, enabling drivers to quickly find parking spaces, optimizing parking lot management, improving user experience and economic benefits, and providing a model for smart city construction.
Smart Images
Figure CN121982921A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent parking guidance technology, specifically relating to an intelligent parking guidance system based on edge computing. Background Technology
[0002] With the acceleration of urbanization and the surge in car ownership, parking difficulties have become a common challenge faced by major cities worldwide. Traditional parking lot management relies on manual guidance or static signage systems, which suffer from inefficiency, resource waste, and poor user experience. According to a 2022 report by the International Parking Institute (IPI), drivers spend an average of 4-7 minutes searching for parking spaces in large parking lots, leading to increased congestion, carbon emissions, and indirectly affecting surrounding traffic flow.
[0003] Meanwhile, the management is quite challenging, with numerous intersections and junctions in the internal roads, making it difficult for drivers to quickly find parking spaces. Parking management methods are not scientific enough, lacking intelligent guidance mechanisms, resulting in a significant imbalance between parking supply and demand. Parking spaces are scattered, with above-ground spaces located in office and residential areas. Furthermore, the complex layout of the buildings within the complex contributes to difficulties in finding parking, hindering the efficient use of limited parking spaces and impeding work efficiency. Summary of the Invention
[0004] (a) Technical problems to be solved The technical problem to be solved by this invention is: how to provide a parking space intelligent guidance system based on edge computing.
[0005] (II) Technical Solution To address the aforementioned technical problems, this invention provides a parking space intelligent guidance system based on edge computing. The parking space intelligent guidance system includes: a multi-source real-time data acquisition system, a distributed edge computing architecture module, and an intelligent decision-making core computing module. The multi-source real-time data acquisition system is the foundational layer of the intelligent parking guidance system. It achieves high-precision, low-latency parking space status perception and global environment modeling through the collaborative work of three major modules: IoT sensor deployment, computer vision enhancement, and auxiliary data source fusion. The distributed edge computing architecture module achieves efficient processing and real-time response to massive parking lot data through the organic combination of layered processing architecture and cloud collaborative computing. This architecture ensures low latency while making full use of cloud computing power for global optimization. The intelligent decision-making core computing module is the central hub of the intelligent parking guidance system. It consists of three major modules: spatiotemporal prediction model, dynamic path planning, and resource allocation optimization. It achieves efficient decision-making through the collaboration of multiple algorithms.
[0006] (III) Beneficial Effects Compared with existing technologies, this invention provides an intelligent parking guidance system based on edge computing, becoming a key component of smart city construction. Its core objective is to achieve optimal allocation of parking resources and path guidance through real-time perception, dynamic decision-making, and multimodal interaction. This invention provides a more convenient, faster, and safer vehicle management method, achieving intelligent parking guidance, improving vehicle traffic efficiency, solving the problem of finding parking spaces, and providing refined and intelligent management tools for vehicles.
[0007] The system of this invention can help drivers quickly find parking spaces and optimize parking lot management efficiency.
[0008] This system achieves a qualitative leap in parking lot operation efficiency through full-chain optimization of perception, decision-making, and execution.
[0009] This invention achieves a transformation of parking lot operation from "experience-driven" to "data-intelligent driven" through a full-chain innovation of precise perception, intelligent decision-making, and humanized interaction. It sets a new industry benchmark in three dimensions: economic benefits, user experience, and social value, and provides a replicable example for the construction of smart city infrastructure. Attached Figure Description
[0010] Figure 1 This is a flowchart of the intelligent parking guidance process for this parking lot.
[0011] Figure 2 This is a flowchart of the overall workflow for intelligent decision-making. Detailed Implementation
[0012] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.
[0013] To address the problems of existing technologies, this invention provides a parking space intelligent guidance system based on edge computing. The parking space intelligent guidance system includes: a multi-source real-time data acquisition system, a distributed edge computing architecture module, and an intelligent decision-making core computing module. The multi-source real-time data acquisition system is the foundational layer of the intelligent parking guidance system. It achieves high-precision, low-latency parking space status perception and global environment modeling through the collaborative work of three major modules: IoT sensor deployment, computer vision enhancement, and auxiliary data source fusion. The distributed edge computing architecture module achieves efficient processing and real-time response to massive parking lot data through the organic combination of layered processing architecture and cloud collaborative computing. This architecture ensures low latency while making full use of cloud computing power for global optimization. The intelligent decision-making core computing module is the central hub of the intelligent parking guidance system. It consists of three major modules: spatiotemporal prediction model, dynamic path planning, and resource allocation optimization. It achieves efficient decision-making through the collaboration of multiple algorithms.
[0014] For the aforementioned multi-source real-time data acquisition system; It is the foundational layer of the intelligent parking guidance system. Through the collaborative work of three major modules—IoT sensor deployment, computer vision enhancement, and auxiliary data source fusion—it achieves high-precision, low-latency parking space status perception and global environment modeling. (1) IoT sensor deployment module, which is used to install low-power geomagnetic sensors (such as LoRaWAN nodes) on the ground of parking spaces to detect the occupancy status of parking spaces in real time; Based on geomagnetic sensing technology, the occupancy status of parking spaces is detected by utilizing the disturbance characteristics of vehicles (metal objects) on the Earth's magnetic field; combined with low-power wide-area network (LPWAN) technology, a distributed sensing network is constructed. First, magnetic signal baseline calibration is performed, which involves continuous sampling for 60 seconds in an empty parking space and calculating the average magnetic field strength. ;
[0015] The real-time detection logic is as follows:
[0016] like If the parking space is occupied for 5 seconds, it will be considered occupied. The IoT sensor deployment module can perform periodic self-tests (automatically recalibrate every 24 hours), and adjacent sensors can cross-verify (e.g., the status is confirmed only when two out of three adjacent sensors trigger).
[0017] (2) Computer vision enhancement module: using existing parking lot cameras + lightweight YOLOv7 model, the video stream is analyzed in real time through edge computing device (NVIDIA Jetson) to detect parking space status and track vehicle trajectory, and the accuracy is improved by fusing with sensor data; the vehicle movement trajectory is captured by the camera, and the parking space status recognition and vehicle tracking are realized by using deep learning model, which complements the geomagnetic sensor. First, a lightweight YOLOv7 model is used for parking space status detection, with the input resolution reduced to [resolution value missing]. The model size was compressed from 73MB to 18MB using channel pruning technology, and the output parking space ID and its occupancy status (0 / 1) were displayed. Based on the DeepSORT algorithm, RelD features were fused to achieve cross-camera target association. The NTP protocol was used for timestamp alignment, and the clock error between the sensor and the camera was <10ms. The mapping relationship between image pixel coordinates and physical coordinates was established by the checkerboard calibration method to achieve data synchronization. (3) Auxiliary data source fusion module, which integrates parking lot gate data (number of vehicles entering), payment system (estimated departure time), and third-party map API (surrounding traffic flow) to construct multi-dimensional data input; integrates multi-dimensional data inside and outside the parking lot, and improves the system's predictive ability through spatiotemporal alignment and feature fusion; the fusion method is as follows: First, define a unified spacetime coordinate system: , Parking lot CAD model For asynchronous data (such as traffic flow at 5-minute intervals), linear interpolation is used to generate continuous time series; then, a spatiotemporal feature matrix is constructed:
[0018] in, For the number of time slices, 3D features (including parking space status, vehicle speed, entrance flow, etc.); a multimodal fusion network employing an attention mechanism: ,
[0019] in, (Sensors, vision, auxiliary data) For global context vectors; The fusion process mainly involves first receiving heterogeneous data from multiple sources through a Kafka message queue, then cleaning the data (removing outliers such as instantaneous jumps in geomagnetic sensors), then performing spatiotemporal alignment (mapping all data onto a unified spatiotemporal network), then performing feature fusion to generate a joint feature vector, and finally outputting the status to update the global parking space status map. The multi-source real-time data acquisition system achieves multi-source complementarity, improves detection accuracy, and has an end-to-end latency of <350ms from data acquisition to status update. It supports deployment at the level of thousands of parking spaces and can be expanded by linearly adding edge nodes. For the distributed edge computing architecture module; By organically combining a layered processing architecture with cloud-based collaborative computing, the system achieves efficient processing and real-time response to massive amounts of parking lot data. This architecture ensures low latency while fully utilizing cloud computing power for global optimization. (1) Layered processing architecture The system adopts a three-tier computing architecture of "edge node - regional gateway - cloud" to offload data processing tasks in layers according to real-time requirements and computational complexity, thereby reducing end-to-end latency and bandwidth consumption. This means that data is processed locally and the computational load is balanced. Edge computing nodes (such as Raspberry Pi 4 + Intel Neural Compute Stick) are deployed every 10-20 parking spaces to process sensor / camera data in the local area in real time with a response latency of <200ms. The layered processing uses the following steps: First, at the edge layer, data filtering is performed to remove sensor noise data, such as instantaneous magnetic field fluctuations. Then, lightweight inference is performed, running a compressed version of the YOLOv7 model (18MB), outputting the parking space status and vehicle coordinates. Emergency obstacle avoidance commands are then executed, such as triggering an audible and visual alarm when a vehicle traveling in the wrong direction is detected. Finally, multi-node data fusion is performed at the gateway layer. ,
[0020] Based on improved A The algorithm generates the optimal path for sub-regions and converts LoRaWAN data packets into MQTT protocol mode. Finally, it performs global state management in the cloud, integrates data from multiple gateways to generate a digital twin of the parking lot, updates STGNN prediction model parameters using historical data, and formulates dynamic pricing and equipment maintenance plans according to long-term strategies. Edge node processing latency will be <150ms (including sensor data + video stream analysis), while the local decision-making ratio will be over 85%. Routine operations will not require cloud intervention, communication bandwidth will be saved, and the amount of uplink data will be reduced by 72% compared to the traditional architecture. (2) Cloud-based collaborative computing By employing a division of labor model of "edge preprocessing + cloud-based deep analysis," a balance between low-latency response and high-precision modeling is achieved, while a federated learning framework ensures data privacy. Preprocessed data from edge nodes is uploaded to the cloud (AWS IoT Greengrass) for global state analysis and large-scale learning inference, processing up to 10TB of data daily. The collaborative mechanism utilizes a task offloading strategy, namely:
[0021] in, This represents the computational complexity (FLOPs). Indicates uplink bandwidth (Mbps). Indicates the task deadline. Indicates local computing efficiency; Metadata, such as parking space status change statistics, is uploaded every 5 minutes to enable periodic synchronization and event trigger synchronization. When an abnormal event (such as a fire alarm) is detected, the entire dataset will be uploaded immediately. The core cloud-based functional modules are shown in the table below. Table 1. Core Functional Modules in the Cloud
[0022] The system will use AES-256 to encrypt sensor data, TLS1.3 to ensure communication links, encrypt data transmission, and perform homomorphic encryption on user license plate information. Only the hash value is retained in the cloud to protect privacy. At the same time, it will use blockchain-based RBAC (role-based access control) for permission management and access control. Taking vehicle entry guidance as an example, let's analyze the scenario. First, the edge layer responds. At this time, the entrance camera recognizes the license plate, and edge node A queries the local database for reservation information. If there is no reservation, it calls the local database D. The Lite algorithm generates a temporary path. Next, cloud-based collaborative work occurs: node A uploads vehicle characteristics (vehicle type, entry time) to the cloud, and the cloud prediction engine returns the vehicle's possible target area, such as near the office building elevator entrance. Then, dynamic adjustments are made: regional gateway B detects sudden congestion in the target area (three vehicles simultaneously searching for parking spaces), triggering a global replanning by issuing new path instructions to relevant edge nodes via MQTT. Finally, feedback optimization is performed. After guidance is complete, each node uploads process data to the cloud, and the federated learning framework updates the STGNN model parameters to improve the accuracy of the next prediction. The following explains how to optimize its performance using a dynamic task allocation algorithm:
[0023] when In case of emergency, some tasks are migrated to low-load nodes; regarding communication protocol optimization, in the normal state of adaptive compression, the system uses Protobuf binary encoding, while in an emergency, it switches to lossless Zstandard compression; in terms of energy, a power consumption model is used:
[0024] The CPU frequency (102GHz 2.4GHz) is adjusted in real time according to the load, achieving an energy saving rate of 38%. For the intelligent decision-making core computing module The intelligent decision-making core computing module is the central hub of the parking guidance system. It consists of three major modules: spatiotemporal prediction model, dynamic path planning, and resource allocation optimization. It achieves efficient decision-making through the collaboration of multiple algorithms. (1) Spatiotemporal prediction model A hybrid Transformer+Graph Neural Network model is employed to capture the spatiotemporal dependencies of parking space occupancy. Inputting historical occupancy data, real-time traffic flow, and surrounding events, along with over 50 feature dimensions, the model predicts parking space availability (MAE < 3%) for each area over the next 30 minutes. Spatially, the topological relationships between parking space nodes exhibit spatial dependencies, such as adjacent parking spaces or spaces within the same area. Temporally, the periodicity and trends of historical occupancy patterns create temporal dependencies. The spatiotemporal graph is defined as follows: Graph structure
[0025] in : A set of nodes (each parking space is a node); Edge set (connecting parking spaces with a distance of <20m); : Adjacency matrix (with weights, reflecting the strength of influence between parking spaces); Spatiotemporal attention mechanism:
[0026] The spatial mask matrix constrains the local attention range; the spatiotemporal convolution module is:
[0027] in This is the normalized adjacency matrix; For temporal convolutional networks, capture temporal patterns; The entire prediction process involves first preprocessing the data and then inputting the feature matrix. Given a historical time slice T, N parking spaces, and D-dimensional features; then, Z-score normalization is performed on each parking space, followed by multimodal fusion. External features This includes weather and holiday markers; then joint training is performed, with the loss function being...
[0028] Among them, the KL divergence term constrains the predicted distribution. With the true distribution Consistency; The output is the future. Occupancy probability of each parking space in a time slice ; The system updates model parameters incrementally every hour to adapt to dynamic changes. The system uses knowledge distillation technology to compress the teacher model into the student model. (2) Dynamic path planning module Based on the improved D Lite algorithm, and incorporating dynamic factors such as real-time parking space status, vehicle speed (obtained through camera tracking), and elevator / staircase location, the globally optimal path is updated every 5 seconds. It supports multi-objective optimization, namely shortest path, minimum turning radius, and low-congestion path. The system integrates the potential field method with multi-objective optimization to achieve real-time path planning in dynamic environments. It uses incremental search, updating only changed areas to avoid global replanning, and balances path length, number of turns, and congestion level. Its dynamic cost function is...
[0029] in Euclidean distance; Steering angle penalty (unit: radians); Congestion coefficient (calculated based on the output of the prediction model); Its potential field function is designed as follows, repulsive potential field:
[0030] Its attractive potential field (target parking space) is: Resultant force direction: ; The system uses B-spline curves to fit the original path, satisfying vehicle kinematic constraints, and can perform real-time collision detection.
[0031] in For the area occupied by vehicles, A collection of obstacles; (3) Resource allocation optimization module Intelligent parking space allocation is achieved by using a combinatorial auction algorithm. The allocation is modeled as a multi-agent bidding problem to achieve optimal global resource allocation. For reserved vehicles, the best parking space (<50m from the target elevator entrance) can be locked 15 minutes in advance. For vehicles on demand, the allocation is dynamically based on vehicle type (distinguishing between regular / SUV / accessible parking spaces). For special vehicles, license plate recognition is used to guide them to charging station parking spaces first. The auction model is shown below, where the valuation function for vehicle i relative to parking space j is:
[0032] in This represents the distance from the vehicle's current position to parking space j. Rate the convenience of parking space j; Winner determination rules: st
[0033] in Assign variables; The allocation weight for special vehicles (such as electric vehicles) has been increased as follows:
[0034] in This is the priority gain coefficient; The allocation process is as follows: First, the system preprocesses the parking spaces, classifying them into different types (regular / SUV / charging parking spaces). Vehicles that have reserved a space must lock it in 15 minutes in advance. In case of conflicts (multiple vehicles bidding for the same space), priority is used to rank the vehicles: Priority=0.6. Valuation +0.4 Arrival time in reverse order; The system can make flexible reservations, that is, reserve 5% of the parking space capacity for VIP vehicles that arrive temporarily, and can also implement anti-speculation penalties, that is, impose bidding cost penalties on vehicles that frequently change targets. Intelligent decision-making overall workflow such as Figure 2 As shown.
[0035] In addition, the core computing module for intelligent decision-making also has the following auxiliary modules: (4) Modal user guidance system 1) Smart terminal interaction The WeChat mini program / app integrates AR navigation (based on ARKit / ARCore), which uses the phone's camera to achieve 3D arrow projection navigation (positioning accuracy up to 0.3 meters) and visually marks key nodes, such as "the third parking space after turning left". It can also provide reverse car-finding guidance (automatically planning the route to the parking space when leaving). 2) On-site guidance hardware The on-site guidance hardware includes laser projection indicators (projecting dynamic arrows on the ground at intersections with visibility >20m), e-ink screen labels (displaying real-time status for each parking space with red / green LEDs and a reservation countdown), and a voice guidance system (playing prompts in key areas via directional speakers, such as "5 parking spaces remaining in Zone A"). 3) Vehicle-to-machine (V2X) Cooperative Solution It supports V2X communication (DSRC / C-V2X), connects directly with the vehicle navigation system, and enables pre-allocation of parking spaces before entry, dynamic updating of the route during driving (without driver operation) and emergency warnings (such as avoidance prompts for oncoming vehicles). (5) Management system and value-added services 1) Digital Twin Monitoring Platform A 3D visualization management system built on Unity can be used to view heat maps of congestion levels in various areas in real time, replay vehicle trajectories, support suspicious behavior analysis, and monitor the health status of equipment, such as automatically triggering alarms when sensors are offline. 2) Dynamic pricing strategy The system uses a game theory model to adjust pricing based on real-time demand: during peak hours (9:00-11:00), rates in core areas increase by 30%; during off-peak hours (14:00-16:00), rates in peripheral areas decrease by 50%; and for long-term parking exceeding 8 hours, tiered rates apply. 3) Value-added service integration Value-added services include automatic charging, smart car washing, and office traffic redirection; after an electric vehicle is parked, the robotic arm automatically connects to the charging pile to achieve automatic charging, or makes appointments for car washing services during off-peak hours, and supports pushing information about the nearest elevator entrance.
[0036] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A parking space intelligent guidance system based on edge computing, characterized in that, The intelligent parking guidance system includes: a multi-source real-time data acquisition system, a distributed edge computing architecture module, and an intelligent decision-making core computing module; The multi-source real-time data acquisition system is the foundational layer of the intelligent parking guidance system. It achieves high-precision, low-latency parking space status perception and global environment modeling through the collaborative work of three major modules: IoT sensor deployment, computer vision enhancement, and auxiliary data source fusion. The distributed edge computing architecture module achieves efficient processing and real-time response to massive parking lot data through the organic combination of layered processing architecture and cloud collaborative computing. This architecture ensures low latency while making full use of cloud computing power for global optimization. The intelligent decision-making core computing module is the central hub of the intelligent parking guidance system. It consists of three major modules: spatiotemporal prediction model, dynamic path planning, and resource allocation optimization. It achieves efficient decision-making through the collaboration of multiple algorithms.
2. The intelligent parking guidance system based on edge computing as described in claim 1, characterized in that, The multi-source real-time data acquisition system includes: (1) IoT sensor deployment module, which is used to install low-power geomagnetic sensors on the ground of parking spaces to detect the occupancy status of parking spaces in real time; Based on geomagnetic sensing technology, the occupancy status of parking spaces is detected by utilizing the disturbance characteristics of vehicles to the Earth's magnetic field; combined with low-power wide-area network technology, a distributed sensing network is constructed. First, magnetic signal baseline calibration is performed, which involves continuous sampling for 60 seconds in an empty parking space and calculating the average magnetic field strength. ; The real-time detection logic is as follows: like If it lasts for 5 seconds, it is considered that the parking space is occupied; The IoT sensor deployment module can perform periodic self-tests, and adjacent sensors can cross-verify each other.
3. The intelligent parking guidance system based on edge computing as described in claim 2, characterized in that, The multi-source real-time data acquisition system also includes: (2) Computer vision enhancement module: using the existing parking lot camera + lightweight YOLOv7 model, the video stream is analyzed in real time through edge computing device to detect parking space status and track vehicle trajectory, and the accuracy is improved by fusing with sensor data; the vehicle movement trajectory is captured by the camera, and the parking space status recognition and vehicle tracking are realized by using deep learning model, which complements the geomagnetic sensor. First, a lightweight YOLOv7 model is used for parking space status detection, with the input resolution reduced to [resolution value missing]. The model size was compressed from 73MB to 18MB using channel pruning technology, and the parking space ID and its occupancy status were output. Based on the DeepSORT algorithm, RelD features were fused to achieve cross-camera target association. The NTP protocol was used for timestamp alignment, and the clock error between the sensor and the camera was <10ms. The mapping relationship between image pixel coordinates and physical coordinates was established by the checkerboard calibration method to achieve data synchronization.
4. The intelligent parking guidance system based on edge computing as described in claim 3, characterized in that, The multi-source real-time data acquisition system also includes: (3) Auxiliary data source fusion module, which connects parking lot gate data, payment system, and third-party map API to construct multi-dimensional data input; integrates multi-dimensional data inside and outside the parking lot, and improves the system's predictive ability through spatiotemporal alignment and feature fusion; the fusion method is as follows: First, define a unified spacetime coordinate system: , Parking lot CAD model For asynchronous data, linear interpolation is used to generate continuous time series; then, a spatiotemporal feature matrix is constructed. in, For the number of time slices, 3D features; multimodal fusion network employing attention mechanism: ， in, , For global context vectors; The main steps of the fusion process are: first, receiving multi-source heterogeneous data through the Kafka message queue; then, cleaning the data, i.e. removing outliers; then, performing spatiotemporal alignment, i.e., mapping all data to a unified spatiotemporal network; then, performing feature fusion to generate a joint feature vector; and finally, outputting the status and updating the global parking space status map. The multi-source real-time data acquisition system achieves multi-source complementarity, improves detection accuracy, and has an end-to-end latency of <350ms from data acquisition to status update. It supports deployment at the level of thousands of parking spaces and can be expanded by linearly adding edge nodes.
5. The intelligent parking guidance system based on edge computing as described in claim 4, characterized in that, The distributed edge computing architecture module includes: (1) Layered processing architecture module The system adopts a three-tier computing architecture of "edge node - regional gateway - cloud", which offloads data processing tasks in layers according to real-time requirements and computational complexity, reducing end-to-end latency and bandwidth consumption. This means that data is processed locally and the computational load is balanced. Edge computing nodes are deployed every 10-20 parking spaces to process sensor / camera data in the local area in real time with a response latency of <200ms. The layered processing uses the following steps: First, at the edge layer, data filtering is performed to remove sensor noise data, such as instantaneous magnetic field fluctuations. Then, lightweight inference is performed, running a compressed version of the YOLOv7 model to output parking space status and vehicle coordinates. Next, an emergency obstacle avoidance command is executed, triggering an audible and visual alarm when a vehicle traveling in the wrong direction is detected. Finally, multi-node data fusion is performed at the gateway layer. ， Based on improved A The algorithm generates the optimal path for sub-regions and converts LoRaWAN data packets into MQTT protocol mode. Finally, it performs global state management in the cloud, integrates data from multiple gateways to generate a digital twin of the parking lot, updates STGNN prediction model parameters using historical data, and formulates dynamic pricing and equipment maintenance plans according to long-term strategies. Edge node processing latency will be <150ms, and the local decision-making ratio will be over 85%. Routine operations will not require cloud intervention, communication bandwidth will be saved, and the amount of uplink data will be reduced by 72% compared to the traditional architecture.
6. The intelligent parking guidance system based on edge computing as described in claim 5, characterized in that, The distributed edge computing architecture module also includes: (2) Cloud-based collaborative computing module By employing a division of labor model of "edge preprocessing + cloud-based deep analysis," a balance between low-latency response and high-precision modeling is achieved, while a federated learning framework ensures data privacy. Preprocessed data from edge nodes is uploaded to the cloud for global state analysis and large-scale learning inference, processing up to 10TB of data daily. The collaborative mechanism utilizes a task offloading strategy, namely: in, This represents the computational complexity (FLOPs). Indicates uplink bandwidth (Mbps). Indicates the task deadline. Indicates local computing efficiency; Metadata, such as parking space status change statistics, is uploaded every 5 minutes to ensure periodic synchronization and event trigger synchronization. When an abnormal event is detected, the entire dataset will be uploaded immediately. AES-256 is used to encrypt sensor data, TLS1.3 is used to ensure communication links, data is encrypted during transmission, and user license plate information is homomorphically encrypted. Only the hash value is retained in the cloud, thus protecting privacy. At the same time, RBAC based on blockchain is used for permission management and access control. When guiding vehicles into the parking lot, scenario analysis is performed. First, the edge layer responds. At this time, the entrance camera recognizes the license plate, and edge node A queries the local database for reservation information. If there is no reservation, it calls the local database D. The Lite algorithm generates a temporary path; secondly, cloud collaboration occurs, with node A uploading vehicle features to the cloud, and the cloud prediction engine returning the vehicle's possible target area, such as near the office building elevator entrance; thirdly, dynamic adjustments are made, with regional gateway B detecting sudden congestion in the target area, triggering global replanning and issuing new path instructions to relevant edge nodes via MQTT; finally, feedback optimization is performed, and after guidance is completed, each node uploads process data to the cloud, and the federated learning framework updates the STGNN model parameters to improve the accuracy of the next prediction; Performance optimization using dynamic task allocation algorithms: when In case of emergency, some tasks are migrated to low-load nodes; regarding communication protocol optimization, in the normal state of adaptive compression, the system uses Protobuf binary encoding, while in an emergency, it switches to lossless Zstandard compression; in terms of energy, a power consumption model is used: The CPU frequency is adjusted in real time according to the load, achieving an energy saving rate of up to 38%.
7. The intelligent parking guidance system based on edge computing as described in claim 6, characterized in that, The intelligent decision-making core computing module includes: (1) Spatiotemporal prediction model A hybrid Transformer+Graph Neural Network model is employed to capture the spatiotemporal dependencies of parking space occupancy. Inputting historical occupancy data, real-time traffic flow, and surrounding events, along with over 50 feature dimensions, the model predicts parking space availability in various areas for the next 30 minutes. Spatially, the topological relationships between parking space nodes exhibit spatial dependencies, including adjacent spaces and spaces within the same area. Temporally, the periodicity and trends of historical occupancy patterns create temporal dependencies. The spatiotemporal graph is defined as follows: Graph structure in : A set of nodes, where each parking space is a node; Edge set, connecting parking spaces with a distance of less than 20m; Adjacency matrix, with weights, reflects the strength of influence between parking spaces; Spatiotemporal attention mechanism: The spatial mask matrix constrains the local attention range; the spatiotemporal convolution module is: in This is the normalized adjacency matrix; For temporal convolutional networks, capture temporal patterns; The entire prediction process involves first preprocessing the data and then inputting the feature matrix. Given a historical time slice T, N parking spaces, and D-dimensional features; then, Z-score normalization is performed on each parking space, followed by multimodal fusion: External features This includes weather and holiday markers; then joint training is performed, with the loss function being... Among them, the KL divergence term constrains the predicted distribution. With the true distribution Consistency; The output is the future. Occupancy probability of each parking space in a time slice ; The system updates model parameters incrementally every hour to adapt to dynamic changes. The system uses knowledge distillation technology to compress the teacher model into the student model.
8. The intelligent parking guidance system based on edge computing as described in claim 7, characterized in that, The intelligent decision-making core computing module also includes: (2) Dynamic path planning module Based on the improved D Lite algorithm, and incorporating dynamic factors such as real-time parking space status, vehicle speed, and elevator / staircase location, the globally optimal path is updated every 5 seconds. It supports multi-objective optimization, namely shortest path, minimum turning radius, and low-congestion path. The system integrates the potential field method with multi-objective optimization to achieve real-time path planning in dynamic environments. It uses incremental search, updating only changed areas to avoid global replanning and balances path length, number of turns, and congestion level. Its dynamic cost function is... in Euclidean distance; Steering angle penalty (unit: radians); Congestion coefficient; Its potential field function is designed as follows, repulsive potential field: Its attractive potential field (target parking space) is: Resultant force direction: ; The system uses B-spline curves to fit the original path, satisfying vehicle kinematic constraints, and can perform real-time collision detection. in For the area occupied by vehicles, A collection of obstacles.
9. The intelligent parking guidance system based on edge computing as described in claim 8, characterized in that, The intelligent decision-making core computing module also includes: (3) Resource allocation optimization module A combinatorial auction algorithm is used to achieve intelligent parking space allocation. The parking space allocation is modeled as a multi-agent bidding problem to achieve optimal global resource allocation. For reserved vehicles, the best parking space can be locked 15 minutes in advance; for vehicles on demand, the allocation is dynamically based on vehicle type; and for special vehicles, license plate recognition is used to guide them to charging station parking spaces first. The auction model is shown below, where the valuation function for vehicle i relative to parking space j is: in This represents the distance from the vehicle's current position to parking space j. Rate the convenience of parking space j; Winner determination rules: s.t. in Assign variables; The allocation weight for special vehicles (such as electric vehicles) has been increased as follows: in This is the priority gain coefficient; The allocation process is as follows: First, the system preprocesses the parking spaces, classifying them into different types. Vehicles that have reserved a space must lock it in 15 minutes in advance. In case of conflicts (multiple vehicles bidding for the same space), priority is used to rank the vehicles: Priority=0.
6. Valuation +0.4 Arrival time in reverse order; The system can make flexible reservations, that is, reserve 5% of the parking space capacity for VIP vehicles that arrive temporarily, and can also implement anti-speculation penalties, that is, impose bidding cost penalties on vehicles that frequently change targets.
10. The intelligent parking guidance system based on edge computing as described in claim 9, characterized in that, The system belongs to the field of intelligent parking guidance technology.