Road section illegal parking space-time prediction system based on deep learning

Abstract

The invention discloses a road section illegal parking space-time prediction system based on deep learning, and relates to the technical field of intelligent traffic management, and the system comprises a multi-source data fusion collection module, a space-time self-attention feature extraction module, a dynamic granularity parameter generation module, a space-time hybrid model prediction module and an illegal parking early warning visualization module. By collecting traffic camera video streams, vehicle-mounted sensor data, mobile phone signaling data and meteorological data in real time, comprehensive perception of the traffic environment is achieved, the multi-source heterogeneous data are mapped to a unified GIS grid after being subjected to space-time coordinate calibration, a multi-dimensional illegal parking influence factor database is constructed, and on the basis, the real-time monitoring of the illegal parking influence factor database is achieved. The space-time self-attention feature extraction module performs efficient space-time feature extraction on multi-source data by using a space-time diagram self-attention model, generates feature vectors with space-time relevance, and solves the problems of single data and insufficient feature extraction in a traditional method.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic management technology, specifically a spatiotemporal prediction system for illegal parking on road sections based on deep learning. Background Technology

[0002] With the acceleration of urbanization, urban traffic problems are becoming increasingly prominent, among which illegal parking is particularly serious. This not only affects road traffic efficiency but also increases the risk of traffic accidents. In order to effectively manage urban traffic and improve road utilization efficiency, accurate prediction of illegal parking has become an urgent problem to be solved. Traditional traffic management methods often rely on manual patrols and fixed camera monitoring. These methods have limitations such as limited monitoring range, slow response speed, and insufficient data utilization. With the rapid development of big data, artificial intelligence, and Internet of Things technologies, using multi-source data for spatiotemporal prediction of illegal parking has become a new research direction. By collecting multi-source heterogeneous data such as traffic camera video streams, vehicle sensor data, mobile phone signaling data, and meteorological data in real time, a more comprehensive understanding of the traffic environment can be achieved, providing rich data support for illegal parking prediction.

[0003] Traditional roadside illegal parking prediction technologies mainly rely on a single data source or simple statistical analysis methods, which have many shortcomings. First, a single data source cannot fully reflect the complexity of the traffic environment, leading to inaccurate prediction results. For example, relying solely on traffic camera video streams cannot obtain information on vehicle driving trajectories and driver behavior, while vehicle-mounted sensor data cannot reflect the surrounding traffic environment and weather conditions. Second, traditional methods often use simple linear models or rule engines in feature extraction and model building, which cannot effectively capture spatiotemporal correlations and nonlinear features, resulting in limited prediction accuracy. In addition, traditional methods lack dynamic adjustment mechanisms and cannot make personalized predictions based on the actual conditions of different road sections, making it difficult to meet the actual needs of urban traffic management.

[0004] Therefore, developing a deep learning-based spatiotemporal prediction system for illegal parking on road sections can provide traffic management departments with effective suggestions for patrol route optimization and facility planning, thereby achieving intelligent and refined management of urban traffic. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a spatiotemporal prediction system for illegal parking on road sections based on deep learning. By fusing and collecting multi-source data and combining spatiotemporal self-attention feature extraction and dynamic granular parameter generation, the system achieves accurate prediction of illegal parking behavior on road sections. The system uses a spatiotemporal graph temporal hybrid model to fuse multi-granular features, outputs the probability of illegal parking for each grid in future time periods, and intuitively displays the prediction results through a visualization module.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a deep learning-based spatiotemporal prediction system for illegal parking on road sections, which includes: a multi-source data fusion acquisition module, a spatiotemporal self-attention feature extraction module, a dynamic granular parameter generation module, a spatiotemporal hybrid model prediction module, and an illegal parking warning visualization module.

[0007] The multi-source data fusion acquisition module collects traffic camera video streams, vehicle sensor data, mobile phone signaling data and meteorological data in real time, desensitizes all external data simultaneously, and maps heterogeneous data to a unified GIS grid with the help of spatiotemporal coordinate calibration to build a multi-dimensional database of illegal parking influencing factors.

[0008] The spatiotemporal self-attention feature extraction module: inputs the preprocessed multi-source data into the spatiotemporal graph self-attention model, generates a joint attention matrix by calculating the spatial correlation matrix and the time dependency matrix, extracts spatiotemporal features from the multi-source data, and obtains spatiotemporal feature vectors;

[0009] The dynamic granularity parameter generation module analyzes the extracted spatiotemporal feature vectors through a fuzzy granularity matching mechanism, calculates the weights of each preset granularity mode based on the road segment characteristics, generates dynamic granularity parameters suitable for different road segments, and passes them to the spatiotemporal hybrid model prediction module.

[0010] The spatiotemporal hybrid model prediction module adopts a spatiotemporal graph temporal hybrid model, combines dynamic granularity parameters and spatiotemporal attention features, aggregates the hidden states of the grid, and fuses multi-granularity features through granularity gating units to output the probability of illegal parking for each grid in future time periods.

[0011] The illegal parking warning visualization module overlays the predicted illegal parking probability onto the electronic map to form an illegal parking hotspot warning system, and generates patrol route optimization and facility planning suggestions.

[0012] Furthermore, the multi-source heterogeneous data acquired in real time by the multi-source data fusion acquisition module includes:

[0013] Traffic camera video stream acquisition: Deploy high-definition cameras to acquire road video in real time at a frame rate of 25 frames per second, covering lanes, roadside parking lanes and sidewalk areas, and simultaneously record the camera's latitude and longitude coordinates and timestamps;

[0014] Vehicle sensor data acquisition: GPS trajectory, vehicle speed, and acceleration data are acquired through the vehicle OBD device and navigation terminal at a sampling frequency of 10Hz, and transmitted to the server in real time using a 4G / 5G communication module;

[0015] Mobile signaling data collection: In cooperation with telecommunications operators, anonymized base station handover records and signal strength data are obtained, with a sampling interval of 5 minutes per base station;

[0016] Meteorological data acquisition: Connects to the meteorological department's API to obtain real-time data on temperature, precipitation intensity, wind speed, and visibility within the region, with an update frequency of 10 minutes per update.

[0017] Furthermore, the specific steps in the multi-source data fusion acquisition module to map heterogeneous data to a unified GIS grid using spatiotemporal coordinate calibration are as follows:

[0018] (1) Preprocess traffic camera video and vehicle sensor multi-source data, and add spatiotemporal labels to each data;

[0019] (2) Divide the study area into dynamic grids, map the data points to the corresponding grids according to their coordinates and generate codes;

[0020] (3) Convert the timestamps of all data to a uniform UTC time, divide the time into 15-minute intervals and align them;

[0021] (4) Extract vehicle density features from multi-source data and standardize them, and construct multi-dimensional feature vectors according to the grid;

[0022] (5) Verify the spatiotemporal consistency of data, mark abnormal data and repair or smooth it through correlation reasoning;

[0023] (6) A distributed time-series database is used to store the data, and a composite index is designed and a multi-dimensional database of factors affecting illegal parking is constructed for storage.

[0024] Furthermore, the spatiotemporal graph self-attention model is constructed in the spatiotemporal self-attention feature extraction module, defining a space-time joint attention matrix. , Let L be the number of grid cells and L be the number of time slices. The joint attention matrix is: ,in It is a spatial correlation matrix, describing the spatial dependencies between grids. It is a time dependency matrix, describing the temporal correlation at different times. The learnable weight matrix is ​​optimized through cross-modal comparative learning, and the formula is: ,in The tags are for illegal parking, with 1 indicating a strong association and 0 indicating no association. It is a matrix index, corresponding to the association pairs of different grids or time slices.

[0025] Furthermore, in the spatiotemporal graph self-attention model It is a spatial incidence matrix. It is a time dependency matrix and a spatial incidence matrix. The calculation formula is: ,in The Euclidean distance between grid cells. The intensity of personnel flow between the two grids, The spatial decay factor, Representing different grids; time dependency matrix The calculation formula is: ,in For periodic parameters, For date type matching functions, It is the time in the time dependency matrix. and Dependency value, time Date type, Representing different points in time, Representative moment Date type.

[0026] Furthermore, in the dynamic granularity parameter generation module, the calculation of dynamic granularity parameters is based on the road segment feature vector. The granularity parameters are generated through a fuzzy rule engine, and the calculation formula is as follows: ,in, This is a time granularity parameter, referring to the size of the time interval for dividing data or conducting analysis along the time dimension. Different time granularities can reflect the characteristics of illegal parking on road sections at different time scales. It is a spatial granularity parameter, referring to the size of the spatial range for dividing data or conducting analysis in a spatial dimension. To preset the number of granularity modes, It is an index variable used to iterate through preset granularity patterns, with a value range of 1 to... , It is the mapping function of the fuzzy rule engine, which generates granular parameters based on the feature vector. It is the first The weights of the preset granularity patterns are calculated using the following formula: , It is the number of dimensions of the road segment feature vector. Features For pattern The membership function is given by the formula: , Dimension index of feature vector and Mean and variance parameters of fine-grained models, It is a characteristic value of the traffic level of a road segment. This represents the natural exponential function. No. The temporal and spatial granularity parameters of the model.

[0027] Furthermore, the construction formula for the spatiotemporal graph temporal series mixture model in the spatiotemporal mixture model prediction module is as follows: ,in, Is the grid g in the future The probability of illegal parking at any given time. It is the Sigmoid activation function, which maps the output to the [0, 1] probability space. GGU is a granular gating unit that fuses features of different granularities. It refers to the number of layers in a parallel spatiotemporal graph neural network. This represents the layer index of a parallel spatiotemporal graph neural network. These are the attention weights of the k-th layer of the network, generated by the spatiotemporal attention matrix. It is a k-th layer spatiotemporal graph neural network that outputs the hidden state of grid g at time t. These represent coarse-grained and fine-grained spatiotemporal features, respectively. It is a spatiotemporal attention matrix that encodes the spatiotemporal dependencies between grids. It is an element-wise multiplication that combines attention weights and feature representations.

[0028] Furthermore, the calculation formula for the granular gating unit (GGU) in the spatiotemporal graph temporal hybrid model is as follows: The gating mechanism is as follows: , It is a granularity fusion weight that dynamically adjusts the contribution of coarse and fine granular features. It is the Sigmoid activation function. These are learnable parameters. This represents the feature concatenation operation; the formula for calculating the spatiotemporal attention matrix is: ,in: It is a spatial incidence matrix, representing a grid. Spatial dependency with other grids, It is a time dependency matrix, representing The relationship between a moment and a historical period It is a learnable weight matrix; the calculation formula for the spatiotemporal graph neural network layer is: ,in: This represents the layer index of a parallel spatiotemporal graph neural network. It is a grid exist The characteristics of time, It is an aggregation function that uses attention-weighted aggregation of neighbor information. No. The weight matrix of a layered spatiotemporal graphical neural network is a learnable parameter used to perform a weighted transformation on the input features. This indicates that an aggregation operation is being performed. Represents a grid Spatial Neighbor Sets In the middle, in the historical time step The feature set, where It is a neighbor grid index. Indicates a historical time step. Grid The set of spatial neighbors, i.e., with the grid In a spatially adjacent set of grids, the size of the time window defined by R determines the time range for considering historical features. No. The bias term in a layered spatiotemporal graph neural network is a learnable parameter used to adjust the computation results.

[0029] Furthermore, the illegal parking warning visualization module uses a combination of heat map and spatiotemporal dynamic map for display. The heat map shows the probability of illegal parking in each grid, and the spatiotemporal dynamic map shows the evolution of illegal parking hotspots. The color mapping of the heat map uses quantile mapping to divide the probability of illegal parking into 5 levels, with each level corresponding to a different color range.

[0030] Compared with existing technologies, this deep learning-based spatiotemporal prediction system for illegal parking on road sections has the following advantages:

[0031] I. This invention achieves comprehensive perception of the traffic environment by real-time acquisition of traffic camera video streams, vehicle sensor data, mobile phone signaling data, and meteorological data. After spatiotemporal coordinate calibration, these multi-source heterogeneous data are mapped onto a unified GIS grid, constructing a multi-dimensional database of factors influencing illegal parking. Based on this, the spatiotemporal self-attention feature extraction module utilizes a spatiotemporal graph self-attention model to efficiently extract spatiotemporal features from the multi-source data, generating feature vectors with spatiotemporal correlation. This not only solves the problems of single data and insufficient feature extraction in traditional methods but also significantly improves the accuracy of illegal parking prediction, providing a more scientific and accurate decision-making basis for urban traffic management.

[0032] Second, this invention utilizes a fuzzy granularity matching mechanism to deeply analyze the extracted spatiotemporal feature vectors. Based on road segment characteristics, it calculates the weights of each preset granularity mode, generating dynamic granularity parameters suitable for different road segments. These parameters are then passed to the prediction module, enabling the prediction model to be personalized according to the actual conditions of different road segments. The spatiotemporal hybrid model prediction module employs a spatiotemporal graph temporal hybrid model, combining dynamic granularity parameters and spatiotemporal attention features to aggregate the hidden states of the grid. Through granularity gating units, it fuses multi-granularity features and outputs the probability of illegal parking for each grid in future time periods. This allows the invention to accurately predict illegal parking behavior for different road segments and time periods. The prediction results are then visually displayed through an illegal parking warning visualization module, providing traffic management departments with effective suggestions for patrol route optimization and facility planning, thereby achieving personalized and precise illegal parking warnings and management.

[0033] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0035] Figure 1 This is a flowchart of a deep learning-based spatiotemporal prediction system for illegal parking on road sections.

[0036] Figure 2 This is a framework diagram of a deep learning-based spatiotemporal prediction system for illegal parking on road sections. Detailed Implementation

[0037] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0038] Example 1:

[0039] Prediction and management of illegal parking in urban commercial center areas.

[0040] The core commercial center of the provincial capital city includes three large shopping malls, five Grade A office buildings, and two subway stations, with an average daily traffic flow of over 100,000 people. During weekday morning and evening rush hours and all day on weekends, the traffic flow is saturated. In recent years, due to the surge in demand for temporary parking, vehicles frequently illegally occupy lanes on both sides of main roads and at intersections of side roads, resulting in an average decrease in vehicle speed of 30% and a 15% year-on-year increase in the incidence of minor collisions. Management departments urgently need to accurately predict hotspots for illegal parking in order to achieve dynamic allocation of patrol resources and optimization of parking facilities.

[0041] Multi-source data fusion acquisition module: 12 4K high-definition cameras are deployed on the main road and surrounding side roads of the commercial center, with lens angles covering both lanes, roadside parking areas, and sidewalks. Video streams are acquired in real time at a frequency of 25 frames per second. Each frame is automatically bound to the camera's latitude and longitude coordinates and a millisecond-level timestamp, ensuring spatial positioning accuracy down to the meter level. In cooperation with Didi Chuxing, data from the onboard OBD devices of 200 taxis in the area is accessed. GPS trajectory, instantaneous speed, and acceleration information are acquired through a 10Hz sampling frequency. 5G network slicing technology enables millisecond-level data transmission, avoiding data loss due to network congestion during peak hours. Anonymized mobile signaling data from telecommunications operators is provided for the area, recording base station switching trajectories and signal strength every 5 minutes. All external data... All data has been anonymized, removing personal identifiers. This anonymized data is then used to analyze the spatiotemporal movement characteristics of shoppers and office workers. Meteorological data, including temperature, precipitation intensity, and visibility, is pushed in real-time via API, updated every 10 minutes. Data is specifically labeled to address the impact of severe weather such as heavy rain and snow on parking behavior. All data timestamps are converted to a unified UTC time, time slices are divided into 15-minute intervals and aligned, vehicle density features are extracted from multi-source data and standardized, and multi-dimensional feature vectors are constructed using a grid. Data spatiotemporal consistency is verified, and abnormal data is marked and repaired or smoothed through association reasoning. A distributed time-series database is used to store the data, and a composite index is designed to construct a multi-dimensional database of factors influencing illegal parking. Figure 1 As shown.

[0042] The spatiotemporal self-attention feature extraction module: inputs the preprocessed video frame vehicle detection results, taxi trajectory data, mobile phone signaling flow patterns, and meteorological indicators into the spatiotemporal graph self-attention model, and defines the space-time joint attention matrix. , Let L be the number of grid cells and L be the number of time slices. The joint attention matrix is: ,in It is a spatial correlation matrix, describing the spatial dependencies between grids. It is a time dependency matrix, describing the temporal correlation at different times. The learnable weight matrix is ​​optimized through cross-modal comparative learning, and the formula is: ,in The tags are for illegal parking, with 1 indicating a strong association and 0 indicating no association. It is a matrix index, corresponding to the association pairs of different grids or time slices. The model captures the spatial dependency relationship between different grids by calculating the spatial association matrix. For example, the correlation between illegal parking in the grid around the shopping mall and the grid at the entrance of the office building will be enhanced due to the flow of commuters. At the same time, the time dependency matrix is ​​used to analyze the temporal pattern of historical illegal parking data, such as the correlation between the probability of illegal parking during the lunch period of 12:00-13:00 on weekdays and the data of the same period of the previous week. Through these two types of matrices, a joint attention matrix is ​​generated, which automatically filters irrelevant features and focuses on high-impact factor data. Finally, the spatiotemporal feature vector containing spatial clustering features and time periodic patterns is output.

[0043] Dynamic granular parameter generation module: Based on the characteristics of "high-density pedestrian flow + tidal traffic flow" in commercial center road sections, it analyzes the spatiotemporal feature vector through a fuzzy granularity matching mechanism. The system presets five granularity modes: "fine-grained during morning peak hours" and "coarse-grained during off-peak hours." Based on the feature vectors of pedestrian density and traffic flow fluctuation rate for the current time period, it calculates the weight of each mode. For example, during the weekend shopping peak period from 2:00 PM to 4:00 PM, the system automatically increases the weight of the "fine-grained spatial division + short time window" mode, generating dynamic granular parameters adapted to this scenario. Let the road segment feature vector be... The granularity parameters are generated through a fuzzy rule engine, and the calculation formula is as follows: ,in, This is a time granularity parameter, referring to the size of the time interval for dividing data or conducting analysis along the time dimension. Different time granularities can reflect the characteristics of illegal parking on road sections at different time scales. It is a spatial granularity parameter, referring to the size of the spatial range for dividing data or conducting analysis in a spatial dimension. To preset the number of granularity modes, It is an index variable used to iterate through preset granularity patterns, with a value range of 1 to... , It is the mapping function of the fuzzy rule engine, which generates granular parameters based on the feature vector. It is the first The weights of the preset granularity patterns are calculated using the following formula: , It is the number of dimensions of the road segment feature vector. Features For pattern The membership function is given by the formula: , Dimension index of feature vector and Mean and variance parameters of fine-grained models, It is a characteristic value of the traffic level of a road segment. This represents the natural exponential function. No. The temporal and spatial granularity parameters of each mode ensure that the prediction accuracy matches real-time traffic changes.

[0044] Spatiotemporal mixture model prediction module: Employs a spatiotemporal graph temporal series mixture model, with the following formula: ,in, Is the grid g in the future The probability of illegal parking at any given time. It is the Sigmoid activation function, which maps the output to the [0, 1] probability space. GGU is a granular gating unit that fuses features of different granularities. It refers to the number of layers in a parallel spatiotemporal graph neural network. This represents the layer index of a parallel spatiotemporal graph neural network. These are the attention weights of the k-th layer of the network, generated by the spatiotemporal attention matrix. It is a k-th layer spatiotemporal graph neural network that outputs the hidden state of grid g at time t. These represent coarse-grained and fine-grained spatiotemporal features, respectively. It is a spatiotemporal attention matrix that encodes the spatiotemporal dependencies between grids. It uses element-wise multiplication, integrates attention weights and feature representations, and inputs dynamic granular parameters and spatiotemporal attention features into the model. The model integrates coarse-grained regional overall trend features and fine-grained local hotspot features through granular gating units. For example, it considers the parking demand trend of the entire commercial center area and the real-time traffic pressure within 50 meters of the entrance of an office building. It aggregates the hidden states of the grid through a multi-layer spatiotemporal graph neural network to capture complex patterns such as "the spatiotemporal correlation between illegal parking within 50 meters of the subway station exit and the saturation of parking spaces in the surrounding shopping mall". Finally, it outputs the predicted value of illegal parking probability for each grid in the next hour, and the probability accuracy can be refined to 15-minute time slices.

[0045] The illegal parking warning visualization module presents the prediction results on the electronic map platform in the form of a heat map overlaid with a spatiotemporal dynamic map. The heat map uses the quantile mapping method to divide the probability of illegal parking into five levels: "extremely low-low-medium-high-extremely high," corresponding to five color intervals: blue, green, yellow, orange, and red. For example, the red area indicates that the probability of illegal parking exceeds 70% in the next 30 minutes. The spatiotemporal dynamic map plays an animation of the migration of illegal parking hotspots at 15-minute intervals, intuitively showing the evolution process such as "the illegal parking hotspot at the entrance of the shopping mall parking lot during lunchtime spreading to the surrounding side roads." Based on the visualization results, the management department will focus the deployment of patrol police forces to the red warning area and plan to add temporary no-parking signs in the orange area.

[0046] In summary, in the prediction and management of illegal parking in urban commercial center areas, the system first collects multi-source data, including road video, taxi trajectories, mobile phone signaling, and meteorological data. Features are extracted using a spatiotemporal graph self-attention model, and dynamic granular parameters are generated by a fuzzy granularity matching mechanism. These parameters are then input into a spatiotemporal graph temporal hybrid model to predict the probability of illegal parking. Finally, the system visualizes the data through heat maps and spatiotemporal dynamic graphs, providing suggestions for patrol optimization and facility planning. This embodiment achieves accurate prediction and effective management of illegal parking in commercial centers, improving traffic management efficiency.

[0047] Example 2:

[0048] Prediction and management of illegal parking on roads surrounding industrial parks.

[0049] A national-level industrial park covers an area of ​​12 square kilometers and houses more than 200 companies, including 3 logistics and warehousing centers and 5 production plants. During the morning and afternoon shifts and peak freight hours, trucks frequently illegally occupy the road for loading and unloading, and private cars temporarily park on the main roads and entrances of the park. This causes an average delay of 15 minutes per logistics vehicle, resulting in a significant risk of traffic accidents. The park management needs to use intelligent prediction to proactively control illegal parking.

[0050] Multi-source data fusion acquisition module: Eight starlight-level high-definition cameras with infrared night vision capabilities are deployed at the east and south gates of the park and the main road of Industrial Avenue. These cameras can clearly capture truck license plates and illegal parking 24 hours a day, acquiring video at 25 frames per second. Coordinate positioning is combined with a high-precision electronic map of the park, with an error controlled within 0.5 meters. Data agreements have been signed with logistics companies in the park, connecting to the onboard OBD devices of 150 freight trucks. GPS trajectory, load status, and braking frequency data are acquired through a 10Hz sampling frequency. 5G private network transmission ensures stability in the industrial environment. Telecommunications operators provide anonymized mobile signaling data within a 3km radius of the park, recording employee commuting trajectories every 5 minutes. Cluster analysis can identify commuter flow patterns in different factory areas. All external data is anonymized, removing personal identifiers. Meteorological department API provides real-time push data on wind speed and precipitation probability, with special emphasis on weather conditions affecting freight safety, such as typhoons and heavy fog, updated every 10 minutes. Figure 2 As shown, the timestamps of all data are converted to a unified UTC time, time slices are divided into 15-minute intervals and aligned, vehicle density features are extracted from multi-source data and standardized, multi-dimensional feature vectors are constructed according to the grid, the spatiotemporal consistency of the data is verified, abnormal data is marked and repaired or smoothed through association reasoning, and the data is stored using a distributed time-series database. A composite index is designed and a multi-dimensional database of illegal parking influencing factors is constructed for storage.

[0051] The spatiotemporal self-attention feature extraction module: inputs truck parking locations obtained from video analysis, load-speed correlation data from onboard sensors, employee commuting trajectories, and meteorological indicators into the spatiotemporal graph self-attention model, and defines a joint spatial-temporal attention matrix. , Let L be the number of grid cells and L be the number of time slices. The joint attention matrix is: ,in The learnable weight matrix is ​​optimized through cross-modal comparative learning, and the formula is: The model uses spatial correlation matrix analysis to analyze, for example, the "illegal parking transmission effect between the grid at the entrance of the warehouse center and the grid on the main road of the park," that is, when the loading and unloading area of ​​the warehouse center is saturated, illegal parking of trucks may spread to the main road; the time dependence matrix captures the periodic patterns of peak freight periods, such as the high incidence of illegal parking caused by concentrated raw material transportation every Monday morning from 9:00 to 10:00. By using a joint attention matrix to weightedly fuse multi-source features, feature vectors containing key information such as "spatiotemporal distribution of truck traffic" and "tidal patterns of employee commuting" are extracted.

[0052] Dynamic granularity parameter generation module: Targeting the characteristics of industrial parks—fixed freight cycles and concentrated commuter traffic—the system presets four granularity modes: "fine-grained during peak freight hours" and "coarse-grained during off-peak hours at night." When the truck traffic on Industrial Avenue exceeds a threshold (e.g., 20 vehicles every 15 minutes), the fuzzy granularity matching mechanism automatically increases the weight of the "large spatial grid + short time window" mode, generating dynamic granularity parameters suitable for different road segments. Let the road segment feature vector be... The granularity parameters are generated through a fuzzy rule engine, and the calculation formula is as follows: For example, the spatial granularity can be adjusted from 50m×50m to 30m×30m, and the time granularity can be shortened from 1 hour to 15 minutes to accurately capture high-frequency areas where trucks temporarily stop.

[0053] Spatiotemporal Hybrid Model Prediction Module: Utilizing a spatiotemporal graph temporal series hybrid model, combined with dynamic granularity parameters and spatiotemporal attention features, it balances the coarse-grained overall freight trend of the park with the fine-grained parking pressure of individual warehouse areas through granular gating units. The formula is as follows: The spatiotemporal graph neural network layer in the model can learn complex patterns such as "the lag correlation between illegal parking of trucks at the South Gate and the remaining number of parking spaces in the warehouse center". For example, when the remaining number of parking spaces in the warehouse center is less than 20%, the probability of illegal parking around the South Gate will increase significantly after 30 minutes. The model will eventually output the probability of illegal parking in each grid in the next 2 hours, providing early warning for logistics scheduling.

[0054] The illegal parking warning visualization module displays the prediction results through the park's smart management platform. The heat map divides the probability of illegal parking into 5 color levels. For example, the yellow area indicates a 40%-50% probability of illegal parking in the next hour, while the red area corresponds to more than 70%. The spatiotemporal dynamic map highlights the migration path of illegal parking hotspots of freight vehicles, such as the dynamic process of "spreading from the east gate to the warehousing center during the morning freight peak period". Based on the warning results, the management department opens temporary waiting areas in the red area in advance to guide trucks to park in an orderly manner. At the same time, the patrol routes are optimized, and additional security personnel are deployed to the employee entrances and exits 30 minutes before the end of the middle shift.

[0055] In summary, to address the issue of illegal parking on roads surrounding industrial parks, the system collects video footage of park roads, freight vehicle trajectories, employee mobile phone signaling, and meteorological data. It extracts features using a spatiotemporal graph self-attention model, generates dynamic parameters through fuzzy granularity matching, and predicts the probability of illegal parking using a spatiotemporal graph temporal hybrid model. The visualized results provide a basis for management, effectively solving the problem of illegal parking of trucks and private cars in industrial parks, reducing traffic delays and accident hazards, and optimizing traffic management around the parks.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A deep learning-based spatiotemporal prediction system for illegal parking on road sections, characterized in that, The system includes: a multi-source data fusion acquisition module, a spatiotemporal self-attention feature extraction module, a dynamic granular parameter generation module, a spatiotemporal hybrid model prediction module, and a parking violation warning visualization module; The multi-source data fusion acquisition module collects traffic camera video streams, vehicle sensor data, mobile phone signaling data and meteorological data in real time, desensitizes all external data simultaneously, and maps heterogeneous data to a unified GIS grid with the help of spatiotemporal coordinate calibration to build a multi-dimensional database of illegal parking influencing factors. The spatiotemporal self-attention feature extraction module: inputs the preprocessed multi-source data into the spatiotemporal graph self-attention model, generates a joint attention matrix by calculating the spatial correlation matrix and the time dependency matrix, extracts spatiotemporal features from the multi-source data, and obtains spatiotemporal feature vectors; The dynamic granularity parameter generation module analyzes the extracted spatiotemporal feature vectors through a fuzzy granularity matching mechanism, calculates the weights of each preset granularity mode based on the road segment characteristics, generates dynamic granularity parameters suitable for different road segments, and passes them to the spatiotemporal hybrid model prediction module. The spatiotemporal hybrid model prediction module adopts a spatiotemporal graph temporal hybrid model, combines dynamic granularity parameters and spatiotemporal attention features, aggregates the hidden states of the grid, and fuses multi-granularity features through granularity gating units to output the probability of illegal parking for each grid in future time periods. The illegal parking warning visualization module overlays the predicted illegal parking probability onto the electronic map to form an illegal parking hotspot warning system, and generates patrol route optimization and facility planning suggestions.

2. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 1, characterized in that, The multi-source heterogeneous data collected in real time by the multi-source data fusion acquisition module is: Traffic camera video stream acquisition: Deploy high-definition cameras to acquire road video in real time at a frame rate of 25 frames per second, covering lanes, roadside parking lanes and sidewalk areas, and simultaneously record the camera's latitude and longitude coordinates and timestamps; Vehicle sensor data acquisition: GPS trajectory, vehicle speed, and acceleration data are acquired through the vehicle OBD device and navigation terminal at a sampling frequency of 10Hz, and transmitted to the server in real time using a 4G / 5G communication module; Mobile signaling data collection: In cooperation with telecommunications operators, anonymized base station handover records and signal strength data are obtained, with a sampling interval of 5 minutes per base station; Meteorological data acquisition: Connects to the meteorological department's API to obtain real-time data on temperature, precipitation intensity, wind speed, and visibility within the region, with an update frequency of 10 minutes per update.

3. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 1, characterized in that, The specific steps in the multi-source data fusion acquisition module to map heterogeneous data to a unified GIS grid using spatiotemporal coordinate calibration are as follows: (1) Preprocess traffic camera video and vehicle sensor multi-source data, and add spatiotemporal labels to each data; (2) Divide the study area into dynamic grids, map the data points to the corresponding grids according to their coordinates and generate codes; (3) Convert the timestamps of all data to a uniform UTC time, divide the time into 15-minute intervals and align them; (4) Extract vehicle density features from multi-source data and standardize them, and construct multi-dimensional feature vectors according to the grid; (5) Verify the spatiotemporal consistency of data, mark abnormal data and repair or smooth it through correlation reasoning; (6) A distributed time-series database is used to store the data, and a composite index is designed and a multi-dimensional database of factors affecting illegal parking is constructed for storage.

4. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 1, characterized in that, The spatiotemporal self-attention feature extraction module constructs a spatiotemporal graph self-attention model and defines a space-time joint attention matrix. , Let L be the number of grid cells and L be the number of time slices. The joint attention matrix is: ,in It is a spatial correlation matrix, describing the spatial dependencies between grids. It is a time dependency matrix, describing the temporal correlation at different times. The learnable weight matrix is ​​optimized through cross-modal comparative learning, and the formula is: ,in The tags are for illegal parking, with 1 indicating a strong association and 0 indicating no association. It is a matrix index, corresponding to the association pairs of different grids or time slices.

5. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 4, characterized in that, In the spatiotemporal graph self-attention model It is a spatial incidence matrix. It is a time dependency matrix and a spatial incidence matrix. The calculation formula is: ,in The Euclidean distance between grid cells. The intensity of personnel flow between the two grids, The spatial decay factor, Representing different grids; time dependency matrix The calculation formula is: ,in For periodic parameters, For date type matching functions, It is the time in the time dependency matrix. and Dependency value, time Date type, Representing different points in time, Representative moment Date type.

6. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 1, characterized in that, The calculation of dynamic granularity parameters in the dynamic granularity parameter generation module is based on the following: Let the road segment feature vector be... The granularity parameters are generated through a fuzzy rule engine, and the calculation formula is as follows: ,in, This is a time granularity parameter, referring to the size of the time interval for dividing data or conducting analysis along the time dimension. Different time granularities can reflect the characteristics of illegal parking on road sections at different time scales. It is a spatial granularity parameter, referring to the size of the spatial range for dividing data or conducting analysis in a spatial dimension. To preset the number of granularity modes, It is an index variable used to iterate through preset granularity patterns, with a value range of 1 to... , It is the mapping function of the fuzzy rule engine, which generates granular parameters based on the feature vector. It is the first The weights of the preset granularity patterns are calculated using the following formula: , It is the number of dimensions of the road segment feature vector. Features For pattern The membership function is given by the formula: , Dimension index of feature vector and Mean and variance parameters of fine-grained models, It is a characteristic value of the traffic level of a road segment. This represents the natural exponential function. No. The temporal and spatial granularity parameters of the model.

7. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 1, characterized in that, The construction formula for the spatiotemporal graph temporal series mixture model in the spatiotemporal mixture model prediction module is as follows: ,in, Is the grid g in the future The probability of illegal parking at any given time. It is the Sigmoid activation function, which maps the output to the [0, 1] probability space. GGU is a granular gating unit that fuses features of different granularities. It refers to the number of layers in a parallel spatiotemporal graph neural network. This represents the layer index of a parallel spatiotemporal graph neural network. These are the attention weights of the k-th layer of the network, generated by the spatiotemporal attention matrix. It is a k-th layer spatiotemporal graph neural network that outputs the hidden state of grid g at time t. These represent coarse-grained and fine-grained spatiotemporal features, respectively. It is a spatiotemporal attention matrix that encodes the spatiotemporal dependencies between grids. It is an element-wise multiplication that combines attention weights and feature representations.

8. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 7, characterized in that, The calculation formula for the parameters in the spatiotemporal graph temporal hybrid model, specifically the granular gating unit (GGU), is as follows: The gating mechanism is as follows: , It is a granularity fusion weight that dynamically adjusts the contribution of coarse and fine granular features. It is the Sigmoid activation function. These are learnable parameters. Indicates feature concatenation operation; The formula for calculating the spatiotemporal attention matrix is: ,in: It is a spatial incidence matrix, representing a grid. Spatial dependency with other grids, It is a time dependency matrix, representing The relationship between a moment and a historical period It is a learnable weight matrix; the calculation formula for the spatiotemporal graph neural network layer is: ,in: This represents the layer index of a parallel spatiotemporal graph neural network. It is a grid exist The characteristics of time, It is an aggregation function that uses attention-weighted aggregation of neighbor information. No. The weight matrix of a layered spatiotemporal graphical neural network is a learnable parameter used to perform a weighted transformation on the input features. This indicates that an aggregation operation is being performed. Represents a grid Spatial Neighbor Sets In the middle, in the historical time step The feature set, where It is a neighbor grid index. Indicates a historical time step. Grid The set of spatial neighbors, i.e., with the grid In a spatially adjacent set of grids, the size of the time window defined by R determines the time range for considering historical features. No. The bias term in a layered spatiotemporal graph neural network is a learnable parameter used to adjust the computation results.

9. The deep learning-based spatiotemporal prediction system for illegal parking on road sections according to claim 1, characterized in that, The illegal parking warning visualization module uses a combination of heat map and spatiotemporal dynamic map for display. The heat map shows the probability of illegal parking in each grid, and the spatiotemporal dynamic map shows the evolution of illegal parking hotspots. The color mapping of the heat map uses the quantile mapping method to divide the probability of illegal parking into 5 levels, with each level corresponding to a different color range.

Images