A parking event detection method and apparatus

Abstract

The application relates to a parking event detection method and device, which comprises the following steps: acquiring vibration signal data of multiple lanes; performing pretreatment and global vibration trend analysis on the vibration signal data to obtain a multi-lane vibration state matrix; determining a candidate area of a suspected parking event according to the multi-lane vibration state matrix, analyzing the neighborhood disturbance features of the candidate area to obtain a neighborhood disturbance feature map; inputting the candidate area and the neighborhood disturbance feature map into a preset graph neural network to obtain a parking event detection result of the candidate area; and through the steps of acquiring vibration signal data based on a grating array, performing pretreatment and analysis to obtain a state matrix, determining a candidate area, analyzing neighborhood disturbance features, and inputting the graph neural network to detect a parking event, parking events on an expressway can be effectively identified, the accuracy and reliability of parking event detection are improved, the false positive rate is reduced, various light and weather conditions can be adapted to, and the expressway full-section coverage monitoring can be realized.

Description

Technical Field

[0001] This invention relates to the field of fiber optic sensor technology, and in particular to a parking event detection method and apparatus. Background Technology

[0002] On highways, where vehicles travel at high speeds, abnormal parking can easily lead to rear-end collisions, chain collisions, and other serious accidents. In particular, when vehicles illegally occupy the emergency lane, they can obstruct the passage of rescue vehicles and delay the critical rescue time. Therefore, it is crucial for the highway system to be able to detect illegally parked vehicles in a timely manner to ensure traffic safety and smooth flow.

[0003] Current common detection methods are based on video surveillance, using fixed cameras or electronic eyes for real-time monitoring, combined with image processing algorithms such as background subtraction, target tracking, and deep learning to identify abnormal parking behavior. These methods have the following problems: First, they have poor environmental adaptability; cameras are affected by changes in lighting conditions such as backlighting and low light at night, as well as weather conditions such as rain, snow, and fog, resulting in a high false detection rate. Second, the monitoring coverage is limited, only monitoring areas within the camera's field of view, failing to achieve seamless coverage of the entire road segment. Finally, they are computationally intensive; AI algorithms based on video images have low computational efficiency in complex scenarios, requiring reference to multiple consecutive frames of data, resulting in insufficient real-time performance. There are also abnormal parking identification methods based on induction coils or infrared sensors; these methods are all point-based sensing and cannot achieve full-area detection of parking on highways. Some methods utilize grating array vibration sensing technology combined with traditional rule-based parking detection methods. This method is simple and fast, but parking events themselves do not generate continuous or significant vibration signals. Vibration analysis methods based on thresholds or simple pattern recognition have poor ability to distinguish between moving vehicles and parked vehicles, and are prone to misjudging traffic flow gaps and sensor noise interference as parking. This threshold-dependent judgment method is difficult to distinguish between parking and low-speed driving, and has poor flexibility, weak generalization ability, and poor accuracy. When deployed on a large scale, it is necessary to adjust parameters interval by interval.

[0004] Therefore, there is an urgent need to propose a parking event detection method and device to solve the technical problem that existing vibration analysis methods based on thresholds or simple pattern recognition have poor ability to distinguish between moving vehicles and parked vehicles, and are prone to misjudging traffic flow gaps and sensor noise interference as parking. Summary of the Invention

[0005] In view of this, it is necessary to provide a parking event detection method and device to solve the technical problem that existing vibration analysis methods based on thresholds or simple pattern recognition have poor ability to distinguish between moving vehicles and parked vehicles, and are prone to misjudging traffic flow gaps and sensor noise interference as parking.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a parking event detection method, comprising:

[0007] Vibration signal data of multiple lanes on a highway are acquired using a grating array.

[0008] The vibration signal data is preprocessed and a global vibration situation analysis is performed to obtain a multi-lane vibration state matrix;

[0009] Candidate regions are determined based on the multi-lane vibration state matrix, and the neighborhood disturbance characteristics of the candidate regions are analyzed to obtain a neighborhood disturbance characteristic map.

[0010] The candidate region and the neighborhood perturbation feature map are input into a preset graph neural network to obtain the parking event detection result of the candidate region.

[0011] In one possible implementation, the analysis of the neighborhood perturbation features of the candidate region to obtain a neighborhood perturbation feature map includes:

[0012] The spatiotemporal neighborhood of the candidate region is analyzed in the multi-lane vibration state matrix to obtain neighborhood disturbance characteristics; the spatiotemporal neighborhood includes the upstream region of the lane where the candidate region is located and the corresponding spatial region of the adjacent lane.

[0013] A neighborhood perturbation feature map is constructed based on the neighborhood perturbation features.

[0014] In one possible implementation, the preprocessing and global vibration state analysis of the vibration signal data to obtain a multi-lane vibration state matrix includes:

[0015] The vibration signal data is preprocessed to obtain preprocessed signal data;

[0016] The comprehensive vibration state value is obtained by calculating the signal data of each grating on each lane in the preprocessed signal data based on the spatiotemporal window.

[0017] The comprehensive vibration state values ​​are sorted at each moment according to the lane as the row and the spatial position of the grating as the column to obtain the multi-lane vibration state matrix.

[0018] In one possible implementation, the step of calculating the signal data of each grating on each lane in the preprocessed signal data based on a spatiotemporal window to obtain a comprehensive vibration state value includes:

[0019] The average absolute amplitude is obtained by calculating the average value of the preprocessed signal data in each spatiotemporal window;

[0020] The proportion of signal energy in each grating on each lane within the preset energy ratio in the preprocessed signal data is calculated to obtain the energy proportion.

[0021] The average absolute amplitude and the energy percentage are normalized and weighted to obtain the comprehensive vibration state value.

[0022] In one possible implementation, determining the candidate region based on the multi-lane vibration state matrix includes:

[0023] Obtain the historical state value set for each grating in each lane;

[0024] The mean and standard deviation of the historical state value set are calculated, and the difference between the mean and the standard deviation is determined as the background state threshold.

[0025] The gratings in the multi-lane vibration state matrix whose state values ​​for a consecutive preset number of periods are less than the background state threshold are marked as candidate regions for suspected parking events.

[0026] In one possible implementation, the analysis of the mode changes of the spatiotemporal neighborhood of the candidate region in the multi-lane vibration state matrix to obtain neighborhood disturbance characteristics includes:

[0027] Determine the spatiotemporal neighborhood of the candidate region; the spatiotemporal neighborhood is the spatial range of multiple gratings taken forward and backward from the grating spatial position as the center, and the same spatial range on the two adjacent lanes to the left and right, and the time range is the continuous time period from the preset time before the state value first falls below the background state threshold to the current time.

[0028] The feature vector corresponding to each grating-time relationship node in the spatiotemporal neighborhood is determined as the neighborhood perturbation feature.

[0029] In one possible implementation, constructing a neighborhood perturbation feature map based on the neighborhood perturbation features includes:

[0030] The relation nodes of the neighborhood perturbation features are determined as graph nodes;

[0031] Using the graph nodes as vertices, nodes whose spatial distance is within multiple gratings and whose time difference is within a preset time are connected to form edges to construct a neighborhood perturbation feature graph; the weight of the edge is the reciprocal of the spatiotemporal distance between nodes.

[0032] In one possible implementation, the step of inputting the candidate region and the neighborhood perturbation feature map into a preset graph neural network to obtain the parking event detection result of the candidate region includes:

[0033] The candidate region and the neighborhood perturbation feature map are input into a preset graph neural network to obtain the event probability of the candidate region; the preset graph neural network is a model obtained by training offline using historical data; the offline historical data includes "real parking" and "fake parking" event samples, as well as the corresponding feature maps;

[0034] The parking event detection result is determined based on the event probability.

[0035] In one possible implementation, determining the parking event detection result based on the event probability includes:

[0036] When the probability of the event is greater than or equal to the preset decision threshold, the parking event detection result is determined to be that the lane and grating spatial position represented by the graph node corresponding to the neighborhood disturbance feature map exist from the time the state value first falls below the background state threshold to the current time.

[0037] Secondly, the present invention also provides a parking event detection device, comprising:

[0038] The signal acquisition module is used to acquire vibration signal data from multiple lanes on a highway based on a grating array.

[0039] The state analysis module is used to preprocess the vibration signal data and perform global vibration situation analysis to obtain a multi-lane vibration state matrix.

[0040] The region determination module is used to determine candidate regions based on the multi-lane vibration state matrix and analyze the neighborhood disturbance characteristics of the candidate regions to obtain a neighborhood disturbance feature map.

[0041] The event detection module is used to input the candidate region and the neighborhood perturbation feature map into a preset graph neural network to obtain the parking event detection result of the candidate region.

[0042] The beneficial effects of this invention are as follows: Vibration signal data from multiple lanes on a highway are acquired based on a grating array; the vibration signal data is preprocessed and a global vibration situation analysis is performed to obtain a multi-lane vibration state matrix; candidate areas for suspected parking events are determined based on the multi-lane vibration state matrix, and the neighborhood disturbance characteristics of the candidate areas are analyzed to obtain a neighborhood disturbance feature map; the candidate areas and the neighborhood disturbance feature map are input into a preset graph neural network to obtain the parking event detection results for the candidate areas; by acquiring vibration signal data based on a grating array, performing preprocessing and analysis to obtain a state matrix, determining candidate areas and analyzing neighborhood disturbance characteristics, and inputting the data into a graph neural network to detect parking events, this invention can effectively identify parking events on highways, improving the accuracy and reliability of parking event detection, reducing false alarm rates, adapting to various lighting and weather conditions, and achieving full-section coverage monitoring of the highway. Attached Figure Description

[0043] Figure 1 This is a schematic flowchart of an embodiment of the parking event detection method provided by the present invention;

[0044] Figure 2 For the present invention Figure 1 A schematic flowchart of an embodiment of step S102;

[0045] Figure 3 For the present invention Figure 2 A schematic diagram of an embodiment of step S202;

[0046] Figure 4 This is a schematic diagram of an embodiment of the parking event detection device provided by the present invention. Detailed Implementation

[0047] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0048] like Figure 1 As shown, a specific embodiment of the present invention discloses a parking event detection method, comprising:

[0049] S101. Obtain vibration signal data of multiple lanes on a highway based on a grating array.

[0050] The grating array in this invention refers to a sensor network composed of multiple fiber optic gratings. This array is typically laid under or beside roads to sense minute vibrations generated by the road surface or passing vehicles in real time, converting these vibrations into measurable optical signals. By demodulating and analyzing these optical signals, vibration information caused by vehicle movement or parking can be obtained. Vibration signal data refers to the raw data collected by the grating array, reflecting the vibration intensity and patterns of multiple lanes on a highway at different time points. This data forms the basis for subsequent analysis and includes potential information such as vehicle type, speed, location, and the presence of abnormal parking.

[0051] S102. Perform preprocessing and global vibration status analysis on the vibration signal data to obtain the multi-lane vibration state matrix.

[0052] Preprocessing in this invention refers to a series of operations performed on the raw vibration signal data to eliminate noise, correct biases, and smooth the data, thereby improving data quality and the accuracy of subsequent analysis. Common preprocessing methods include filtering, detrending, and normalization. Global vibration status analysis refers to a holistic analysis of the preprocessed vibration signal data to understand the overall vibration status and trends of multiple lanes on a highway over a period of time. This analysis can identify abnormal vibration patterns, providing macroscopic background information for subsequent parking event detection. A multi-lane vibration state matrix is ​​a data structure used to represent the vibration state of multiple lanes on a highway at different times and spatial locations. This matrix typically uses time as the dimension and lane and grating spatial location as dimensions, recording the comprehensive vibration state value of each grating at each moment, thus forming a multi-dimensional vibration state view.

[0053] S103. Based on the multi-lane vibration state matrix, candidate regions are determined, and the neighborhood disturbance characteristics of the candidate regions are analyzed to obtain a neighborhood disturbance characteristic map.

[0054] In this embodiment of the invention, the candidate region refers to the spatiotemporal range in the multi-lane vibration state matrix that may contain parking events, identified through preliminary analysis. The vibration patterns in these regions may differ significantly from those of normally moving vehicles, requiring further refined analysis for confirmation. Neighborhood disturbance characteristics refer to the changes in vibration patterns relative to normal conditions or the surrounding environment within the spatiotemporal neighborhood of the suspected parking event candidate region. These characteristics can reflect the impact of a parking event on surrounding traffic flow and road surface vibration, such as local disturbances caused by vehicle deceleration, stopping, or restarting.

[0055] S104. Input the candidate region and neighborhood perturbation feature map into the preset graph neural network to obtain the parking event detection result of the candidate region.

[0056] The preset graph neural network in this invention is a trained deep learning model specifically designed for processing graph-structured data. This network can learn complex spatial and temporal dependencies from neighborhood perturbation feature maps, thereby classifying or regressing candidate regions of suspected parking events to determine whether they are genuine parking events. The parking event detection result refers to the final judgment on the candidate regions of suspected parking events output by the preset graph neural network. This result typically includes the probability of the event occurring, the event type (e.g., parking, low-speed driving), and the time and spatial location of the event, providing a basis for decision-making by traffic management departments.

[0057] First, vibration signal data from multiple lanes of the highway is acquired using a grating array. This step is fundamental to the entire detection process. Grating array sensors deployed beneath the highway surface can continuously and in real-time collect vibration information generated by passing vehicles. For example, the grating array can be designed to be linearly arranged along the lane direction, with each grating responsible for monitoring the vibration in its covered area. When a vehicle passes, its weight and speed cause minute deformations in the road surface, resulting in a shift in the center wavelength of the grating. These shifts are captured by the grating demodulator and converted into electrical signals, forming the raw vibration signal data. This data can be directly stored or transmitted to the data processing unit via wired or wireless means. For example, four independent vibration intensity-space-time signals D1(x,t), D2(x,t), D3(x,t), and D4(x,t) for each lane, where x is the grating number and t is time.

[0058] Secondly, the vibration signal data is preprocessed and a global vibration state analysis is performed to obtain a multi-lane vibration state matrix. After acquiring the raw vibration signal data, a series of processes are required to extract useful information. Preprocessing may include simple low-pass filtering of the raw signal to remove high-frequency noise interference; or baseline correction to eliminate errors caused by sensor drift. Global vibration state analysis may involve simply averaging the vibration amplitude of each grating within a fixed time window to roughly reflect the activity level of the area. Subsequently, these average vibration amplitudes can be organized into a matrix, where rows represent different lanes, columns represent different grating spatial locations, and each cell stores the average vibration amplitude of the corresponding grating within a specific time window, thus forming a multi-lane vibration state matrix.

[0059] Next, candidate areas for suspected parking events are determined based on the multi-lane vibration state matrix, and the neighborhood disturbance characteristics of these candidate areas are analyzed to obtain a neighborhood disturbance feature map. After obtaining the multi-lane vibration state matrix, a fixed vibration threshold can be set. For example, when the vibration amplitude of a certain grating is continuously lower than the threshold for a period of time, it is initially identified as a candidate area for a suspected parking event. This threshold can be manually set based on empirical values ​​or historical data. For these identified candidate areas, the vibration patterns around them can be further analyzed. For example, the changes in vibration amplitude of adjacent gratings around the candidate area within the same time period can be statistically analyzed, and these changes can be used as neighborhood disturbance features. Subsequently, these neighborhood disturbance features can be represented as a simple two-dimensional grid diagram, where each grid point represents a grating, the connection between grid points represents spatial adjacency, and the value on the grid point represents its disturbance characteristics, thus forming a neighborhood disturbance feature map.

[0060] Finally, the candidate region and its neighborhood perturbation feature map are input into a pre-set graph neural network to obtain the parking event detection result for the candidate region. After constructing the candidate region information and neighborhood perturbation feature map, this data is fed into a pre-trained graph neural network model. This graph neural network can be a simple multilayer perceptron, whose input layer receives the identification information of the candidate region and a flattened representation of the neighborhood perturbation feature map. Through the network's calculation, a probability value indicating whether the candidate region is a parking event can be output. For example, if the output probability is higher than a certain preset value, it is determined that a parking event exists in the candidate region; otherwise, it is determined to be a non-parking event. This detection result can be a binary classification result, indicating whether a parking event exists.

[0061] Compared with existing technologies, this embodiment provides a method for acquiring vibration signal data from multiple lanes of a highway based on a grating array; preprocessing and analyzing the vibration signal data to obtain a multi-lane vibration state matrix; determining candidate areas for suspected parking events based on the multi-lane vibration state matrix, and analyzing the neighborhood disturbance characteristics of the candidate areas to obtain a neighborhood disturbance feature map; inputting the candidate areas and the neighborhood disturbance feature map into a preset graph neural network to obtain parking event detection results for the candidate areas; by acquiring vibration signal data based on a grating array, preprocessing and analyzing it to obtain a state matrix, determining candidate areas and analyzing neighborhood disturbance characteristics, and inputting the data into a graph neural network to detect parking events, this method can effectively identify parking events on highways, improve the accuracy and reliability of parking event detection, reduce false alarm rates, adapt to various lighting and weather conditions, and achieve full-section coverage monitoring of the highway.

[0062] In some embodiments of the present invention, such as Figure 2 As shown, step S102 includes:

[0063] S201. Preprocess the vibration signal data to obtain preprocessed signal data.

[0064] This preprocessing aims to eliminate noise, interference, and unwanted signal components in the original vibration signal data to improve the signal-to-noise ratio and the accuracy of subsequent analysis. For example, preprocessing may include filtering the original signal, such as using a low-pass filter to remove high-frequency noise or a band-pass filter to retain vibration information within a specific frequency range. Furthermore, preprocessing may also include baseline drift correction, outlier removal, and other operations to ensure signal stability and validity.

[0065] S202. Based on the spatiotemporal window, calculate the signal data of each grating on each lane in the preprocessed signal data to obtain the comprehensive vibration state value.

[0066] This step aims to extract representative quantitative indicators from local, instantaneous signal data that reflect the vibration characteristics of the current spatiotemporal region. The introduction of a spatiotemporal window allows the evaluation of vibration state to consider not only the vibration of a single grating at a specific moment but also its vibration trend over a period of time and within adjacent spatial ranges. The comprehensive vibration state value can be a single numerical value obtained by comprehensively considering multiple dimensions such as signal amplitude, energy, and frequency characteristics. For example, it can calculate the root mean square value, peak value, or energy of a specific frequency band within the spatiotemporal window.

[0067] S203. Sort the comprehensive vibration state values ​​at each time step according to the lane as the row and the spatial position of the grating as the column to obtain the multi-lane vibration state matrix.

[0068] This step aims to organize the discrete, distributed vibration state values ​​across different lanes and gratings into a structured, easily analyzable two-dimensional data representation. Each row of this matrix represents a lane, and each column represents the spatial location of a grating within that lane. Each element in the matrix represents the combined vibration state value of the corresponding lane and grating at the current moment. This matrix representation can intuitively display the overall vibration status of the highway at a given moment, facilitating subsequent pattern recognition and anomaly detection.

[0069] This invention first preprocesses the raw vibration signal data to ensure the data quality for subsequent analysis. Then, the concept of a spatiotemporal window is introduced to calculate the preprocessed signal data for each grating Pj on each lane Li for a duration T of 1 second, thereby obtaining a comprehensive vibration state value S(i, j,t) that comprehensively reflects the local vibration characteristics. This spatiotemporal window-based calculation method effectively captures the continuity and correlation of the vibration signal in time and space, avoiding errors that may arise from judging based solely on instantaneous or single grating data. Finally, these comprehensive vibration state values ​​are arranged in an orderly manner according to the spatial positions of the lanes and gratings at each moment, constructing a two-dimensional multi-lane vibration state matrix M[t]. This matrix, in its structured form, clearly presents the global vibration situation of the highway at a specific moment, providing a stable, reliable, and spatiotemporally contextualized input for subsequent parking event detection. This systematic data organization and feature extraction method makes it possible to accurately identify abnormal parking events from complex grating vibration signals.

[0070] In some embodiments of the present invention, such as Figure 3 As shown, step S202 includes:

[0071] S301. Calculate the average value of the preprocessed signal data in each spatiotemporal window to obtain the average absolute amplitude.

[0072] This step aims to quantify the overall strength of the signal. The average absolute amplitude is an effective indicator of signal energy or strength, reflecting the degree of physical disturbance experienced by the grating array when a vehicle passes or stops. This average absolute amplitude can be obtained by summing the absolute values ​​of all preprocessed signal data within the spatiotemporal window and then dividing by the total number of data points; alternatively, it can be calculated by integrating the absolute values ​​of the signal within the spatiotemporal window and then dividing by the length of the time window.

[0073] S302. Calculate the proportion of signal energy of each grating in each lane within the preset energy ratio in the preprocessed signal data to obtain the energy ratio.

[0074] This step is for a more refined analysis of the signal's energy distribution characteristics. This indicator helps distinguish between different types of vibration events, such as the continuous vibration of a vehicle in motion versus the weak vibration or vibration within a specific frequency range that may occur when the vehicle is stationary. This energy percentage can be calculated by first calculating the total energy of the signal within a spatiotemporal window, then determining the number of data points whose energy values ​​fall within a preset proportion range (e.g., a frequency band of 5-30 Hz), and finally dividing this number by the total number of data points. Alternatively, the signal can be analyzed in the frequency domain to calculate the energy of different frequency components, and then the proportion of the frequency component energy within the preset energy proportion range to the total energy can be calculated.

[0075] S303. Normalize and weight the average absolute amplitude and energy percentage to obtain the comprehensive vibration state value.

[0076] The normalization process aims to eliminate differences in dimensions and numerical ranges between different features, ensuring fairness in subsequent calculations. Normalization methods can include Min-Max normalization or Z-score normalization. Weighted summation assigns different weights to the average absolute amplitude and energy percentage based on their importance in parking event detection, thus fusing these two complementary features into a single, more representative comprehensive vibration state value. The weights can be set based on experience or optimized using machine learning methods.

[0077] This invention combines two indicators reflecting vibration characteristics from different dimensions—mean absolute amplitude and energy percentage—and performs normalization and weighted summation to more comprehensively and accurately characterize the vibration characteristics S(i, j, t) of each grating within each spatiotemporal window. The S(i, j, t) ranges between [0, 1], with larger values ​​indicating more active vibration. The mean absolute amplitude directly reflects the vibration intensity, while the energy percentage reveals the intrinsic structure and pattern of the vibration signal. By fusing these two complementary features, the limitations of a single indicator can be avoided. For example, vibration amplitude alone may not be sufficient to distinguish between the weak vibrations of a vehicle moving slowly and when it is stopped. This comprehensive calculation method provides more reliable basic data for the subsequent construction of a multi-lane vibration state matrix, thereby significantly improving the accuracy of identifying parking events.

[0078] In some embodiments of the present invention, step S103 includes:

[0079] Obtain the set of historical state values ​​for each grating in each lane.

[0080] The historical state value set refers to the sequence data of vibration state values ​​recorded by a specific grating on a specific lane over a certain period of time. This data reflects the vibration characteristics of the grating under normal traffic flow. This set can be obtained through continuous monitoring and data storage.

[0081] The mean and standard deviation of the historical state value set are calculated, and the difference between the mean and standard deviation is determined as the background state threshold.

[0082] The calculation of the mean and standard deviation is used to quantify the average vibration level and fluctuation range of the grating under normal operating conditions. The mean represents the typical vibration intensity of the grating under normal conditions, while the standard deviation reflects the dispersion or uncertainty of the vibration intensity. These statistics allow for the establishment of a dynamic vibration baseline that reflects the current environmental context.

[0083] The gratings in the multi-lane vibration state matrix whose state values ​​are less than the background state threshold for a consecutive preset number of cycles are marked as candidate regions for suspected parking events.

[0084] This step aims to establish an adaptive judgment benchmark. When the vibration state value of the grating is below this threshold, it indicates that its vibration level is significantly lower than the normal background, which may indicate the occurrence of a parking event.

[0085] This invention overcomes the problem of fixed thresholds being susceptible to interference in complex highway environments by dynamically establishing a background vibration state threshold θ(i, j) for each grating. Specifically, the system first continuously collects historical vibration state values ​​for each grating in each lane. These historical data reflect the vibration characteristics of the grating under normal traffic conditions (e.g., a set of historical 10-minute state values ​​for the i-th lane and the j-th grating position). Subsequently, by statistically analyzing these historical state value sets, their mean and standard deviation are calculated, providing a quantitative benchmark for assessing the current vibration state. Based on these statistics, such as the difference between the mean and the standard deviation, the background state threshold for that grating is determined. This threshold is adaptive and can be dynamically adjusted according to changes in actual traffic and environmental conditions. When the current state value of a grating in the multi-lane vibration state matrix is ​​lower than its corresponding background state threshold for several consecutive cycles, it indicates that the vibration level in that area is persistently abnormally low, which is consistent with the characteristic of vehicles being stationary for a long time. Therefore, this grating is marked as a candidate region for a suspected parking event. For example, if the state value S of a grating (Li, Pj) is lower than its background threshold θ(i, j) for 5 consecutive cycles during an element-wise scan of the real-time matrix M[t], then the grating is marked as a "persistent hole candidate", denoted as event Ecand = (Li, Pj, tstart), where tstart is the time when the state first falls below the threshold. This method can effectively filter out transient interference, ensuring that the identified candidate regions have high reliability, and providing more accurate input for subsequent neighborhood disturbance feature analysis and graph neural network detection.

[0086] In some embodiments of the present invention, step S103 further includes:

[0087] The mode changes of the spatiotemporal neighborhood of the candidate region in the multi-lane vibration state matrix are analyzed to obtain the neighborhood disturbance characteristics; the spatiotemporal neighborhood includes the upstream region of the lane where the candidate region is located and the corresponding spatial region of the adjacent lane.

[0088] Construct a neighborhood perturbation feature map based on neighborhood perturbation characteristics.

[0089] This study analyzes the pattern changes of the spatiotemporal neighborhood of candidate regions within the multi-lane vibration state matrix to obtain neighborhood disturbance features. The aim is to capture vibration pattern changes in the surrounding environment of candidate regions from both temporal and spatial dimensions, generating neighborhood disturbance features that reflect the contextual information of potential parking events. Parking events are often accompanied by vehicle deceleration, stopping, and changes in traffic flow in surrounding lanes or upstream areas; these changes manifest as specific patterns in the vibration signal. By analyzing the pattern changes in the spatiotemporal neighborhood, a more comprehensive and accurate understanding can be gained as to whether the abnormal state of the candidate region is indeed caused by parking. This analysis can be implemented in various ways.

[0090] In some embodiments of the present invention, the mode changes of the spatiotemporal neighborhood of the candidate region in the multi-lane vibration state matrix are analyzed to obtain neighborhood disturbance characteristics, including:

[0091] Determine the spatiotemporal neighborhood of the candidate region; the spatiotemporal neighborhood is the spatial range of multiple gratings taken forward and backward from the grating spatial position and the same spatial range on the two adjacent lanes to the left and right; the time range is the continuous time period from the preset time before the state value first falls below the background state threshold to the current time.

[0092] The feature vector corresponding to each grating-time relationship node in the spatiotemporal neighborhood is determined as the neighborhood perturbation feature.

[0093] This step in the embodiments of the present invention aims to provide comprehensive contextual information for candidate areas of suspected parking events, facilitating subsequent analysis of traffic pattern changes in the surrounding area. The determination of the spatiotemporal neighborhood considers not only the spatial location of the candidate area itself but also extends to its upstream, downstream, and adjacent lanes, while also covering the time period before and after the event. This expanded perspective helps capture dynamic changes in vehicle behavior, such as vehicle deceleration, stopping, restarting, or avoidance behavior by vehicles in adjacent lanes. Specifically, the spatiotemporal neighborhood can be defined as a spatial range extending forward (upstream) and backward (downstream) by a predetermined number of gratings, for example, five gratings each, centered on the grating spatial location Pj where the candidate area is located. This spatial range also includes the corresponding spatial areas on the two lanes adjacent to the lane where the candidate area is located. In the time dimension, the time range tstart of the spatiotemporal neighborhood can start from a predetermined time point before the state value first falls below the background state threshold, for example, 5 seconds before the event occurs, and continue until the current moment. Another approach is to dynamically adjust the spatial range based on actual traffic flow and vehicle speed, for example, expanding the spatial range when driving at high speeds and shrinking it when congested; the temporal range can also be adaptively adjusted based on the duration of the event or the type of vehicle.

[0094] Furthermore, the feature vector corresponding to each grating's relationship node with time in the spatiotemporal neighborhood is determined as the neighborhood disturbance feature, aiming to structure the complex traffic state information within the spatiotemporal neighborhood into a numerical representation that can be processed by machine learning models. Each "relationship node" represents the state of a specific grating at a specific time point within the spatiotemporal neighborhood. By extracting "feature vectors" for these nodes, the vibration mode and traffic conditions of the grating at that moment can be quantified. The feature vector can contain various information, such as the comprehensive vibration state value of the grating at that moment, the rate of change of that state value, the historical average state value of the grating, and a binary indication of whether a vehicle has passed. Another approach is to include more complex statistical features in the feature vector, such as the spectral characteristics of the vibration signal, energy distribution characteristics, or state differences with neighboring gratings and adjacent time points, to more comprehensively describe the local disturbance at that node. The set of these feature vectors constitutes a quantitative description of the neighborhood disturbance mode.

[0095] In the aforementioned parking event detection method, vibration signal data is first acquired through a grating array, and after preprocessing and global vibration situation analysis, a multi-lane vibration state matrix is ​​generated. Based on this matrix, and combined with the background state threshold determined by historical state values, candidate areas for suspected parking events can be preliminarily identified. To further improve the accuracy of parking event detection, this application conducts in-depth analysis of the neighborhood disturbance characteristics of these candidate areas. Specifically, for each candidate area Ecand of a suspected parking event, the system first determines a specific spatiotemporal neighborhood surrounding the candidate area. This spatiotemporal neighborhood not only covers the upstream and downstream areas of the lane where the candidate area is located, but also extends to adjacent lanes, and traces back to a continuous period of time before the event occurred. This comprehensive spatiotemporal range ensures that various traffic dynamics related to parking events can be captured, such as vehicle deceleration, stopping, restarting, and the avoidance or normal passage patterns of vehicles in adjacent lanes. Once the spatiotemporal neighborhood is determined, the system extracts the corresponding feature vector for each grating within the neighborhood at each time point, and combines the relationship nodes of these gratings with time and their feature vectors to form neighborhood disturbance characteristics. These feature vectors can quantitatively describe the vibration state, variation trend, and correlation with other gratings of each grating at a specific time. In this way, this application can characterize traffic disturbance patterns around candidate areas from multiple dimensions and perspectives, thus providing rich and detailed input information for subsequent graph neural network analysis. This method can effectively distinguish real parking events from similar signals caused by other factors (such as brief congestion, sensor noise, or slow vehicle passage), significantly enhancing the robustness and accuracy of parking event detection.

[0096] In some embodiments of the present invention, constructing a neighborhood perturbation feature map based on neighborhood perturbation features includes:

[0097] The relational nodes based on neighborhood perturbation features are defined as graph nodes;

[0098] Using graph nodes as vertices, nodes whose spatial distance is within multiple gratings and whose time difference is within a preset time are connected to form edges to construct a neighborhood perturbation feature map; the weight of the edge is the reciprocal of the spatiotemporal distance between nodes.

[0099] In this context, defining the relational nodes of neighborhood perturbation features as graph nodes means abstracting the state or features of each grating in the spatiotemporal neighborhood at a specific point in time as nodes in a graph. The purpose of this abstraction is to structure complex spatiotemporal data, making it processable using graph theory methods. Using graph nodes as vertices, connecting nodes within a spatial distance of multiple gratings (e.g., two) and with a time difference within a preset time (e.g., 2 seconds) to form edges constitutes the neighborhood perturbation feature graph G. This involves establishing connections (edges) between graph nodes, representing spatiotemporal correlations. By setting thresholds for spatial distance and time difference, strongly correlated node pairs can be selected for connection, thus forming a graph structure that reflects local spatiotemporal dynamics. The edge weight is the reciprocal of the spatiotemporal distance between nodes, meaning the edge weight is used to quantify the strength of the association between nodes. Using the reciprocal of the spatiotemporal distance as the weight implies that the closer (more spatiotemporally close) the nodes, the stronger their association and the greater their weight. This aligns with physical intuition, that is, events that are close together and occur within a short timeframe are more likely to influence each other.

[0100] This application first identifies the feature vector corresponding to each grating-time relationship node in the neighborhood perturbation features as a graph node (each node's feature vector consists of its vibration state value S, spatiotemporal coordinate offset relative to the center point, and time-domain features such as zero-crossing rate and spectral centroid extracted from the original signal), thereby transforming discrete spatiotemporal information into processable graph elements. Subsequently, by setting thresholds for spatial distance and time difference, connections (edges) are established between these graph nodes, explicitly characterizing interactions within a local spatiotemporal range. For example, a vibration change on one grating is likely related to the vibration changes of its neighboring gratings within a similar timeframe. In this way, the method effectively filters out irrelevant events occurring at long distances or over long time intervals, focusing on truly influential local spatiotemporal correlations. Furthermore, the edge weights are set to the reciprocal of the spatiotemporal distance between nodes, making the connection weights larger for nodes that are closer together. This physically reflects the influence of spatiotemporal proximity on the strength of event correlations. This weighting mechanism allows the subsequent graph neural network to prioritize more closely related spatiotemporal information during processing. By transforming the original neighborhood perturbation features into a structured neighborhood perturbation feature map, this application provides a representation that can effectively capture the complex spatiotemporal dynamics around parking events, providing richer and more structured inputs for subsequent graph neural networks, thereby significantly improving the accuracy and robustness of parking event detection.

[0101] In some embodiments of the present invention, step S104 includes:

[0102] The candidate region and its neighborhood perturbation feature map are input into a preset graph neural network to obtain the event probability of the candidate region. The preset graph neural network is a model obtained by training offline using historical data. The offline historical data includes "real parking" and "fake parking" event samples and their corresponding feature maps.

[0103] The parking event detection result is determined based on the event probability.

[0104] The preset graph neural network in this embodiment of the invention is as follows: A module consisting of five layers, each composed of a graph convolutional layer and a ReLU activation layer, is constructed. The penultimate layer uses Dropout to prevent overfitting, and the output layer uses log_softmax to calculate the log probability, outputting the event probability p, representing the probability that the candidate event is "parking". This GNN model needs to be trained offline using historical data. Training data is collected by manually annotating or verifying with video data, collecting 500 samples each of "real parking" and "fake parking" events, and constructing corresponding feature maps. Training uses a negative log-likelihood loss function and the Adam optimizer. The training set, validation set, and test set are set in an 8:1:1 ratio. Training uses a negative log-likelihood loss function and the Adam optimizer, with a learning rate of 0.01 and 200 training epochs. The model with the highest accuracy on the validation set is selected as the fully trained graph neural network.

[0105] In some embodiments of the present invention, determining the parking event detection result based on the event probability includes:

[0106] When the event probability is greater than or equal to the preset decision threshold, the parking event detection result is determined to be the existence of a parking event from the time the state value first falls below the background state threshold to the current time, as represented by the graph node corresponding to the neighborhood disturbance feature map.

[0107] The event probability refers to the likelihood that a candidate region will be identified as a parking event by the pre-defined graph neural network. It is typically represented by a value between 0 and 1, with a higher value indicating a higher probability of a parking event occurring in that region. The pre-defined decision threshold is a pre-set value used to convert the event probability into a binary decision (i.e., "yes" or "no"). This threshold can be determined based on historical data analysis, expert experience, or through optimization algorithms (e.g., based on receiver operating characteristic curves or precision-recall curves) to balance detection precision and recall.

[0108] This invention transforms probabilistic judgments into definitive parking event detection results by judging the event probabilities output by a preset graph neural network. Specifically, when the event probability p output by the preset graph neural network for a candidate region reaches or exceeds a preset judgment threshold P (P is 0.85), i.e., p>= P, the system determines that a parking event does indeed exist in the candidate region. This judgment mechanism effectively filters out low-probability false alarms, ensuring the reliability of the detection results. Simultaneously, this scheme further clarifies the location and duration of the parking event. The location is precisely given by the lane and raster spatial positions represented by the graph nodes corresponding to the neighborhood perturbation feature map, while the duration is traced back to the moment when the state value of the region first falls below the background state threshold and continues until the current moment, i.e., a parking event has occurred at lane Li, position Pj, starting from tstart, outputting the lane number, precise station number, event start time, and confidence level. This mechanism, combining probabilistic judgment and precise spatiotemporal positioning, makes the parking event detection results not only accurate but also highly operable, providing timely and comprehensive information for subsequent traffic management and emergency response. In this way, the solution transforms abstract probability values ​​into concrete event reports, solving the problem that probability values ​​alone cannot directly guide practical operations, and significantly improving the practicality and efficiency of parking event detection.

[0109] To better implement the parking event detection method in the embodiments of the present invention, correspondingly, the embodiments of the present invention also provide a parking event detection device, such as... Figure 4 As shown, the parking event detection device 400 includes:

[0110] The signal acquisition module 401 is used to acquire vibration signal data of multiple lanes on a highway based on a grating array;

[0111] State analysis module 402 is used to preprocess vibration signal data and perform global vibration situation analysis to obtain a multi-lane vibration state matrix;

[0112] The region determination module 403 is used to determine candidate regions based on the multi-lane vibration state matrix and analyze the neighborhood disturbance characteristics of the candidate regions to obtain a neighborhood disturbance characteristic map.

[0113] The event detection module 404 is used to input the candidate region and the neighborhood disturbance feature map into the preset graph neural network to obtain the parking event detection result of the candidate region.

[0114] The parking event detection device 400 provided in the above embodiments can implement the technical solutions described in the above parking event detection method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above parking event detection method embodiments, and will not be repeated here.

[0115] The parking event detection method and device provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A parking event detection method, characterized in that, include: Vibration signal data of multiple lanes on a highway are acquired using a grating array. The vibration signal data is preprocessed and a global vibration situation analysis is performed to obtain a multi-lane vibration state matrix; Candidate regions are determined based on the multi-lane vibration state matrix, and the neighborhood disturbance characteristics of the candidate regions are analyzed to obtain a neighborhood disturbance characteristic map. The candidate region and the neighborhood perturbation feature map are input into a preset graph neural network to obtain the parking event detection result of the candidate region; The step of determining the candidate region based on the multi-lane vibration state matrix includes: Obtain the historical state value set for each grating in each lane; The mean and standard deviation of the historical state value set are calculated, and the difference between the mean and the standard deviation is determined as the background state threshold. The grating in the multi-lane vibration state matrix whose state value is less than the background state threshold for a consecutive preset number of cycles is marked as a candidate region for a suspected parking event; The spatiotemporal neighborhood of the candidate region is analyzed in the multi-lane vibration state matrix to obtain neighborhood disturbance characteristics, including: Determine the spatiotemporal neighborhood of the candidate region; the spatiotemporal neighborhood is the spatial range of multiple gratings taken forward and backward from the grating spatial position as the center, and the same spatial range on the two adjacent lanes to the left and right, and the time range is the continuous time period from the preset time before the state value first falls below the background state threshold to the current time. The feature vector corresponding to each grating-time relationship node in the spatiotemporal neighborhood is determined as the neighborhood perturbation feature.

2. The parking event detection method according to claim 1, characterized in that, The analysis of the neighborhood perturbation features of the candidate region to obtain a neighborhood perturbation feature map includes: The spatiotemporal neighborhood of the candidate region is analyzed in the multi-lane vibration state matrix to obtain neighborhood disturbance characteristics; the spatiotemporal neighborhood includes the upstream region of the lane where the candidate region is located and the corresponding spatial region of the adjacent lane. A neighborhood perturbation feature map is constructed based on the neighborhood perturbation features.

3. The parking event detection method according to claim 1, characterized in that, The preprocessing and global vibration state analysis of the vibration signal data yields a multi-lane vibration state matrix, including: The vibration signal data is preprocessed to obtain preprocessed signal data; The comprehensive vibration state value is obtained by calculating the signal data of each grating on each lane in the preprocessed signal data based on the spatiotemporal window. The comprehensive vibration state values ​​are sorted at each moment according to the lane as the row and the spatial position of the grating as the column to obtain the multi-lane vibration state matrix.

4. The parking event detection method according to claim 3, characterized in that, The calculation of the signal data of each grating on each lane in the preprocessed signal data based on the spatiotemporal window yields a comprehensive vibration state value, including: The average absolute amplitude is obtained by calculating the average value of the preprocessed signal data in each spatiotemporal window; The proportion of signal energy in each grating on each lane within the preset energy ratio in the preprocessed signal data is calculated to obtain the energy proportion. The average absolute amplitude and the energy percentage are normalized and weighted to obtain the comprehensive vibration state value.

5. The parking event detection method according to claim 1, characterized in that, The step of constructing a neighborhood perturbation feature map based on the neighborhood perturbation features includes: The relation nodes of the neighborhood perturbation features are determined as graph nodes; Using the graph nodes as vertices, nodes whose spatial distance is within multiple gratings and whose time difference is within a preset time are connected to form edges to construct a neighborhood perturbation feature graph; the weight of the edge is the reciprocal of the spatiotemporal distance between nodes.

6. The parking event detection method according to claim 5, characterized in that, The step of inputting the candidate region and the neighborhood perturbation feature map into a preset graph neural network to obtain the parking event detection result of the candidate region includes: The candidate region and the neighborhood perturbation feature map are input into a preset graph neural network to obtain the event probability of the candidate region; the preset graph neural network is a model obtained by training offline using historical data; the offline historical data includes "real parking" and "fake parking" event samples, as well as the corresponding feature maps; The parking event detection result is determined based on the event probability.

7. The parking event detection method according to claim 6, characterized in that, Determining the parking event detection result based on the event probability includes: When the probability of the event is greater than or equal to the preset decision threshold, the parking event detection result is determined to be that the lane and grating spatial location represented by the graph node corresponding to the neighborhood disturbance feature map exist from the time the state value first falls below the background state threshold to the current time.

8. A parking event detection device, characterized in that, include: The signal acquisition module is used to acquire vibration signal data from multiple lanes on a highway based on a grating array. The state analysis module is used to preprocess the vibration signal data and perform global vibration situation analysis to obtain a multi-lane vibration state matrix. The region determination module is used to determine candidate regions based on the multi-lane vibration state matrix and analyze the neighborhood disturbance characteristics of the candidate regions to obtain a neighborhood disturbance feature map. The event detection module is used to input the candidate region and the neighborhood perturbation feature map into a preset graph neural network to obtain the parking event detection result of the candidate region; The step of determining the candidate region based on the multi-lane vibration state matrix includes: Obtain the historical state value set for each grating in each lane; The mean and standard deviation of the historical state value set are calculated, and the difference between the mean and the standard deviation is determined as the background state threshold. The grating in the multi-lane vibration state matrix whose state value is less than the background state threshold for a consecutive preset number of cycles is marked as a candidate region for a suspected parking event; The spatiotemporal neighborhood of the candidate region is analyzed in the multi-lane vibration state matrix to obtain neighborhood disturbance characteristics, including: Determine the spatiotemporal neighborhood of the candidate region; the spatiotemporal neighborhood is the spatial range of multiple gratings taken forward and backward from the grating spatial position as the center, and the same spatial range on the two adjacent lanes to the left and right, and the time range is the continuous time period from the preset time before the state value first falls below the background state threshold to the current time. The feature vector corresponding to each grating-time relationship node in the spatiotemporal neighborhood is determined as the neighborhood perturbation feature.

Images