A city parking global trusted perception method based on crowdsourcing data

Abstract

This invention relates to a reliable perception method for urban parking across the entire region based on crowdsourced data, comprising the following steps: S1, constructing a parking space detection model based on deep learning; S2, defining four influencing factors and combining them with a simulation scenario to study and analyze the influence of different combinations of influencing factors on the parking space detection model and the correlation between perception success rate; S3, evaluating the four defined influencing factors, establishing a prediction model for the probability of parking space status perception failure under different external interference factors, and predicting the reliability of the parking space detection model's recognition; S4, building a multi-vehicle perception simulation scenario, analyzing the influence of vehicle fleet size on the parking space detection model, constructing a data freshness model, defining a reliable fusion method for crowdsourced data, and completing the reliable perception of urban roadside parking space status. Compared with existing technologies, this invention has advantages such as improving the accuracy of reliable perception of roadside parking space status, increasing reliability, and strong real-time performance.

Description

Technical Field

[0001] This invention relates to the field of traffic information perception and analysis technology, and in particular to a reliable perception method for urban parking across the entire region based on crowdsourced data. Background Technology

[0002] In recent years, with the improvement of the social and economic level, the number of private cars in my country has been increasing year by year. In some large and medium-sized cities, the phenomenon of parking space shortage is particularly prominent, and disputes caused by parking problems such as illegal parking, illegal parking, obstruction of roads, and cutting in are also common. The surge in the number of motor vehicles and the increasingly prominent supply and demand contradiction of parking spaces are only the surface phenomenon. What is exposed behind it is the inability of traditional parking lot (space) information to be shared, resulting in low utilization rate of vacant parking spaces, causing serious waste of parking space resources, and making the shortage of parking spaces in cities increasingly larger.

[0003] Comprehensively advancing digital transformation is an inevitable requirement for the modernization of governance in megacities. To break down information silos and alleviate urban parking difficulties, it is necessary to accelerate the construction of public parking information platforms. However, real-time status monitoring of roadside parking spaces across a wide area remains an open challenge. Existing solutions based on roadside IoT smart devices offer advantages such as constant data collection speed and high reliability, but deployment across the entire city presents problems including high construction costs, susceptibility to damage, and high maintenance costs. Currently, with the large-scale application of ride-hailing services, sensors such as in-vehicle panoramic surround view and millimeter-wave radar can perceive the status of roadside parking spaces in real time. Furthermore, ride-hailing crowdsourced data offers advantages such as wide coverage, rapid update frequency, and low collection costs. Efficiently utilizing crowdsourced data provides a more feasible approach for real-time status monitoring of roadside parking spaces across a wide area. Reliable perception methods for the status of city-level roadside parking spaces based on crowdsourced data fall within the research scope of data fusion. Factors such as road conditions, driving conditions, and the performance of onboard sensors can cause recognition errors and incomplete information. Assessing the credibility of crowdsourced data is a prerequisite for improving fusion accuracy. Simultaneously, the spatiotemporal randomness of vehicle trajectories in road networks directly affects the reliability and real-time performance of estimation results for different road segments due to the coverage area and update frequency. Addressing the challenges of the high discreteness and variability of massive amounts of intelligent connected vehicle data is crucial for wide-area parking space status estimation. Therefore, there is an urgent need to develop a reliable perception method for roadside parking spaces that considers numerous influencing factors, in order to improve the accuracy and credibility of urban-level roadside parking space perception results based on crowdsourced data. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a reliable perception method for urban parking based on crowdsourced data.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for reliable perception of urban parking across the entire area based on crowdsourced data, comprising the following steps:

[0007] S1. Construct a parking space detection model based on deep learning;

[0008] S2. Define four influencing factors and combine them with simulation scenarios to study the influence of different combinations of influencing factors on the parking space detection model, and analyze the correlation between the influencing factors and the perception success rate of the parking space detection model.

[0009] S3. Use ensemble learning to evaluate the four defined influencing factors, establish a prediction model for the failure probability of parking space status perception under different external interferences, and predict the reliability of the parking space detection model under the interference of different influencing factors.

[0010] S4. Build a multi-vehicle perception simulation scenario, analyze the impact of fleet size on parking space detection model, construct a data freshness model, define a reliable fusion method for crowdsourced data, and complete the reliable perception of the status of roadside parking spaces in the city.

[0011] Furthermore, the simulation scenario sets sensor parameters on the simulated vehicle; the sensor parameters include position, shooting frequency, and image pixels; the simulation scenario extracts macroscopic system data and microscopic vehicle data under different combinations of influencing factors; the crowdsourced data includes sampling time, vehicle self-positioning, and the relative positions of surrounding vehicles.

[0012] Furthermore, S1 specifically includes the following steps:

[0013] S101. Define parking space endpoints p1 and p2 and a parking space entrance line recognition method. If it is the entrance line of a parallel parking space, then t1≤||p1p2||≤t2; if it is a perpendicular or inclined parking space, then t3≤||p1p2||≤t4. The parameters t1, t2, t3, and t4 are determined based on prior knowledge of different parking space entrance line lengths. Furthermore, if one endpoint is missed between the two endpoints and a valid parking space entrance line cannot be formed, recognition is performed on the local image formed by the two endpoints.

[0014] S102. Calculate the complete coordinates of the parking spaces. Each parking space consists of four points: p1, p2, p3, and p4. p1 and p2 are the two endpoints that form the entrance line of the parking space, and p3 and p4 are the other two endpoints of the parking space that are not visible in the image. The coordinate calculation process is as follows:

[0015]

[0016]

[0017] Where, α i and d i The angle and depth of the parking space are respectively: the angle of the perpendicular parking space and the parallel parking space is α1, and the depth is d1 and d2 respectively; the depth of the angled parking space is d3, and the angle is set to α2 and α3 respectively depending on whether it is an acute angle or an obtuse angle.

[0018] S103. Classify parking lot occupancy using regularization to maximize classification performance. Employ perspective transformation to cut and distort parking spaces into uniform 120×46 pixel pixels based on their position in the image. The four boundary points of the parking spaces in the image serve as source points, while the target points are the four vertices of a fixed 120×46 pixel rectangle. Divide the resulting series of labeled images into positive and negative samples, where positive samples represent vacant parking spaces and negative samples represent occupied parking spaces. Rotate the images by 180° to further increase the number of training samples. Finally, use different colored markers to distinguish parking space occupancy.

[0019] Furthermore, S2 specifically refers to:

[0020] S201. Build a simulation scenario, setting road interaction elements including the number of parking spaces, parking space locations, moving vehicles, parked vehicles, and obstacles; set sensor parameters including location, image capture frequency, and image pixels;

[0021] S202. Define and quantify four influencing factors: road width, wear and tear of parking space entrance lines, traffic flow in parking spaces, and camera clarity.

[0022] S203. Define four different application scenarios. In each application scenario, simulate and calculate the failure probability of the parking space detection model under different combinations of influencing factors, and analyze the influence of different influencing factors on the parking space detection model. The four different application scenarios include sunny days with normal lighting, nighttime with weak lighting, sunny days with glaring light, and rainy / foggy days with low visibility.

[0023] S204. Collect microscopic data obtained from simulations under different combinations of influencing factors, and use regression analysis to establish a hierarchical regression model to evaluate the influencing factors; gradually add four influencing factors: road width, wear degree of parking space entrance line, parking space traffic flow and camera clarity, and analyze the impact of the four influencing factors on the perception success rate of the parking space detection model.

[0024] Furthermore, S202 specifically refers to:

[0025] The road width D is defined as the straight-line distance between a vehicle equipped with a surround-view camera and the entrance line of a roadside parking space. The following classification standards are set: D=1 indicates that the detected vehicle is less than 2.5m away from the entrance line of the roadside parking space, that is, the detected vehicle is traveling on or near the trajectory of the center line of the adjacent lane of the parking space; D=2 indicates that the detected vehicle is within the range of 2.5 to 5m away from the entrance line of the roadside parking space, at which point the trajectory of the detected vehicle basically coincides with the left lane line of the adjacent lane of the roadside parking space.

[0026] Define the grading standard for the wear level A at the entrance of the parking space: A=1, the white line of the parking space is clearly visible with only a small amount of wear; A=2, some parking space lines are worn to the point that the white line is no longer visible, causing the parking space recognition model to fail.

[0027] The traffic flow C for a single parking space is defined using a multivariate linear model, as shown in the following formula:

[0028] C i =∑ω r ×k r

[0029] Among them, C i Let r represent the traffic flow at the i-th parking space, r represent the type of obstacle, and ω represent the traffic flow at the ith parking space. r This represents the weighting coefficient corresponding to the obstacle, which is proportional to the area occupied by the individual building, k. r Indicates the number of obstacles;

[0030] The definition of overall traffic flow is as follows:

[0031]

[0032] Where n represents the total number of parking spaces in the model; the classification criteria for overall traffic flow are as follows:

[0033]

[0034] Define camera resolution S to grade and evaluate image quality: when the captured image data is clear, S = 1; otherwise, when the image is blurry, S = 2.

[0035] Furthermore, S3 specifically refers to:

[0036] S301. Using the ensemble learning XGBoost algorithm, classify and predict the results corresponding to any combination of influencing factors, and construct a credibility prediction model for roadside parking space identification for four different application scenarios.

[0037] S302. Perform hierarchical K-fold partitioning on the dataset, cross-validate the reliability prediction model, predict the reliability of the parking space detection model under the interference of different influencing factors, and compare the importance of the influencing factors.

[0038] S303. Based on the credibility prediction results obtained in S302, manually correct the data, retrain the credibility prediction model, and perform cross-validation to improve prediction accuracy and model generalization ability.

[0039] Furthermore, S4 specifically refers to:

[0040] S401. Build a multi-vehicle simulation scenario, continuously increase the number of vehicles detected in the simulation, collect accuracy data of the parking space detection model in the simulation, and analyze the influence of the fleet size on the parking space detection model.

[0041] S402. Construct a data freshness model, set a dynamic freshness time window, define a reliable fusion method for crowdsourced data, and construct a roadside parking space status perception method based on crowdsourced data.

[0042] S403. Simulate a multi-vehicle detection environment to study how a crowdsourced data-based roadside parking space status perception method can improve the accuracy of urban-level roadside parking space status perception.

[0043] Further, S401 specifically involves: setting up n parking spaces in a simulation scenario, placing a detection vehicle, designated Car_1, and sequentially detecting the n parking spaces using a parking space detection model. The accuracy of the detection results is then calculated and recorded as p1. Next, another detection vehicle, designated Car_2, is placed, collecting the same amount of data as Car_1. At this point, each parking space is detected by sensors on both Car_1 and Car_2. The two results containing this parking space are compared and analyzed, and the accuracy of the n parking spaces is calculated and recorded as p2. This process continues, with one additional detection vehicle added each time, repeating the comparison and analysis steps, and recording the accuracy p of the parking space detection model's detection results. i The study analyzed the improvement in accuracy of the parking space detection model after implementing multi-vehicle collaborative detection.

[0044] Furthermore, S402 specifically refers to:

[0045] Constructing a freshness model: Assuming that within a relatively short period of time, the number of vehicles arriving at a certain road segment approximately follows a Poisson distribution with parameter λ, and the number of vehicles leaving approximately follows a Poisson distribution with parameter μ, and ignoring higher-order terms, the freshness function of a single piece of historical sensing data can be obtained using the law of total probability:

[0046]

[0047] Where t is the time since the last historical data point; a freshness threshold is set, and a dynamic freshness time window is set based on the statistically obtained parameters λ and μ. During the reliable data fusion process, only data falling within the time window is considered; a reliable fusion method for crowdsourced data is defined: Let (p i ,t i ) is a piece of sensing data, which represents the data at time t. i The probability that a parking space is not occupied at any given time is p. i A t Let A be the event that a parking space is not occupied at time t. Given a dynamic freshness time window of length t0, event A can be mathematically analyzed. t Conditional probability of occurrence:

[0048]

[0049] Furthermore, S403 specifically involves: setting up n parking spaces in a simulation scenario, using a reliable fusion method for crowdsourced data to construct a roadside parking space status perception method based on crowdsourced data, and performing collaborative processing of multi-vehicle detection results to achieve reliable perception and improved accuracy of urban-level roadside parking spaces.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] I. This invention combines a deep learning-based parking space detection model with traffic simulation to analyze the correlation between different influencing factors and the success rate of the parking space detection model. Furthermore, it uses an ensemble learning method to establish a prediction model for the failure probability of parking space status perception under different external interferences, thereby effectively improving the reliability and accuracy of the parking space status perception model.

[0052] Second, this invention leverages the advantages of crowdsourced data, such as its wide coverage, fast update frequency, and low collection cost, by introducing a freshness index and employing a reliable fusion method for crowdsourced data within a dynamically set time frame. This establishes a reliable perception method for the status of roadside parking spaces at the city level based on crowdsourced data, and further improves the accuracy of the reliable perception method for the status of roadside parking spaces by utilizing the reliable fusion of crowdsourced data. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0054] Figure 2 This is a schematic diagram illustrating the construction of the local image recognition coordinate system according to the present invention;

[0055] Figure 3This is a complete parking space deduction diagram of the three types of parking spaces: vertical, parallel, and inclined, according to the present invention, wherein (a) is a vertical parking space, (b) is a parallel parking space, (c) is an acute-angle inclined parking space, and (d) is an obtuse-angle inclined parking space;

[0056] Figure 4 This is a diagram illustrating the application of the parking space detection model recognition of the present invention;

[0057] Figure 5 This is a schematic diagram illustrating the wear classification of parking space entrances according to the present invention, where A=1.

[0058] Figure 6 This is a schematic diagram illustrating the wear classification of parking space entrances according to the present invention, where A=2.

[0059] Figure 7 This is a schematic diagram illustrating the dynamic time window setting of the present invention;

[0060] Figure 8 This is a schematic diagram of the simulation scene constructed for this invention;

[0061] Figure 9 This is a graph of the data freshness function in this invention;

[0062] Figure 10 This demonstrates the improvement in detection accuracy resulting from the reliable fusion of crowdsourced data in this invention. Detailed Implementation

[0063] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0064] Example

[0065] This example is based on the traffic simulation software PreScan.

[0066] like Figure 1 As shown, a reliable perception method for urban parking based on crowdsourced data is described, and the method includes the following steps:

[0067] S1. Construct a parking space detection model based on deep learning.

[0068] S101. Define parking space endpoints p1 and p2 and a parking space entrance line recognition method. Consider that the two identified parking space endpoints p4 and p2 meet certain confidence requirements, that is, p1 and p2 can form a valid parking space entrance line, and the geometric distance between the two endpoints should meet certain constraints. If it is a parallel parking space entrance line, then t1≤||1p2||≤t2; if it is a perpendicular or inclined parking space, then t3≤||1p2||≤t4, where the parameters t1, t2, t3, and t4 are determined based on prior knowledge of different parking space entrance line lengths. Furthermore, if one endpoint is missed and a valid parking space entrance line cannot be formed, recognition is performed on the local image formed by the two endpoints. Figure 2 As shown, first, establish the origin with the midpoint p(x,y) of p1(x,y) and p2(x,y), and then... Let R be a local coordinate system along the X-axis. Within this system, a rectangular region R is defined with its length along the X-axis as w1 = ||1p2|| + Δw, and its length along the Y-axis as h1. The specific calculation formula is as follows:

[0069]

[0070]

[0071]

[0072] Where Δw is a hyperparameter that controls the width of the rectangular region, and Δh is a hyperparameter that controls the height of the rectangular region.

[0073] S102. Calculate the complete parking space coordinates. Often, video or image data collected from surround-view cameras cannot fully display the parking space; therefore, it is necessary to reason based on geometry and prior knowledge to obtain the complete parking space coordinates. Figure 3 As shown, each parking space consists of four points: p1, p2, p3, and p4. p1 and p2 are the two endpoints that form the entrance line of the parking space, while p3 and p4 are the other two endpoints of the parking space that are not visible in the image. Their coordinates are calculated as follows:

[0074]

[0075]

[0076] Among them, such as Figure 3 As shown in (a), 3(b), 3(c) and 3(d), α i and d iThe angle and depth of the parking space are respectively: the angle of the perpendicular parking space and the parallel parking space is α1, and the depth is d1 and d2 respectively; the depth of the inclined parking space is d3, and the angle is set to α2 and α3 respectively depending on whether it is an acute angle or an obtuse angle.

[0077] S103. Parking lot occupancy is classified. Considering the varying sizes of parking spaces in the panoramic images, regularization is used to maximize classification performance. Perspective transformation is employed to cut and distort parking spaces into uniform 120×46 pixel pixels based on their position in the image. The four boundary points of the parking spaces in the image serve as source points, while the target points are the four vertices of a fixed 120×46 pixel rectangle. The resulting series of labeled images are divided into positive and negative samples, where positive samples represent vacant parking spaces and negative samples represent occupied parking spaces. The images are then rotated 180° to further increase the number of training samples. Finally, different color markers are used to distinguish parking space occupancy. Figure 4 As shown.

[0078] S2. Define four influencing factors and combine them with a simulation scenario to study the influence of different combinations of influencing factors on the parking space detection model, and analyze the correlation between the influencing factors and the perception success rate of the parking space detection model; the simulation scenario sets the sensor parameters on the simulated vehicle; the sensor parameters include position, shooting frequency and image pixels; the simulation scenario extracts macroscopic data of the system and microscopic data of the vehicle under different combinations of influencing factors.

[0079] S201, such as Figure 8 As shown, a simulation scenario was built, and the road-related elements included the number of parking spaces, parking space locations, moving vehicles, parked vehicles, and obstacles. A total of 200 parking spaces were placed on both sides of the five main roads in the simulation, including 100 perpendicular parking spaces and 100 parallel parking spaces. The placement, orientation, and spacing of buildings were randomly set. Furthermore, traffic light waiting times, pedestrian yielding mechanisms, and traffic order at intersections were irrelevant to parking space detection and were not considered in the simulation to simplify the model. At the same time, a row of perpendicular parking spaces can be seen on the left side of the road. To better simulate the real-world scenario, some parking spaces were randomly occupied, and there were pedestrians and delivery electric vehicles between some parking spaces.

[0080] The sensor parameters are set, including location, shooting frequency, and image resolution. A camera is mounted 30m directly above the detection vehicle, with its shooting angle pointing vertically downwards, to collect detection photos. The input image size for the parking space detection model is set to 600 pixels × 600 pixels, so the sensor output image size is also set to 600 pixels × 600 pixels.

[0081] S202 defines and quantifies four influencing factors: road width, wear and tear on parking space entrance lines, traffic flow in parking spaces, and camera clarity.

[0082] The road width D is defined as the straight-line distance between a vehicle equipped with a surround-view camera and the entrance line of a roadside parking space. The following classification standards are set: D=1 indicates that the detected vehicle is less than 2.5m away from the entrance line of the roadside parking space, that is, the detected vehicle is traveling on or near the trajectory of the center line of the adjacent lane of the parking space; D=2 indicates that the detected vehicle is within the range of 2.5 to 5m away from the entrance line of the roadside parking space, at which point the trajectory of the detected vehicle basically coincides with the left lane line of the adjacent lane of the roadside parking space.

[0083] Define the grading criteria for wear and tear level A at parking space entrances: such as Figure 5 As shown, A=1, the white lines of the parking space are clearly visible with only minor wear; Figure 6 As shown, A=2, some parking space lines are worn down to the point that the white lines are no longer visible, causing the parking space recognition model to fail;

[0084] The traffic flow C for a single parking space is defined using a multivariate linear model, as shown in the following formula:

[0085] C e =∑ω r ×k r

[0086] Among them, C i Let r represent the traffic flow at the i-th parking space, r represent the type of obstacle, and ω represent the traffic flow at the ith parking space. r This represents the weighting coefficient corresponding to the obstacle, which is proportional to the area occupied by the individual building, k. r Indicates the number of obstacles;

[0087] In the PreScan simulation experiment, obstacles are divided into three categories: a. pedestrians (0.4m × 0.7m), b. electric vehicles (2.3m × 0.8m), and c. boxes, roadblocks, etc. (1.0m × 1.0m), and assigned weights of 0.1, 0.6, and 0.3 respectively based on their area. Therefore, the traffic flow calculation formula for a single roadside parking space is:

[0088] C i =f(x,y,z) = 0.1 + 0.6 + 0.3

[0089] The definition of overall traffic flow is as follows:

[0090]

[0091] Where n represents the total number of parking spaces in the model, n = 200; when classifying the overall traffic flow, i.e., conducting a macroscopic failure probability study, the classification criteria are as follows:

[0092]

[0093] In subsequent studies of traffic flow at individual parking spaces, i.e., identifying the type of micro-level individual parking spaces, there is no need for hierarchical classification; it can be done through specific C... i The numerical value represents the traffic flow of the i-th parking space. The number of each parking space is extracted from the semantically segmented image. The image is acquired using the semantic segmentation sensor in the simulation software. The number of different obstacles in each parking space can be directly output for traffic flow calculation.

[0094] Define camera resolution S to grade and evaluate image quality: when the captured image data is clear, S = 1; otherwise, when the image is blurry, S = 2.

[0095] S203. Define four different application scenarios. In each application scenario, simulate and calculate the failure probability of the parking space detection model under different combinations of influencing factors, and analyze the influence of different influencing factors on the parking space detection model. The four different application scenarios include sunny days with normal lighting, nighttime with weak lighting, sunny days with glaring light, and rainy / foggy days with low visibility. Secondly, according to the definition and classification of influencing factors, there are 2^4 = 16 different combinations of influencing factors and their corresponding failure probabilities in each application scenario. Therefore, there are a total of 64 macroscopic data points for the four different application scenarios.

[0096] S204. Collect microscopic data obtained from simulations under different combinations of influencing factors. In the application scenario of a sunny day with normal lighting, randomly extract frames from the overhead driving video obtained by image stitching from the four cameras in the PreScan simulation experiment, i.e., use the video results as images, sampling a total of 518 frames. Then, statistically analyze the data of four influencing factors corresponding to each frame image: road width, wear condition of parking space entrance line, complexity, and camera clarity. Determine whether the roadside parking space is accurately detected. Summarize the above data to obtain the dataset for this module. Use regression analysis to establish a hierarchical regression model to evaluate the influencing factors. Gradually add four influencing factors: road width, wear condition of parking space entrance line, parking space traffic flow, and camera clarity, and analyze the impact of the four influencing factors on the perception success rate of the parking space detection model.

[0097] S3. Using ensemble learning methods, evaluate the four defined influencing factors, establish a prediction model for the failure probability of parking space status perception under different external interferences, and predict the reliability of the parking space detection model under the interference of different influencing factors.

[0098] S301. Using the ensemble learning XGBoost algorithm, the results corresponding to any combination of influencing factors are classified and predicted, and a credibility prediction model for roadside parking space identification is constructed for four different application scenarios.

[0099] S302. Divide the 518 frames of data collected in S204 into training and test sets in a 3:1 ratio. Perform hierarchical K-fold partitioning on the dataset to ensure the stability and reliability of the final model. Cross-validate the credibility prediction model to predict the credibility of the parking space detection model under the interference of different influencing factors and compare the importance of the influencing factors.

[0100] S303. Based on the credibility prediction results obtained in S302, manually correct the data, retrain the credibility prediction model, and perform cross-validation to improve prediction accuracy and model generalization ability.

[0101] S4. Build a multi-vehicle perception simulation scenario, analyze the influence of fleet size on the parking space detection model, construct a data freshness model, define a reliable fusion method for crowdsourced data, and complete the reliable perception of the status of urban roadside parking spaces; the crowdsourced data includes sampling time, vehicle self-positioning, and the relative position of surrounding vehicles.

[0102] S401. Build a multi-vehicle simulation scenario, continuously increase the number of vehicles detected in the simulation, collect accuracy data of the parking space detection model in the simulation, and analyze the influence of vehicle fleet size on the parking space detection model.

[0103] In a simulation scenario, n parking spaces are set up, and a detection vehicle, designated Car_1, is placed. The parking space detection model sequentially detects the n parking spaces in the scenario, and the accuracy of the detection results is recorded as p1. Next, another detection vehicle, designated Car_2, is placed, collecting the same amount of data as Car_1. At this point, each parking space is detected by the sensors on both Car_1 and Car_2. The two results containing this parking space are compared and analyzed, and the accuracy of the n parking spaces is recorded as p2. This process is repeated, adding one detection vehicle each time, and the comparison and analysis steps are repeated, recording the accuracy p of the parking space detection model. i The analysis shows the improvement in accuracy of the parking space detection model after multi-vehicle collaborative detection; in this embodiment, the number of parking spaces is set to 65.

[0104] S402. Construct a data freshness model, set a dynamic freshness time window, define a reliable fusion method for crowdsourced data, and construct a roadside parking space status perception method based on crowdsourced data.

[0105] Constructing a freshness model: Assuming that within a relatively short period of time, the number of vehicles arriving at a certain road segment approximately follows a Poisson distribution with parameter λ, and the number of vehicles leaving approximately follows a Poisson distribution with parameter μ, and ignoring higher-order terms, the freshness function of a single piece of historical sensing data can be obtained using the law of total probability:

[0106]

[0107] Where t is the time from the present to the present for this single historical sensing data point, and μ is set to 0.2, changing the value of λ can yield the following results: Figure 9 The image shown;

[0108] like Figure 6 As shown, a freshness threshold of 0.2 is set, and a dynamic freshness time window is set based on the statistically obtained parameters λ and μ. During the reliable data fusion process, only data falling within the time window is considered. A reliable fusion method for crowdsourced data is defined: Let (p...) i ,t i ) is a piece of sensing data, which represents the data at time t. i The probability that a parking space is not occupied at any given time is p. i A t Let A be the event that a parking space is not occupied at time t. Given a dynamic freshness time window of length t0, event A can be mathematically analyzed. t Conditional probability of occurrence:

[0109]

[0110] S403. Simulate a multi-vehicle detection environment to study how a crowdsourced data-based roadside parking space status perception method can improve the accuracy of urban-level roadside parking space status perception.

[0111] In a simulation scenario, n parking spaces are set up. A roadside parking space status perception method based on crowdsourced data is constructed using a trusted fusion method for crowdsourced data. This method enables collaborative processing of multi-vehicle detection results. The trusted fusion method for crowdsourced data improves detection accuracy as follows: Figure 10 As shown, although the amount of crowdsourced data in this experiment is relatively small, it still reflects that the above-mentioned crowdsourced data-based roadside parking space status perception method can effectively improve the reliability and accuracy of urban-level roadside parking space perception.

[0112] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A reliable perception method for urban parking across the entire region based on crowdsourced data, characterized in that, The method includes the following steps: S1. Construct a parking space detection model based on deep learning; S2. Define four influencing factors and combine them with simulation scenarios to study the influence of different combinations of influencing factors on the parking space detection model, and analyze the correlation between the influencing factors and the perception success rate of the parking space detection model. S3. Use ensemble learning to evaluate the four defined influencing factors, establish a prediction model for the failure probability of parking space status perception under different external interferences, and predict the reliability of the parking space detection model under the interference of different influencing factors. S4. Build a multi-vehicle perception simulation scenario, analyze the impact of fleet size on parking space detection model, construct a data freshness model, define a reliable fusion method for crowdsourced data, and complete the reliable perception of the status of urban roadside parking spaces. S1 specifically includes the following steps: S101, Define parking space endpoints and And the method for recognizing parking space entrance lines; if it is the entrance line of a parallel parking space, then... If it is a perpendicular parking space or an angled parking space, then , where parameters , , , It is determined based on prior knowledge of the length of different parking space entrance lines; and if there is an endpoint that is missed between the two endpoints and a valid parking space entrance line cannot be formed, it is completed by recognizing the local image formed by the two endpoints. S102. Calculate the complete coordinates of each parking space, where each parking space is... , , and It consists of four points, among which and These are the two endpoints that form the entrance line to the parking space. and The coordinates of the other two endpoints of the parking space, which are not visible in the image, are calculated as follows: in, and These represent the angle and depth of the parking space, respectively; the angles of perpendicular and parallel parking spaces are both... The depths are respectively and The depth of the angled parking spaces is... The angle is set according to whether it is acute or obtuse. and ; S103. Classify parking lot occupancy, using regularization to maximize classification performance. Employ perspective transformation technology to cut and distort parking spaces into uniform 120×46 pixel pixels based on their position in the image. The four boundary points of the parking spaces in the image are used as source points, while the target points are the four vertices of a fixed 120×46 pixel rectangle. Divide the resulting series of labeled images into positive and negative samples, where positive samples represent vacant parking spaces and negative samples represent occupied parking spaces. Rotate the images by 180° to further increase the number of training samples. Finally, use different colored markers to distinguish parking space occupancy. Specifically, S2 is: S201. Build a simulation scenario, setting road interaction elements including the number of parking spaces, parking space locations, moving vehicles, parked vehicles, and obstacles; set sensor parameters including location, image capture frequency, and image pixels; S202. Define and quantify four influencing factors: road width, wear and tear of parking space entrance lines, traffic flow in parking spaces, and camera clarity. S203. Define four different application scenarios. In each application scenario, simulate and calculate the failure probability of the parking space detection model under different combinations of influencing factors, and analyze the influence of different influencing factors on the parking space detection model. The four different application scenarios include sunny days with normal lighting, nighttime with weak lighting, sunny days with glaring light, and rainy / foggy days with low visibility. S204. Collect microscopic data obtained from simulations under different combinations of influencing factors, and use regression analysis to establish a hierarchical regression model to evaluate the influencing factors; gradually add four influencing factors: road width, wear degree of parking space entrance line, parking space traffic flow and camera clarity, and analyze the impact of the four influencing factors on the perception success rate of the parking space detection model. Specifically, S3 refers to: S301. Using the ensemble learning XGBoost algorithm, classify and predict the results corresponding to any combination of influencing factors, and construct a credibility prediction model for roadside parking space identification for four different application scenarios. S302. Perform hierarchical K-fold partitioning on the dataset, cross-validate the reliability prediction model, predict the reliability of the parking space detection model under the interference of different influencing factors, and compare the importance of the influencing factors. S303. Based on the credibility prediction results obtained in S302, manually correct the data, retrain the credibility prediction model, and perform cross-validation to improve prediction accuracy and model generalization ability. Specifically, S4 is: S401. Build a multi-vehicle simulation scenario, continuously increase the number of vehicles detected in the simulation, collect accuracy data of the parking space detection model in the simulation, and analyze the influence of the fleet size on the parking space detection model. S402. Construct a data freshness model, set a dynamic freshness time window, define a reliable fusion method for crowdsourced data, and construct a roadside parking space status perception method based on crowdsourced data. S403. Simulate a multi-vehicle detection environment to study how a crowdsourced data-based roadside parking space status perception method can improve the accuracy of urban-level roadside parking space status perception.

2. The urban parking full-domain reliable perception method based on crowdsourced data according to claim 1, characterized in that, The simulation scenario sets sensor parameters on the simulated vehicle; the sensor parameters include position, shooting frequency, and image pixels; the simulation scenario extracts macroscopic system data and microscopic vehicle data under different combinations of influencing factors; the crowdsourced data includes sampling time, vehicle self-positioning, and the relative positions of surrounding vehicles.

3. The urban parking full-domain reliable perception method based on crowdsourced data according to claim 1, characterized in that, Specifically, S202 is as follows: The road width D is defined as the straight-line distance between a vehicle equipped with a surround-view camera and the entrance line of a roadside parking space. The following classification standards are set: D=1 indicates that the detected vehicle is less than 2.5m away from the entrance line of the roadside parking space, that is, the detected vehicle is traveling on or near the trajectory of the center line of the adjacent lane of the parking space; D=2 indicates that the detected vehicle is within the range of 2.5~5m away from the entrance line of the roadside parking space, at which point the trajectory of the detected vehicle basically coincides with the left lane line of the adjacent lane of the roadside parking space. The grading standard for the wear level A at the entrance of a parking space is defined as follows: A=1, the white line of the parking space is clearly visible, with only a small amount of wear; A=2, some parking space lines are worn down to the point that the white lines are no longer visible, causing the parking space recognition model to fail; The traffic flow C for a single parking space is defined using a multivariate linear model, as shown in the following formula: in, Let r represent the traffic flow at the i-th parking space, and r represent the type of obstacle. This indicates the weighting coefficient corresponding to the obstacle, which is proportional to the area occupied by the individual building. Indicates the number of obstacles; The definition of overall traffic flow is as follows: Where n represents the total number of parking spaces in the model; the classification criteria for overall traffic flow are as follows: ； Define camera resolution S to grade and evaluate image quality: S=1 when the captured image data is clear; otherwise, S=2 when the image is blurry.

4. The urban parking full-area reliable perception method based on crowdsourced data according to claim 1, characterized in that, S401 specifically involves: setting up n parking spaces in a simulation scenario, placing a detection vehicle (Car_1), sequentially detecting the n parking spaces using a parking space detection model, and statistically analyzing the accuracy of the detection results, denoted as . ; Following this, another detection vehicle, designated Car_2, is placed, collecting the same amount of data as Car_1. At this point, each parking space is detected by sensors on both Car_1 and Car_2. The two results containing this parking space are compared and analyzed. Then, the accuracy rate for n parking spaces is statistically analyzed and denoted as . Similarly, one additional testing vehicle is added to each experiment, the comparative analysis steps are repeated, and the accuracy of the parking space detection model's detection results is recorded. The study analyzed the improvement in accuracy of the parking space detection model after implementing multi-vehicle collaborative detection.

5. The urban parking full-domain reliable perception method based on crowdsourced data according to claim 1, characterized in that, Specifically, S402 is as follows: Constructing a freshness model: Assume that within a relatively short period of time, the number of vehicles arriving on a certain road segment approximately follows the parameter... λ The number of vehicles leaving follows a Poisson distribution with parameters approximately equal to 1. μ The Poisson distribution, if we ignore higher-order terms, can be used with the law of total probability to obtain the freshness function of a single piece of historical sensory data: in t The time elapsed since the present is considered for this single historical data entry; Set a freshness threshold and base it on the parameters obtained from statistics. λ and μ Set a dynamic freshness time window, and only consider data falling within the time window during the trusted data fusion process; define a trusted fusion method for crowdsourced data: [describe...] For a single piece of sensing data, it represents in The probability that a parking space is not occupied at any given time is , for t The dynamic freshness time window is the length of an event where a parking space is not occupied at any given time. Then the event can be mathematically analyzed. Conditional probability of occurrence: 。 6. The urban parking full-domain reliable perception method based on crowdsourced data according to claim 1, characterized in that, Specifically, S403 involves setting up n parking spaces in a simulation scenario, using a reliable fusion method for crowdsourced data to construct a roadside parking space status perception method based on crowdsourced data, and performing collaborative processing of multi-vehicle detection results to achieve reliable perception and improved accuracy of urban-level roadside parking spaces.

Images