A method and system for predicting parking lot empty parking spaces

Abstract

The present application provides a method and system for predicting vacant parking spaces in a parking lot. The method comprises: generating a spatiotemporal topological map of a special-shaped parking lot by fusing ramp curvature parameters extracted from the topological structure data of the special-shaped parking lot with historical vehicle turning radius data; aggregating the parking space occupancy status data of all special-shaped parking lots and the spatiotemporal topological map of the special-shaped parking lot through a federated transfer learning framework to obtain target feature extraction model parameters; combining the time-varying characteristics of the channel flow of the special-shaped parking lot in the spatiotemporal topological map of all special-shaped parking lots with the target feature extraction model parameters to generate a dynamic parking space allocation strategy; mapping the dynamic parking space allocation strategy to the real-time navigation path of the spatiotemporal topological map of the special-shaped parking lot so that the vehicles to be parked can reach the vacant parking spaces through the real-time navigation path. The present application generates a dynamic parking space allocation strategy and performs real-time navigation by fusing the topological structure of the special-shaped parking lot with historical data, thereby optimizing parking space utilization and parking efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer application, and in particular, relates to a parking lot idle parking space prediction method and system. BACKGROUND

[0002] In a special-shaped parking lot (such as a spiral ramp parking lot, a multi-layer stereo parking lot, etc.), due to reasons such as curved lanes, complex ramps, dynamic changes in vehicle flow, and difficulty in intuitively identifying parking spaces, the time for vehicles to find idle parking spaces increases, thereby affecting the overall operation efficiency of the parking lot. In order to cope with this challenge, a technical solution is needed that can monitor and analyze the occupancy status of each parking space in the parking lot in real time, in order to improve parking space utilization, reduce vehicle detention time in the parking lot, and optimize the allocation of space resources in the parking lot.

[0003] The existing scheme uses sensor technology and wireless communication networks to monitor parking space occupancy. Specifically, the existing scheme can collect real-time parking occupancy status data by installing a geomagnetic sensor or an ultrasonic sensor at each parking space, and transmit these data to a central control system for processing.

[0004] The existing scheme can usually only provide parking occupancy status data, resulting in insufficient responsiveness to dynamic traffic flow changes. Moreover, the existing scheme is difficult to meet the demand for joint management of multiple special-shaped parking lots, and cannot adjust parking allocation strategies and navigation paths in real time, resulting in low efficiency of vehicles searching for idle parking spaces during peak hours or sudden situations. SUMMARY

[0005] The present application provides a parking lot idle parking space prediction method to solve the problem of insufficient responsiveness to dynamic traffic flow changes, difficulty in joint management of multiple special-shaped parking lots, and insufficient real-time adjustment of parking allocation strategies and navigation paths, resulting in low efficiency of vehicles searching for idle parking spaces during peak hours or sudden situations.

[0006] In a first aspect, the present application provides a parking lot idle parking space prediction method, comprising: obtaining parking occupancy status data and special-shaped parking lot topology structure data of a plurality of special-shaped parking lots;

[0007] For each special-shaped parking lot, a special-shaped parking lot space-time topology graph is generated by fusing a ramp curvature parameter extracted from the special-shaped parking lot topology structure data and historical vehicle turning radius data.

[0008] The parking occupancy status data and the special-shaped parking lot space-time topology graph of all the special-shaped parking lots are aggregated through a federated transfer learning framework to obtain target feature extraction model parameters.

[0009] The time variation characteristics of the channel flow of the special-shaped parking lot in the special-shaped parking lot space-time topology graph and the target feature extraction model parameters are combined to generate a dynamic parking space allocation strategy.

[0010] The dynamic parking space allocation strategy is mapped to the real-time navigation path of the special-shaped parking lot space-time topology graph, so that the vehicle to be parked reaches the idle parking space through the real-time navigation path.

[0011] Optionally, the federated transfer learning framework aggregates the parking space occupancy state data of all the special-shaped parking lots and the special-shaped parking lot space-time topology graph to obtain the target feature extraction model parameters, including:

[0012] The parking space occupancy state data of each special-shaped parking lot and the special-shaped parking lot space-time topology graph are input into the corresponding local node in the federated transfer learning framework, so as to store the parking space occupancy state data and the special-shaped parking lot space-time topology graph of the corresponding special-shaped parking lot in each local node. One local node corresponds to one special-shaped parking lot;

[0013] In each local node, the spiral ramp geometry parameters in the special-shaped parking lot space-time topology graph are extracted, including lane curvature;

[0014] Based on the parking space occupancy state data and the spiral ramp geometry parameters, a local feature extraction model on the local node is trained to obtain local feature extraction model parameters related to lane curvature;

[0015] In the central server in the federated transfer learning framework, all the local feature extraction model parameters are aggregated to generate global feature extraction model parameters;

[0016] The global feature extraction model parameters are subjected to convergence processing to obtain target feature extraction model parameters.

[0017] Optionally, the aggregation of all the local feature extraction model parameters to generate global feature extraction model parameters includes:

[0018] All the local feature extraction model parameters are fused by using a weighted average strategy to generate intermediate feature extraction model parameters. The weighted average strategy includes a weight distribution scheme determined based on the data volume, data quality or model performance indicators of each local node;

[0019] The intermediate feature extraction model parameters are verified, and the verified intermediate feature extraction model is determined as the global feature extraction model parameters.

[0020] Optionally, the time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0021] The time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0022] The time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0023] The time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0024] The time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0025] The time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0026] The time variation characteristics of the channel flow of all the special-shaped parking lots in the special-shaped parking lot space-time topology graph are combined with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including:

[0027] Optionally, the local feature extraction model on the local node is trained based on the parking space occupancy state data and the spiral ramp geometric parameters to obtain local feature extraction model parameters related to lane curvature, including:

[0028] The local feature extraction model is initialized.

[0029] The input data set of the local feature extraction model is constructed based on the parking space occupancy state data and the spiral ramp geometric parameters.

[0030] The input data set is input into the initialized local feature extraction model to obtain the local feature extraction model parameters related to the lane curvature through training.

[0031] Optionally, the parking space occupancy state data of a plurality of special-shaped parking lots is obtained, including:

[0032] The flexible pressure film sensor array acquires the pressure distribution data of each parking space in the corresponding special-shaped parking lot in real time. One special-shaped parking lot corresponds to one flexible pressure film sensor array. Flexible pressure film sensors are arranged on each parking space in the special-shaped parking lot, and flexible pressure film sensors are arranged on the curved lane in the special-shaped parking lot. The flexible pressure film sensor array is arranged in a conformal manner according to the parking space contour and the curved lane contour of the special-shaped parking lot.

[0033] The pressure distribution data of the parking space is converted into parking space occupancy state data.

[0034] Optionally, the pressure distribution data of the parking space is converted into parking space occupancy state data, comprising:

[0035] According to a preset analysis rule, the pressure distribution data is analyzed to obtain vehicle parking position and weight distribution information;

[0036] According to the vehicle parking position and the weight distribution information, a parking space occupancy identifier is generated;

[0037] The parking space occupancy identifier is combined with the vehicle weight distribution information to generate parking space occupancy state data.

[0038] In a second aspect, the application provides a rain alarm valve group operation state remote monitoring system, comprising:

[0039] An acquisition module is configured to acquire parking space occupancy state data and special-shaped parking lot topology structure data of a plurality of special-shaped parking lots.

[0040] A generation module is configured to generate a special-shaped parking lot space-time topology graph by fusing a ramp curvature parameter extracted from the special-shaped parking lot topology structure data and historical vehicle turning radius data for each special-shaped parking lot.

[0041] An extraction module is configured to aggregate the parking space occupancy state data and the special-shaped parking lot space-time topology graph of all the special-shaped parking lots through a federal transfer learning framework to obtain target feature extraction model parameters.

[0042] A generation module is configured to generate a dynamic parking space allocation strategy by combining the time variation characteristics of the channel flow of the special-shaped parking lot in all the special-shaped parking lot space-time topology graphs and the target feature extraction model parameters.

[0043] A mapping module is configured to map the dynamic parking space allocation strategy to a real-time navigation path of the special-shaped parking lot space-time topology graph, so that a vehicle to be parked reaches an idle parking space through the real-time navigation path.

[0044] In a third aspect, the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component, and realize the method for predicting idle parking spaces of a parking lot according to any one of the first aspect.

[0045] In a fourth aspect, the present application provides a computer storage medium, which stores a computer program; when the computer program is executed by a computer, the method for predicting idle parking spaces of a parking lot according to any one of the first aspect is realized.

[0046] In the present application, a method for predicting idle parking spaces of a parking lot is provided, which comprises: acquiring parking space occupancy state data and topology structure data of a plurality of special-shaped parking lots; for each special-shaped parking lot, generating a special-shaped parking lot space-time topology graph by fusing a ramp curvature parameter extracted from the topology structure data of the special-shaped parking lot and historical vehicle turning radius data; aggregating parking space occupancy state data and the special-shaped parking lot space-time topology graph of all the special-shaped parking lots by using a federal transfer learning framework to obtain target feature extraction model parameters; generating a dynamic parking space allocation strategy by combining time variation characteristics of channel flow of the special-shaped parking lot in all the special-shaped parking lot space-time topology graphs and the target feature extraction model parameters; and mapping the dynamic parking space allocation strategy to a real-time navigation path of the special-shaped parking lot space-time topology graph, so that a vehicle to be parked reaches an idle parking space through the real-time navigation path.

[0047] The application realizes comprehensive integration and utilization of multi-source data by obtaining parking space occupation state data and topological structure data of multiple special-shaped parking lots; accurate special-shaped parking lot space-time topology graphs are generated by fusing ramp curvature parameters and historical vehicle turning radius data, providing a reliable basis for dynamic parking space allocation; the data of multiple parking lots are aggregated by a federal transfer learning framework to generate target feature extraction model parameters, supporting collaborative management and resource optimization of multiple special-shaped parking lots. Dynamic parking space allocation strategies are generated by combining the time variation characteristics of channel flow with the target feature extraction model parameters, improving the real-time response capability to traffic flow changes. The dynamic parking space allocation strategies are mapped into real-time navigation paths, enabling vehicles to be parked to quickly reach idle parking spaces, improving parking efficiency, especially during peak periods or in emergency situations. By adjusting the parking space allocation strategy and navigation path in real time, the complex management requirements of multiple special-shaped parking lots are effectively met, and the intelligent level of the overall system is improved. Further, the parking space occupation state data and space-time topology graphs of each special-shaped parking lot are input into the local nodes of the federal transfer learning framework, each local node stores the data of the corresponding parking lot, and extracts the spiral ramp geometric parameters (such as lane curvature) in the space-time topology graph; based on the parking space occupation state data and the spiral ramp geometric parameters, a local feature extraction model is trained to obtain local model parameters related to the lane curvature; in the central server, all local model parameters are aggregated using a weighted average strategy to generate intermediate feature extraction model parameters, and the global feature extraction model parameters are determined through verification, and the target feature extraction model parameters are finally obtained. Through the federal transfer learning framework, distributed storage and collaborative training of multi-special-shaped parking lot data are realized, the accuracy and adaptability of the feature extraction model are improved by combining the spiral ramp geometric parameters and the parking space occupation state data; the local model parameters are aggregated using the weighted average strategy to ensure the robustness and generalization ability of the global model, and the finally generated target feature extraction model parameters can effectively support the optimization of dynamic parking space allocation and navigation strategies, improving the intelligent level and efficiency of parking management.

[0048] These aspects or other aspects of the application will be more apparent in the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and other drawings can also be obtained by those skilled in the art without any creative effort.

[0050] Figure 1 A flowchart of a parking lot idle parking space prediction method provided by an embodiment of the application;

[0051] Figure 2 A structural schematic diagram of a parking lot idle parking space prediction system provided for an embodiment of the present application is shown in the figure.

[0052] Figure 3 A structural schematic diagram of a computing device provided for an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0053] In order to enable personnel in the technical field to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application.

[0054] In some processes described in the specification and claims of the present application and the above-mentioned drawings, a plurality of operations appearing in a specific order are included, but it should be clearly understood that these operations can be executed or performed in parallel without the order in which they appear in the text, and the serial numbers of the operations such as 11, 12, etc. are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and the operations can be executed in sequence or in parallel. It should be noted that the "first", "second", etc. described herein are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence. Also, "first" and "second" are different types.

[0055] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0056] Figure 1 A flowchart of a parking lot idle parking space prediction method provided for an embodiment of the present application is shown in the figure, which includes: Figure 1

[0057] S11, acquiring parking space occupancy state data and topology structure data of a plurality of special-shaped parking lots.

[0058] The parking space occupancy state data includes the real-time occupancy state (such as idle, occupied, reserved, etc.) of each parking space in the parking lot, which is usually collected by sensors, cameras or Internet of Things devices. The topology structure data of the special-shaped parking lot is the physical structure information describing the parking lot, including the geometric characteristics of the driveway, the ramp, the turning area, the entrance and exit position, etc., specifically including the ramp curvature parameter, which is usually obtained by architectural design drawings or three-dimensional modeling. The ramp can refer to the driveway passed by the entrance, and also can refer to the parking ramp with parking spaces.

[0059] ​In the embodiments of the present application, the flexible pressure film sensor array can be used to collect real-time parking space pressure data in the shaped parking lot, and multi-modal data fusion can be performed by combining cameras and Internet of Things devices to determine the real-time occupancy state of the parking space. For example, if the pressure data corresponding to the parking space is greater than or equal to the preset pressure threshold, it indicates that there is a vehicle on the parking space, so the real-time occupancy state of the parking space is occupied. If the pressure data corresponding to the parking space is less than the preset pressure threshold, it indicates that there is no vehicle on the parking space, and the camera can be used to confirm that there is no vehicle on the parking space, and then the Internet of Things device can be used to check whether the parking space is reserved in the current time period. If yes, the real-time occupancy state is reserved, otherwise it is idle. The present application can generate a digital twin model of the parking lot based on a building information model or a three-dimensional laser scan. The present application can use the least squares method to fit a circular arc to determine the lane curvature radius, and determine the slope of the ramp, the entrance and exit positions, etc. according to the triangulation method.

[0060] For example, a certain commercial complex contains three shaped parking lots (A, B and C), and the parking spaces of the three shaped parking lots are all installed with flexible pressure film sensors with a resolution of 5cm x 5cm. The spiral ramp of parking lot A has a curvature radius of 3.2m and a slope of 12° determined by modeling. After the sensor data is analyzed by the edge computing node, it is determined that the parking space A-07 has a pressure value of 2100kg (above the threshold) at 12:00, which is marked as "large vehicle occupancy". The data is uploaded to the cloud through the protocol, and the data of parking lots B and C form an initial data set.

[0061] S12, for each shaped parking lot, generate a shaped parking lot space-time topology graph by fusing the ramp curvature parameters extracted from the shaped parking lot topology structure data and the historical vehicle turning radius data.

[0062] The ramp curvature parameter is used to describe the geometric characteristics of the parking lot ramp, including the curvature radius, slope, length, etc., which are used to evaluate the difficulty and safety of vehicle driving. The historical vehicle turning radius data is used to record the actual turning radius of different types of vehicles in the parking lot, which is used to optimize path planning and parking space allocation. The shaped parking lot space-time topology graph is a dynamic graph structure that combines the physical topology information of the parking lot and the traffic change characteristics in the time dimension, which is used to describe the possible driving path and bottleneck area of the vehicle in the parking lot.

[0063] In the embodiments of the present application, in order to fuse the ramp curvature parameter and the historical vehicle turning radius data, the curvature radius, slope and length can be encoded into a vector, for example, 3.2m, 12°, 50m, in order to eliminate the dimension, normalization processing can also be performed to obtain [0.32, 0.12, 0.5]. The embodiments can also divide the vehicles into small cars (turning radius 4m), medium cars (6m) and large cars (8m) based on K-means clustering. If the ramp curvature radius is greater than or equal to the vehicle turning radius, the vehicle is marked as “passable”. Then, the embodiments can use a graph database to store a dynamic graph structure, the nodes are parking spaces, passages, etc., and the edge weights are real-time travel times. The real-time travel time of the ramp provides data support for the calculation of subsequent road congestion and road scoring.

[0064] For example, the spiral ramp with a curvature radius of 3.2m in parking lot A matches the historical vehicle data, and only allows small cars with a turning radius of ≤3.5m to pass. In the spatiotemporal topology graph, the travel time of the ramp at 12:00 is the passage length 50m ÷ the reference vehicle speed 1.39m / s × (1 + congestion sensitivity parameter 0.8 × current flow 15 vehicles / channel capacity 20 vehicles) = 48 seconds. Correspondingly, the congestion coefficient is 15 / 20 = 0.6, so the dynamic weight adjustment can be triggered, resulting in a decrease in the score of the path.

[0065] S13, aggregate the parking space occupancy state data of all special-shaped parking lots and the spatiotemporal topology graph of the special-shaped parking lots through a federated transfer learning framework to obtain target feature extraction model parameters.

[0066] The federated transfer learning framework is a distributed machine learning framework that allows multiple local nodes to collaboratively train models without sharing raw data, protecting data privacy. The target feature extraction model parameters are global model parameters generated through federated learning, which are used to extract common features of parking lot data to support dynamic parking space allocation and path planning.

[0067] In this embodiment, each parking lot serves as a federated learning node for distributed training. The input of the graph attention network is a spatio-temporal topology graph, and the output is the associated features of lane curvature and parking space occupancy. The loss function of this network can be the sum of regularization and mean square error. The central server adopts a weighted average strategy, which can determine the weight size of the model parameters provided by this federated learning node according to the data volume, data quality and model accuracy. For example, in terms of quantity, parking lot A has 1000, parking lot B has 800, and parking lot C has 1200. In terms of data quality, the data quality of parking lot A is 0.9, the data quality of parking lot B is 0.7, and the data quality of parking lot C is 0.8. In terms of model accuracy, the model accuracy of parking lot A is 95%, the model accuracy of parking lot B is 90%, and the model accuracy of parking lot C is 92%. Finally, the weight corresponding to parking lot A is 0.35, the central server aggregates the model parameters of all parking lots to generate global model parameters, and the validation set accuracy is improved to support cross-parking lot feature generalization.

[0068] S14, combine the time-varying characteristics of the channel flow of the special-shaped parking lot in the spatio-temporal topology graph of the special-shaped parking lot with the target feature extraction model parameters to generate a dynamic parking space allocation strategy.

[0069] The time-varying characteristics of the channel flow describe the distribution of vehicle flow in different time periods in the parking lot, such as congestion areas during peak hours and idle areas during off-peak hours. The dynamic parking space allocation strategy dynamically adjusts the parking space allocation scheme according to real-time traffic flow and target feature extraction model parameters to optimize parking efficiency and resource utilization.

[0070] For example, when the score of the idle parking space 12 of parking lot C is 3.0 and the score of the idle parking space 05 of parking lot B is 2.2, the dynamic allocation strategy preferentially guides the vehicle to be parked to parking lot C.

[0071] S15, map the dynamic parking space allocation strategy to the real-time navigation path of the spatio-temporal topology graph of the special-shaped parking lot, so that the vehicle to be parked reaches the idle parking space through the real-time navigation path.

[0072] The real-time navigation path is a vehicle driving path generated according to the dynamic parking space allocation strategy, including turn prompts, slope prompts and parking space guidance information.

[0073] In the embodiments of the present application, the improved A* algorithm can be used, and the cost function is the sum of the actual same row time, the estimated remaining time and the path accessibility score.

[0074] For example, after the vehicle enters parking lot C, the ramp travel time is determined to be 48 seconds based on the real-time topological map, and the planned path is the entrance, ramp lane C3, and then to the empty parking space 12. The navigation instructions can be updated synchronously through the display screen and the vehicle terminal, prompting "front spiral ramp, speed limit 5 km / h".

[0075] The following is a specific example:

[0076] Assume that there are three special-shaped parking lots (A, B, and C) in a city, each with complex topological structures containing multiple spiral ramps and narrow turning areas. During peak hours, parking lots A and B have high occupancy rates, while parking lot C is relatively empty. The occupancy state data of the three parking lots is obtained through a sensor network, and the topological structure data is generated through digital modeling. The ramp curvature parameters of parking lots A, B, and C are extracted, and combined with historical vehicle turning radius data, the respective spatiotemporal topological maps are generated. Through a federated transfer learning framework, the data of the three parking lots is aggregated, and the target feature extraction model parameters are trained. Combined with the time-varying features in the spatiotemporal topological map (such as the congestion of parking lots A and B during peak hours), a dynamic parking space allocation strategy is generated, which preferentially guides vehicles to parking lot C. The dynamic parking space allocation strategy is mapped into the real-time navigation path, and the navigation system guides the driver to efficiently reach the empty parking space in parking lot C.

[0077] By performing S11-S15, the present embodiment achieves coordinated management and resource optimization of multiple special-shaped parking lots through deep fusion of multi-source heterogeneous data, a federated transfer learning framework, and a dynamic path planning algorithm. The system can real-time perceive and respond to dynamic traffic flow changes, dynamically generate and adjust parking space allocation strategies and navigation paths based on global feature extraction model parameters and spatiotemporal topological maps, thereby improving parking efficiency and resource utilization. Especially during peak hours or sudden scenarios, the system effectively alleviates parking lot congestion through intelligent path guidance and parking space allocation mechanisms, providing drivers with more efficient, intelligent, and convenient parking service experiences.

[0078] In one possible embodiment, S13 aggregates the parking space occupancy state data and the spatiotemporal topological map of all special-shaped parking lots through a federated transfer learning framework to obtain target feature extraction model parameters, including:

[0079] Step 131 inputs the parking space occupancy state data and the spatiotemporal topological map of each special-shaped parking lot into the corresponding local node in the federated transfer learning framework to store the parking space occupancy state data and the spatiotemporal topological map of the corresponding special-shaped parking lot in each local node, with one local node corresponding to one special-shaped parking lot.

[0080] Among them, the local node is a distributed computing unit in the federal transfer learning framework, each node is responsible for storing and processing the data of one special parking lot, ensuring data privacy and security.

[0081] In practical applications, each parking lot is deployed with a corresponding edge computing node and configured with a corresponding encryption communication module. The data processing environment of different parking lots can also be isolated by virtualization technology. The central server issues initial model parameters, and each node loads a local model with the same structure.

[0082] Step 132, in each local node, extract the spiral ramp geometric parameters in the special parking lot spatio-temporal topology graph, including lane curvature.

[0083] Among them, the spiral ramp geometric parameters are used to describe the geometric characteristics of the parking lot spiral ramp, including lane curvature, slope, length, etc., which are used to evaluate the difficulty and safety of vehicle driving. Lane curvature is a parameter used to describe the degree of curvature of the ramp, which directly affects the turning radius and driving speed of the vehicle.

[0084] The present application embodiment can extract the ramp geometric data through the analysis tool, fit the ramp center line circle through the least square method, calculate the curvature radius, and normalize the curvature, slope, and length to the range of [0, 1]. The center line of the spiral ramp of parking lot A is fitted to get the curvature radius r=3.2m ( ), the slope is 12°, and the length is 50m, which is normalized to the vector [0.3125, 0.6, 0.5] after encoding.

[0085] Step 133, based on the parking space occupancy state data and the spiral ramp geometric parameters, train the local feature extraction model on the local node to obtain the local feature extraction model parameters related to the lane curvature.

[0086] Among them, the local feature extraction model is a machine learning model (such as convolutional neural network or graph neural network) used to extract features from parking space occupancy state data and geometric parameters. The local feature extraction model parameters are the weight and bias parameters of the local feature extraction model after training, which are used to describe the relationship between lane curvature and parking space occupancy state.

[0087] In the present application embodiment, the model architecture on each node is a 3-layer graph attention network with 128 hidden units per layer and ReLU activation function. The past 7 days of parking space occupancy time series data (5-minute interval) and ramp parameters are concatenated as graph node features, then used as training data, and trained using an optimizer with a learning rate of 0.001 for 50 rounds until convergence. For example, the local model of parking lot A can obtain parameters including attention layer weights and output layer bias, which is used to adjust the baseline of the predicted value.

[0088] Step 134, in the central server within the federated transfer learning framework, aggregate all local feature extraction model parameters to generate global feature extraction model parameters.

[0089] Wherein, the global feature extraction model parameters are generated by aggregating the model parameters of multiple local nodes to extract the common features of the parking lot data. The weighted average strategy is a model parameter aggregation method that assigns weights based on the data volume, data quality, or model performance of each local node to ensure the robustness of the global model. For example, the convolution kernel weights, attention coefficients, and other parameters are weighted and averaged according to the weights to generate global parameters.

[0090] Step 135, perform convergence processing on the global feature extraction model parameters to obtain target feature extraction model parameters.

[0091] Wherein, the convergence processing is achieved by iterative optimization to make the model parameters reach a stable state, ensuring the generalization ability and performance of the model. The target feature extraction model parameters are the final model parameters after convergence processing, which are used to support dynamic parking space allocation and path planning.

[0092] Here is a specific example:

[0093] Suppose there are 3 irregular parking lots (A, B, C) in a city, each with complex topological structure, including multiple spiral ramps and narrow turning areas. The parking space occupancy state data and spatio-temporal topology graph of parking lots A, B, and C are input into 3 local nodes of the federated transfer learning framework, and each node stores the data of the corresponding parking lot. In each local node, the spiral ramp geometric parameters (such as lane curvature, slope) in the spatio-temporal topology graph are extracted. Based on the parking space occupancy state data and spiral ramp geometric parameters, a local feature extraction model is trained to obtain local model parameters related to lane curvature. In the central server, the weighted average strategy is used to aggregate the model parameters of the 3 local nodes to generate global feature extraction model parameters. The global feature extraction model parameters are subjected to convergence processing to obtain target feature extraction model parameters, which are used to support dynamic parking space allocation and path planning.

[0094] By performing steps 131-135, the present embodiment introduces a federated transfer learning framework to achieve distributed training and collaborative optimization of multi-shape parking lot data, effectively solving the data island problem in traditional centralized learning. By fusing the spiral ramp geometric parameters (such as lane curvature, slope) and real-time parking space occupancy state data, the system can accurately capture the dynamic characteristics and spatial constraints of the parking lot, and then generate high-precision target feature extraction model parameters. This parameter not only supports real-time optimization of dynamic parking allocation strategies, but also enables intelligent path planning by combining the spatio-temporal topology graph, thereby improving parking efficiency and resource utilization. Especially during peak hours or sudden scenarios, the system effectively alleviates parking congestion by adaptively adjusting parking allocation and navigation paths, providing drivers with a more efficient, intelligent, and convenient parking service experience while ensuring data privacy and security.

[0095] In one possible embodiment, step 134, aggregating all local feature extraction model parameters to generate global feature extraction model parameters, includes:

[0096] Step a1, using a weighted average strategy to fuse all local feature extraction model parameters to generate intermediate feature extraction model parameters.

[0097] The weighted average strategy is a model parameter aggregation method that assigns weights to each local node's model parameters, calculates the weighted average, and generates global model parameters. The local feature extraction model parameters are the model weight and bias parameters obtained after training from each local node, which describe the characteristics of local data. The intermediate feature extraction model parameters are preliminary global model parameters generated by the weighted average strategy and have not been verified and optimized.

[0098] Step a2, wherein the weighted average strategy includes a weight allocation scheme based on the data volume, data quality, or model performance indicators of each local node.

[0099] The data volume is based on the data size of the parking space occupancy state data and the spatio-temporal topology graph stored in the local node. The data quality is based on the accuracy, completeness, and timeliness of the data, usually evaluated by data cleaning and validation indicators. The model performance indicator refers to the performance of the local model on the validation set, such as accuracy, recall rate, F1 score, etc. The weight allocation scheme assigns higher weights to local nodes with larger data volumes, for example, node A has twice the data volume of node B, so node A has twice the weight of node B. The data quality is evaluated by data cleaning and validation indicators (such as missing value proportion, abnormal value proportion), and nodes with higher quality are assigned higher weights. According to the performance (such as accuracy) of the local model on the validation set, nodes with better performance are assigned higher weights. Finally, the weight allocation scheme is normalized to ensure that the sum of all weights is 1.

[0100] Step a3, validate the intermediate feature extraction model parameters, and determine the validated intermediate feature extraction model as the global feature extraction model parameters.

[0101] Wherein, the validation is to evaluate the performance of the intermediate feature extraction model parameters through the validation set, to ensure its generalization ability and robustness. The validation of the intermediate feature extraction model parameters is to evaluate the performance of the intermediate model by using an independent validation set (data not involved in training), and to calculate indicators such as accuracy, recall rate, F1 score, etc. If the model performance does not reach the preset threshold, return to step a1 to adjust the weight allocation scheme or re-aggregate the local model parameters. If the model performance meets the standard, the intermediate feature extraction model parameters are determined as the global feature extraction model parameters, which are used for subsequent dynamic parking space allocation and path planning.

[0102] For example, by weighted average fusion of multi-node features, combined with dynamic weight allocation of data quality and model performance, the system generates intermediate parameters that take into account global commonality and local characteristics; and through stress testing of an independent validation set, a high-robustness global model is selected. For example, in a certain federated learning, the accuracy rates of the local models of 3 parking lots are 92%, 85%, and 90% respectively. The central server generates intermediate parameters by weighted average (weights 0.4, 0.2, 0.4), and the validation set accuracy rate is improved to 93%. Finally, the accuracy rate of parking lot C's spiral ramp area parking space allocation is improved by 20%, and the vehicle shunting efficiency during peak hours is improved by 35%.

[0103] Here is a specific example:

[0104] Suppose there are 3 irregular parking lots (A, B, C) in a city, and the data volume and model performance of each parking lot are as follows: parking lot A: data volume 1000, model accuracy 95%. Parking lot B: data volume 800, model accuracy 90%. Parking lot C: data volume 1200, model accuracy 92%. The model parameters of the 3 local nodes are fused by using the weighted average strategy. Based on the data volume and model accuracy, the weights are allocated, for example, the weight of parking lot A is 0.35, the weight of parking lot B is 0.25, and the weight of parking lot C is 0.40. The intermediate feature extraction model parameters are validated, and if the accuracy rate on the validation set reaches 93%, they are determined as the global feature extraction model parameters.

[0105] By performing steps a1~a3, the embodiment of the present application realizes efficient fusion of local model parameters of multiple heterogeneous parking lots through a weighted average strategy. Combined with data volume, data quality and model performance indicators, high-precision global feature extraction model parameters are generated. After verification and optimization, the parameters can effectively support dynamic parking space allocation and path planning, improve parking efficiency and resource utilization, especially in peak hours or sudden situations, and provide drivers with a more intelligent and convenient parking experience. In one possible embodiment, S14, in combination with the time variation characteristics of the channel flow of the heterogeneous parking lot in the space-time topology graph of all heterogeneous parking lots and the target feature extraction model parameters, a dynamic parking space allocation strategy is generated, including:

[0106] Step 141, using a pre-trained machine learning algorithm, extracts the time variation characteristics of the channel flow of the corresponding heterogeneous parking lot from the space-time topology graph of each heterogeneous parking lot, including the vehicle flow data of each channel in different time periods.

[0107] Wherein, the time variation characteristics of the channel flow of the heterogeneous parking lot include the vehicle flow data of each channel in different time periods, which are used to describe the flow pattern and dynamic change trend of the parking lot.

[0108] Step 142, input the time variation characteristics of the channel flow of all heterogeneous parking lots into the target feature extraction model using the target feature extraction model parameters to obtain the integrated space-time feature representation, which is used to describe the flow pattern and dynamic change trend of each heterogeneous parking lot.

[0109] Wherein, the integrated space-time feature representation is the integration of the time variation characteristics of the channel flow of each heterogeneous parking lot, forming a unified space-time feature representation, which is used to describe the flow pattern and dynamic change trend of each parking lot.

[0110] Step 143, obtain real-time channel flow data at the current time, combine the integrated space-time feature representation, and predict the idle parking space distribution of each heterogeneous parking lot in the future time period to obtain an idle parking space distribution map, which contains the location information and estimated available time of each idle parking space.

[0111] Wherein, the real-time channel flow data is the vehicle flow data of each channel at the current time. The idle parking space distribution map contains the location information and estimated available time of each idle parking space. The idle parking space distribution map is generated according to the prediction result, and contains the location information and estimated available time of each idle parking space.

[0112] Step 144, based on the idle parking space distribution map, combine the helical ramp geometric parameters to calculate the accessibility score value of each idle parking space.

[0113] wherein the accessibility score value is calculated based on the location of the free parking space and the spiral ramp geometry parameters, for evaluating the difficulty of the vehicle to reach the free parking space. The parameter acquisition is to acquire the spiral ramp geometry parameters from the parking lot design drawings or management system. The accessibility calculation is to calculate the accessibility score value of each free parking space by using an algorithm in combination with the free parking space distribution map and the spiral ramp geometry parameters.

[0114] Step 145, according to the accessibility score value of each free parking space, load balancing analysis is performed on all special-shaped parking lots to obtain a candidate parking space list of the vehicle to be parked.

[0115] wherein the candidate parking space list is selected according to the accessibility score value. The accessibility score value is sorted according to the accessibility score value of all free parking spaces to obtain a free parking space list from high to low. The load balancing is further filtered according to the current load situation of each parking lot (such as the number of parked vehicles, channel congestion, etc.), to ensure load balancing.

[0116] Step 146, the driving trajectory information of the vehicle to be parked is acquired, and a dynamic parking space allocation strategy is generated according to the driving trajectory information and the candidate parking space list of the vehicle to be parked.

[0117] wherein the dynamic parking space allocation strategy is generated according to the optimal parking path, including the specific parking space location and the predicted arrival time.

[0118] For example, input the channel flow data every 15 minutes in the past 7 days, use the long short-term memory network to capture periodic patterns, such as 20% higher traffic during the morning peak on weekdays than on weekends, and output time encoding features, such as peak period label vectors. The long short-term memory network output of each parking lot is spliced with geometric parameters such as lane curvature and slope, and input into the graph convolution network to generate a 128-dimensional feature vector representing global flow patterns and spatial constraints. The real-time channel flow (such as 3 vehicles per minute at the current entrance) and the spatio-temporal feature vector are input into the time convolution network to output the probability and duration of each parking space being free, generating a distribution map in the form of a heat map. From the path from the entrance to the target parking space, the reciprocal of the curvature and the slope ratio of each ramp are accumulated, the smaller the curvature (the sharper the turn) or the larger the slope, the lower the score. According to the formula: The load rate of parking lot A is 90%, and the load rate of parking lot C is 70%, so the high-score parking space of parking lot C is preferentially allocated. According to the real-time position of the vehicle, the candidate parking space list, and the real-time topological map (such as sudden congestion on channel C2), the model selects the optimal path with the shortest time and a score ≥ 3.5, and pushes the instruction through the navigation system: “turn right into ramp P1, and drive straight for 50 meters to reach parking space C-08”.

[0119] The following is a specific example: In a special-shaped parking lot of a large commercial complex, the system needs to predict the distribution of idle parking spaces in each parking lot in real time, and allocate the best parking space for vehicles entering the parking lot. Through sensors installed in each channel, real-time vehicle flow data is collected and cleaned and preprocessed. The pre-trained model is used to extract the time-varying characteristics of channel flow from the spatiotemporal topology graph and integrate them into a unified spatiotemporal feature representation. Combined with real-time channel flow data and integrated spatiotemporal feature representation, a deep learning model is used to predict the idle parking space distribution in the future time period. According to the idle parking space distribution map and the spiral ramp geometric parameters, the accessibility score value of each idle parking space is calculated. Combined with the current load of each parking lot, a candidate parking space list for the vehicle to be parked is selected. By obtaining the driving trajectory information of the vehicle to be parked, a specific parking space allocation strategy is generated to guide the vehicle to park in the best idle parking space.

[0120] By performing steps 141-146, the system of the present application embodiment can predict the distribution of idle parking spaces in each special-shaped parking lot in real time, and dynamically allocate the best parking space according to the driving trajectory and parking demand of the vehicle. This not only improves the utilization rate of the parking lot and reduces the time for vehicles to find parking spaces, but also effectively avoids congestion in the parking lot and improves the user experience.

[0121] In one possible embodiment, step 133, based on the parking space occupancy state data and the spiral ramp geometric parameters, trains a local feature extraction model on the local node to obtain local feature extraction model parameters related to the lane curvature, including:

[0122] Step b1, initialize the local feature extraction model.

[0123] Where initialization is to assign initial values to the weight and bias parameters of the model, usually using random initialization or pre-training model parameters for initialization to ensure stability at the beginning of training. The initialized model provides a basis for subsequent training.

[0124] In the present application embodiment, the PyTorch framework can be used to build the local feature extraction model, and the model structure is a 3-layer fully connected network (input layer 256 dimensions, hidden layer 128 dimensions, output layer 64 dimensions), the activation function uses ReLU, and the output layer connects the Sigmoid function to generate the parking space occupancy probability. The hidden layer is initialized with Xavier normal distribution to avoid gradient disappearance. The bias is initialized to zero. If there is a historical model, download the pre-training parameters (such as the feature extraction layer of ResNet-18) from the central server, and only fine-tune the top layer by freezing the bottom layer parameters.

[0125] For example, after the local model of parking lot A is initialized, the input layer receives 256-dimensional features (128-dimensional time series of parking space state + 128-dimensional ramp geometry parameters), the hidden layer weight matrix has a dimension of 128*256, and the bias vector has a dimension of 128*1.

[0126] Step b2, based on the parking space occupancy state data and the spiral ramp geometry parameters, an input data set of the local feature extraction model is constructed.

[0127] In the formula, the input data set is the feature engineering processing of the parking space occupancy state data and the spiral ramp geometry parameters, and is used to construct the input data required for model training. The construction of the input data set is to extract the parking space occupancy state data from the local node, convert it into time series features (such as parking occupancy changes in the past 1 hour), extract the spiral ramp geometry parameters, and perform standardization processing (such as normalization to the range of [0, 1]). The parking space occupancy state data and the spiral ramp geometry parameters are spliced to form the input feature vector of the model. For example, the parking space state in the past 1 hour (1 piece per second) is read from the local database, the occupancy rate change is counted in a 5-minute window, and a 128-dimensional time series vector is generated (for example, the occupancy rate in the window is reduced from 80% to 70%, which is coded as 0.8, 0.78,..., 0.7). The curvature , slope (5°~15°), length (20~100m) is extracted, and is normalized to [0, 1] through MinMax, and is spliced into a 128-dimensional vector, such as curvature 0.3 converted to 0.3 / 0.5=0.6, slope 12° converted to 12 / 15=0.8. The parking space state vector and the ramp parameter vector are spliced by column to form a 256-dimensional input feature. Then, the input data set is divided into a training set, a validation set and a test set, with a proportion of 70%:15%:15%. For another example, among the 1000 data of parking lot A, 700 are used for training, 150 are used for validation, and 150 are used for testing. Single data sample: input feature = [0.8, 0.78,..., 0.7 (time series), 0.6, 0.8,... (ramp parameters)], label = 1 (parking occupancy).

[0128] Step b3, input the input data set into the initialized local feature extraction model, and obtain the local feature extraction model parameters related to the lane curvature through training.

[0129] Wherein, the training is to adjust the model parameters through optimization algorithm (such as gradient descent method) to minimize the loss function (such as mean square error or cross entropy loss). The specific process of training is to input the initialized local feature extraction model into the input data set, and the model calculates the predicted value through forward propagation. The loss between the predicted value and the true value (such as using mean square error loss function) is calculated, the gradient is calculated through back propagation algorithm, and the model parameters are updated using the optimizer (such as Adam optimizer). Repeat the above process until the performance of the model on the validation set reaches the preset threshold or the training round ends. Finally, the local feature extraction model parameters related to the lane curvature are obtained, which will be uploaded to the central server for aggregation.

[0130] For example, after 50 rounds of local model training of parking lot A, the validation set accuracy reaches 95%, and the loss function converges to 0.12. The extracted model parameters include: part of the value of the first layer weight W1 is 0.32, -0.15, 0.28,..., which shows the negative effect of lane curvature parameters (128-dimensional input features) on hidden layer activation value. The output layer bias b3=0.18 is used to adjust the predicted baseline.

[0131] The following is a specific example: suppose the parking space occupancy state data of parking lot A includes the following content: parking space identifier (Identifier, ID) is 001, state is idle, timestamp is: 2023-10-01, 12:00:00, parking space ID is 002, state is occupied, timestamp is 2023-10-01, 12:00:00, and the spiral ramp geometric parameters of parking lot A include the following content: ramp ID is 001, lane curvature is 0.5, slope is 10%, and length is 50 meters. Initialize the local feature extraction model (such as graph neural network), use the initialization method to construct the input data set, concatenate the parking space occupancy state data and the spiral ramp geometric parameters to form the feature vector. Input the input data set into the model, and get the local feature extraction model parameters related to the lane curvature through training.

[0132] By performing steps b1~b3, the embodiment of the application generates local feature extraction model parameters related to lane curvature by initializing local feature extraction model, constructing input data set and training. These parameters can effectively capture the dynamic features and spatial constraints of parking lot data, providing high-quality basis for subsequent federated learning and global model optimization, and improving the intelligent level and efficiency of parking management.

[0133] In one possible embodiment, S11, the parking space occupancy state data of a plurality of special-shaped parking lots is obtained, including:

[0134] Step 111, obtain the pressure distribution data of each parking space in the corresponding irregular parking lot collected by each flexible pressure film sensor array in real time. One irregular parking lot corresponds to one flexible pressure film sensor array, and a flexible pressure film sensor is arranged on each parking space in the irregular parking lot, and a flexible pressure film sensor is arranged on the curved lane in the irregular parking lot. The flexible pressure film sensor array is arranged in a conformal manner according to the contour of the parking space and the curved lane of the irregular parking lot.

[0135] Among them, the flexible pressure film sensor array is an array composed of multiple flexible pressure film sensors, used for detecting pressure distribution. Each sensor can be bent and conformally arranged on the parking space and curved lane of the irregular parking lot. The pressure distribution data is the pressure data collected by the sensor in real time on each parking space and lane, reflecting the position and weight distribution of the vehicle. The sensor arrangement is to install flexible pressure film sensors on each parking space and curved lane in the irregular parking lot, ensuring that the sensor array is arranged in a conformal manner according to the contour of the parking space and lane. Data acquisition is to collect pressure data in real time by the sensor, and transmit it to the central data processing unit through wireless communication.

[0136] Step 112, convert the pressure distribution data of the parking space into parking space occupancy state data.

[0137] Among them, the parking space occupancy state data represents the state information of whether each parking space is currently occupied, usually represented in binary form (0 represents idle, 1 represents occupied). Data conversion is to convert the judgment result into parking space occupancy state data and store it in the database for subsequent use.

[0138] The following is a specific example:

[0139] In an irregular parking lot of a large commercial complex, the system needs to monitor the occupancy state of each parking space in real time in order to manage and navigate the parking space. Flexible pressure film sensors are installed on each parking space and curved lane in the parking lot, ensuring that the sensor array is arranged in a conformal manner according to the contour of the parking space and lane. The sensor collects pressure data in real time and transmits it to the central data processing unit through wireless communication. The central unit cleans and preprocesses the received data to remove noise and outliers. The central processing unit analyzes the pressure distribution data of each parking space, sets a pressure threshold, and judges whether each parking space is occupied. The judgment result is converted into parking space occupancy state data and stored in the database for use by the parking space management and navigation system.

[0140] By executing steps 111~112, the system of the present application can obtain the pressure distribution data of each parking space in the irregular parking lot in real time and convert it into parking space occupancy state data. This not only improves the efficiency and accuracy of parking management, but also provides real-time parking navigation services for drivers, improving user experience.

[0141] In one possible embodiment, step 112 of converting the pressure distribution data of the parking space into parking space occupancy state data comprises:

[0142] Step c1, according to the preset analysis rule, the pressure distribution data is analyzed to obtain the vehicle parking position and weight distribution information.

[0143] The preset analysis rule can be a set of pre-defined data processing rules and algorithm logic for extracting vehicle parking position and weight distribution information from original pressure data. The rule set can include data filtering rules, outlier rejection rules, coordinate conversion rules, pressure peak detection rules, weight calculation rules, etc., wherein the data filtering rules can filter out invalid signals such as environmental vibration and transient pressure caused by pedestrians stepping on; the outlier rejection rules can refer to identifying abnormal data through standard deviation analysis or box plot method; the coordinate conversion rule can refer to mapping the physical coordinates of the sensor array to the global coordinate system of the parking lot; the pressure peak detection rule can refer to identifying the local maximum value point of the pressure through a sliding window (such as a 5x5 grid); the weight calculation rule can refer to calculating the total weight of the vehicle based on the pressure integral and sensor calibration parameters. The vehicle parking position is the vehicle center point coordinate (such as X / Y axis coordinate) located by the sensor array coordinate. The weight distribution information is the pressure data distribution map of the vehicle on the sensor, reflecting the wheel position and weight concentration area. The weight estimation is the sum of the pressure values in the pressure distribution area through the integral algorithm, combined with the sensor sensitivity parameter (such as 0.5V / N) to convert into the total weight of the vehicle. The position analysis is based on the pressure peak detection algorithm (such as convolutional neural network), which identifies the local maximum value point in the pressure distribution map, and calculates the vehicle center position combined with the sensor grid coordinates.

[0144] Step c2, generating a parking space occupancy identifier based on the vehicle parking position and weight distribution information.

[0145] The parking space occupancy identifier is a data structure containing binary state (0 / 1), timestamp, and confidence score, used to represent whether the parking space is occupied. The weight state determination is a dynamic threshold method (such as historical data sliding window mean ± 3σ), which marks the occupancy when the weight exceeds the threshold (such as > 50 kg) and the duration > 5 seconds. The generation process of the parking space occupancy identifier is as follows: according to the vehicle parking position, the corresponding parking space ID is matched. According to the weight distribution information, the vehicle type (such as small car, large car) is judged, and the parking space occupancy identifier is generated combined with the preset rule. For example, if the weight exceeds the threshold, it is marked as "large car occupied". The parking space occupancy identifier is stored in the parking space state database for subsequent use.

[0146] Step c3, combining the parking space occupancy identifier with the vehicle weight distribution information to generate parking space occupancy state data.

[0147] Wherein, the parking space occupancy state data is through the structured data packet, including occupancy state, weight distribution heat map, time series change curve and other dimensions. The parking space occupancy identifier and vehicle weight distribution information are combined by time and space alignment to bind the occupancy identifier and weight data through timestamp and parking space number, ensuring the matching of the two types of data at the same time and the same parking space. For example, when the occupancy identifier of parking space B-07 is 1, the corresponding weight data needs to match the sensor collection range structure of the parking space to generate a data packet containing multiple dimensions, such as "Parking Space Number": "B-07", "Occupancy State": 1, "Confidence": 98%, "Total Weight": 2100kg, "Weight Distribution Heat Map": "Base64 encoded matrix data", "Timestamp": "2025-03-24T10:15:30". Then the reliability of the occupancy identifier is verified through dynamic verification by weight data. For example, if the identifier is "occupied" but the weight is below the threshold (such as <50kg), an abnormal detection mechanism is triggered. The deep combination of parking space occupancy identifier and weight distribution information can avoid misjudgment (such as temporary parking of shopping cart), and the accuracy is improved to more than 99%, and the weight data is combined to optimize the vehicle transfer strategy (such as the logic of transferring the vehicle to the conveyor belt by the skateboard device in webpage 1); the complete weight heat map is stored to assist in accident responsibility determination (such as restoring the vehicle position when a scratch occurs).

[0148] The following is a specific example: A certain airport multi-level parking garage collects real-time pressure data of 2000 irregular parking spaces through a conformal flexible pressure sensor array (resolution 5cm x 5cm), with a sampling frequency of 10Hz. The system uses a pre-processing module to filter temperature interference (adapted to extreme environments of -20℃~60℃), and accurately identifies the pressure distribution pattern of four-wheel vehicles through the model, outputting lightweight structured data in data exchange format (such as vehicle coordinates (x=3.2m, y=5.7m) and weight 1.8t). The dynamic threshold algorithm combines with the pressure distribution characteristics to determine that parking space A12 in area B is in an occupied state (confidence 98%), and generates a scalable vector graphic with timestamp to real-time overlay the electronic map, supporting visual parking space state tracking. The data integration module constructs a comprehensive data packet containing 256 color scale heat map, occupancy duration and vehicle weight distribution, and pushes it to the central dispatching system through the protocol, with communication delay controlled within 200ms, meeting the high real-time requirement. Through multi-dimensional data fusion and low-latency transmission, the system realizes accurate monitoring and efficient scheduling of irregular parking spaces.

[0149] By performing steps c1-c3, the application embodiment realizes the accurate extraction of vehicle parking position and weight distribution information through the analysis and processing of pressure distribution data, and generates real-time parking space occupancy state data in combination with the parking space occupancy identifier. The data can effectively support dynamic parking space allocation and path planning, improve the intelligent level and efficiency of parking management, especially in peak periods or sudden situations, and provide drivers with a more intelligent and convenient parking experience.

[0150] Figure 2 A structural schematic diagram of a parking lot idle parking space prediction system provided by the application embodiment is shown in Figure 2 The system comprises:

[0151] The acquisition module 21 is configured to acquire parking space occupancy state data and topology structure data of the plurality of special-shaped parking lots.

[0152] The generation module 22 is configured to generate, for each special-shaped parking lot, a special-shaped parking lot space-time topology graph by fusing a slope curvature parameter extracted from the special-shaped parking lot topology structure data and historical vehicle turning radius data.

[0153] The extraction module 23 is configured to aggregate the parking space occupancy state data and the special-shaped parking lot space-time topology graph of all special-shaped parking lots through a federal transfer learning framework to obtain target feature extraction model parameters.

[0154] The generation module 24 is configured to generate a dynamic parking space allocation strategy in combination with a time variation feature of a channel flow of the special-shaped parking lot in the special-shaped parking lot space-time topology graph and the target feature extraction model parameters.

[0155] The mapping module 25 is configured to map the dynamic parking space allocation strategy to a real-time navigation path of the special-shaped parking lot space-time topology graph, so that a vehicle to be parked reaches an idle parking space through the real-time navigation path.

[0156] Figure 2 The parking lot idle parking space prediction method system can perform Figure 1 The implementation principle and technical effects of the parking lot idle parking space prediction method described in the embodiment shown in are not described again. For the specific manner in which each module, unit of the parking lot idle parking space prediction system described in the above embodiment performs operations, it has been described in detail in the embodiment related to the method, and will not be described in detail here.

[0157] In one possible design, Figure 2 The parking lot idle parking space prediction system of the embodiment shown in can be implemented as a computing device, as shown in Figure 3 The computing device can include a storage component 31 and a processing component 32.

[0158] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32.

[0159] The processing component 32 is configured to: obtain parking space occupancy state data and topology structure data of a plurality of special-shaped parking lots; for each special-shaped parking lot, generate a special-shaped parking lot space-time topology graph by fusing a ramp curvature parameter extracted from the special-shaped parking lot topology structure data and historical vehicle turning radius data; aggregate the parking space occupancy state data and the special-shaped parking lot space-time topology graph of all special-shaped parking lots through a federal transfer learning framework to obtain target feature extraction model parameters; generate a dynamic parking space allocation strategy in combination with time variation characteristics of channel flow of the special-shaped parking lot in the special-shaped parking lot space-time topology graph and the target feature extraction model parameters; and map the dynamic parking space allocation strategy to a real-time navigation path in the special-shaped parking lot space-time topology graph, so that a vehicle to be parked reaches an idle parking space through the real-time navigation path.

[0160] The processing component 32 can include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic elements for executing the above method.

[0161] The storage component 31 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or nonvolatile storage devices or a combination thereof, such as a random access memory (RAM), a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic storage, a flash memory, a magnetic disk, or a optical disk.

[0162] Of course, the computing device can also include other components, such as an input / output interface, a display component, a communication component, etc.

[0163] The input / output interface provides an interface between the processing component and peripheral interface modules, which can be output devices, input devices, etc.

[0164] The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, etc.

[0165] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform, and the processing component, the storage component, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0166] The embodiments of the present application also provide a computer storage medium storing a computer program, and the computer program can implement the above-mentioned Figure 1 The embodiment shown provides a parking lot idle parking space prediction method.

[0167] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0168] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0169] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0170] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting vacant parking spaces in a parking lot, characterized in that: include: Obtain parking space occupancy status data and topological structure data of multiple special-shaped parking lots; For each special-shaped parking lot, a spatiotemporal topological map of the special-shaped parking lot is generated by fusing ramp curvature parameters extracted from the special-shaped parking lot topological structure data with historical vehicle turning radius data; Aggregating the parking space occupancy status data of all the irregular parking lots and the spatiotemporal topology map of the irregular parking lots through a federated transfer learning framework to obtain target feature extraction model parameters; Combining the time-varying characteristics of the channel flow of all the irregular parking lots in the spatiotemporal topology of the irregular parking lots with the target feature extraction model parameters, a dynamic parking space allocation strategy is generated; Mapping the dynamic parking space allocation strategy to the real-time navigation path of the spatiotemporal topology of the special-shaped parking lot so that the parked vehicles can reach the vacant parking spaces through the real-time navigation path; The method of aggregating the parking space occupancy status data of all the irregular parking lots and the spatiotemporal topology map of the irregular parking lots through the federated transfer learning framework to obtain target feature extraction model parameters includes: Inputting the parking space occupancy status data and the spatiotemporal topology map of each of the irregular parking lots into a corresponding local node within the federated transfer learning framework, so as to store the parking space occupancy status data and the spatiotemporal topology map of the corresponding irregular parking lot in each of the local nodes, with one local node corresponding to one irregular parking lot; In each of the local nodes, extracting geometric parameters of the spiral ramp in the spatiotemporal topological graph of the special-shaped parking lot, the geometric parameters of the spiral ramp including lane curvature; Based on the parking space occupancy status data and the spiral ramp geometric parameters, training a local feature extraction model on a local node to obtain local feature extraction model parameters related to lane curvature; In a central server within the federated transfer learning framework, aggregating all the local feature extraction model parameters to generate global feature extraction model parameters; The global feature extraction model parameters are subjected to convergence processing to obtain target feature extraction model parameters.

2. The method according to claim 1, characterized in that The aggregating all the local feature extraction model parameters to generate global feature extraction model parameters includes: A weighted average strategy is used to fuse all the local feature extraction model parameters to generate intermediate feature extraction model parameters; wherein the weighted average strategy includes a weight distribution scheme determined based on the data volume, data quality or model performance index of each local node; The intermediate feature extraction model parameters are verified, and the verified intermediate feature extraction model is determined as the global feature extraction model parameters.

3. The method according to claim 1, characterized in that The method combines the time variation characteristics of the channel flow of all the irregular parking lots in the spatiotemporal topology of the irregular parking lots with the target feature extraction model parameters to generate a dynamic parking space allocation strategy, including: Using a pre-trained machine learning algorithm, the time-varying characteristics of the channel traffic of each special-shaped parking lot are extracted from the spatiotemporal topology map of each special-shaped parking lot. The time-varying characteristics include the vehicle flow data of each channel in the corresponding special-shaped parking lot at different time periods. Input the time-varying characteristics of the channel flow of all irregular-shaped parking lots into the target feature extraction model using the target feature extraction model parameters to obtain an integrated spatiotemporal feature representation, which is used to describe the flow pattern and dynamic change trend of each irregular-shaped parking lot; Acquire real-time channel traffic data at the current moment, combine it with the integrated spatiotemporal feature representation, predict the distribution of vacant parking spaces in each special-shaped parking lot in the future time period, and obtain a vacant parking space distribution map, wherein the vacant parking space distribution map includes the location information and estimated availability time of each vacant parking space; Calculating an accessibility score for each vacant parking space based on the vacant parking space distribution map and the spiral ramp geometric parameters; Performing load balancing analysis on all the special-shaped parking lots according to the accessibility score of each vacant parking space to obtain a list of candidate parking spaces for vehicles to be parked; Acquire the driving trajectory information of the vehicle to be parked, and generate a dynamic parking space allocation strategy based on the driving trajectory information and the candidate parking space list of the vehicle to be parked.

4. The method according to claim 1, wherein The method of training a local feature extraction model on a local node based on the parking space occupancy status data and the spiral ramp geometric parameters to obtain local feature extraction model parameters related to lane curvature includes: Initialize the local feature extraction model; constructing an input data set for a local feature extraction model based on the parking space occupancy status data and the spiral ramp geometric parameters; The input data set is input into the initialized local feature extraction model, and local feature extraction model parameters related to lane curvature are obtained through training.

5. The method according to claim 1, wherein The method of obtaining parking space occupancy status data of a plurality of special-shaped parking lots includes: Obtaining real-time pressure distribution data for each parking space in the corresponding irregularly shaped parking lot, collected by each flexible pressure film sensor array. Each irregularly shaped parking lot corresponds to one flexible pressure film sensor array. Each parking space in the irregularly shaped parking lot is equipped with a flexible pressure film sensor, and each curved lane in the irregularly shaped parking lot is equipped with a flexible pressure film sensor. The flexible pressure film sensor arrays are arranged in a conforming manner according to the contours of the parking spaces and curved lanes in the irregularly shaped parking lot. The pressure distribution data of the parking space is converted into parking space occupancy status data.

6. The method according to claim 5, characterized in that The converting the pressure distribution data of the parking space into parking space occupancy status data includes: Analyzing the pressure distribution data according to preset analysis rules to obtain vehicle parking position and weight distribution information; generating a parking space occupancy indicator according to the vehicle parking position and the weight distribution information; The parking space occupancy identification is combined with the vehicle weight distribution information to generate parking space occupancy status data.

7. A system for predicting vacant parking spaces in a parking lot, characterized in that: include: An acquisition module is used to obtain parking space occupancy status data and topological structure data of multiple special-shaped parking lots; A generation module is used to generate a spatiotemporal topological map of each special-shaped parking lot by fusing ramp curvature parameters extracted from the special-shaped parking lot topological structure data with historical vehicle turning radius data; an extraction module, configured to aggregate the parking space occupancy status data of all the irregular parking lots and the spatiotemporal topology map of the irregular parking lots through a federated transfer learning framework to obtain target feature extraction model parameters; A generation module, configured to combine the time-varying characteristics of the channel flow of all the irregular parking lots in the spatiotemporal topology graph of the irregular parking lots with the target feature extraction model parameters to generate a dynamic parking space allocation strategy; A mapping module, configured to map the dynamic parking space allocation strategy to a real-time navigation path of the spatiotemporal topology of the special-shaped parking lot, so that the parked vehicle can reach an available parking space via the real-time navigation path; The method of aggregating the parking space occupancy status data of all the irregular parking lots and the spatiotemporal topology map of the irregular parking lots through the federated transfer learning framework to obtain target feature extraction model parameters includes: Inputting the parking space occupancy status data and the spatiotemporal topology map of each of the irregular parking lots into a corresponding local node within the federated transfer learning framework, so as to store the parking space occupancy status data and the spatiotemporal topology map of the corresponding irregular parking lot in each of the local nodes, with one local node corresponding to one irregular parking lot; In each of the local nodes, extracting geometric parameters of the spiral ramp in the spatiotemporal topological graph of the special-shaped parking lot, the geometric parameters of the spiral ramp including lane curvature; Based on the parking space occupancy status data and the spiral ramp geometric parameters, training a local feature extraction model on a local node to obtain local feature extraction model parameters related to lane curvature; In a central server within the federated transfer learning framework, aggregating all the local feature extraction model parameters to generate global feature extraction model parameters; The global feature extraction model parameters are subjected to convergence processing to obtain target feature extraction model parameters.

8. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for predicting vacant parking spaces in a parking lot as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the method for predicting vacant parking spaces in a parking lot according to any one of claims 1 to 6 is implemented.

Images