Method for detecting group vehicle cooperative illegal parking based on vehicle-road multi-element reliability modeling

Abstract

The present application relates to the field of intelligent transportation system and vehicle cooperative perception technology, and particularly relates to a group vehicle cooperative illegal parking detection method based on vehicle-road multi-element reliability modeling, comprising: S1: obtaining sensor data by a plurality of sensing vehicles and generating a space-time perception map for the sensing vehicles; S2: calculating a road environment reliability score of a target vehicle detected by the sensing vehicles based on the space-time perception map of the sensing vehicles; S3: obtaining local detection results of all sensing vehicles through a server; S4: constructing a multi-vehicle-multi-target detection matrix and a road environment reliability score matrix; S5: constructing a probability generation model; S6: performing joint inference on the probability generation model through an expectation maximization algorithm, and outputting a judgment result of an illegal parking state of all target vehicles. The method of the present application can improve the accuracy, recall rate and overall robustness of illegal parking detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems and vehicle cooperative perception technology, specifically to a group vehicle cooperative illegal parking detection method based on vehicle-road multi-dimensional reliability modeling. Background Technology

[0002] With the continuous surge in motor vehicle ownership, illegal parking has become increasingly frequent. It not only occupies limited road resources, exacerbating traffic congestion and reducing traffic efficiency, but also obstructs the view of other drivers and pedestrians, creating serious traffic safety hazards. Currently, the mainstream detection methods mainly rely on fixed roadside cameras and manual patrols. While fixed cameras can achieve a certain degree of automation, their deployment is limited by infrastructure conditions such as power supply and network, and they have significant blind spots, failing to cover the entire road network. Furthermore, the high costs of equipment procurement, installation, debugging, and long-term maintenance also restrict their large-scale adoption. Manual patrols are inefficient and constrained by human resources and working hours, making it impossible to conduct effective, large-scale, and round-the-clock monitoring. Both of these traditional methods are insufficient to meet the urgent need for real-time, accurate, and comprehensive detection of illegal parking in a large-scale, highly dynamic urban road network environment.

[0003] To overcome the limitations of traditional fixed sensing models, and thanks to the rapid development of IoT, V2X, and AI technologies, a new detection paradigm—mobile vehicle-based collaborative sensing—has emerged. The core idea of ​​this technology is to transform a large number of ordinary vehicles on the road (such as taxis, buses, and ride-hailing vehicles) into mobile sensing nodes, leveraging their extensive and natural spatiotemporal coverage to compensate for the shortcomings of fixed sensing. These vehicles, equipped with various onboard sensors (such as forward-facing cameras, GPS, and IMU inertial measurement units), continuously collect data and perform local computation on the road environment (especially roadside areas) during their journey. By deploying lightweight deep learning models (such as YOLO and SSD object detection algorithms) on onboard computing units or mobile edge devices, video streams can be analyzed in real time to initially identify and locate potentially illegally parked vehicles, generating local detection events containing information such as time, location, and image evidence. Aggregating the detection results from multiple sensing vehicles can further improve the detection range and reliability. This model, known as collaborative illegal parking detection, has inherent advantages such as wide coverage, flexible deployment, and low cost.

[0004] However, existing, especially early, collaborative vehicle-to-vehicle parking violation detection schemes have revealed a fundamental technical flaw in large-scale deployments. Many early studies and system designs tended to employ a relatively simple and direct multi-source information aggregation strategy, most notably the "majority voting" mechanism, which simply assumes that a target reported simultaneously by a majority of vehicles is a genuine violation. However, in the complex urban traffic environment, the reliability of detection results varies significantly among different vehicles due to differences in sensor performance, installation orientation, and instantaneous environmental factors (such as obstruction by other vehicles, changes in lighting, and weather conditions). Blindly aggregating detection reports from all vehicles not only fails to improve accuracy but may also lead to the paradoxical phenomenon that "the more vehicles involved, the lower the overall detection accuracy."

[0005] Therefore, how to effectively evaluate and integrate detection results of different reliability is the core technical challenge in achieving high-precision collaborative perception among vehicles. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a group vehicle cooperative illegal parking detection method based on vehicle-road multi-dimensional reliability modeling. This method aims to solve the core problem that "the more vehicles involved, the lower the overall detection accuracy" caused by ignoring the reliability differences between sensing nodes. By introducing a systematic reliability quantification framework and intelligent fusion algorithm, this invention achieves efficient utilization of multi-source heterogeneous detection data, thereby significantly improving the accuracy, recall rate, and overall robustness of illegal parking detection.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A group-vehicle cooperative illegal parking detection method based on vehicle-road multivariate reliability modeling includes:

[0009] S1: Acquire sensor data through several sensing vehicles, and generate a spatiotemporal perception map for each sensing vehicle that includes the target vehicles it detects based on the sensor data.

[0010] S2: Calculate the road environment reliability score of the detected target vehicle based on the spatiotemporal perception map of the perceived vehicle.

[0011] S3: Obtain local detection results of all perceived vehicles, including the road environment reliability scores of the target vehicles they detect, through the server;

[0012] S4: Construct a multi-vehicle-multi-target detection matrix and a road environment reliability score matrix based on the local detection results of all perceived vehicles through the server;

[0013] S5: Construct a probabilistic generative model based on the detection matrix of multiple vehicles and multiple targets and the road environment reliability score matrix;

[0014] S6: Perform joint inference on the probabilistic generation model using the expectation-maximization algorithm to output the illegal parking status judgment results of all target vehicles and the sensing reliability evaluation results of all sensing vehicles.

[0015] Preferably, in step S1, a spatiotemporal perception map is generated for the sensing vehicle through the following steps:

[0016] S101: Acquire sensor data, including in-vehicle video, camera parameters, and its own pose data, by sensing the vehicle.

[0017] S102: Determine the initial sensing sector area based on the camera metadata of the sensing vehicle;

[0018] S103: Perform vehicle target detection and tracking on each frame of the in-vehicle video of the sensing vehicle, and generate detection boxes for surrounding vehicles.

[0019] S104: Based on the changes in the detection boxes of surrounding vehicles, combined with camera parameters and its own pose data, determine its motion state, and take the surrounding vehicles that are stationary as the target vehicles.

[0020] S105: Calculate the real-time occlusion angle range caused by the geometric contours of surrounding vehicles, subtract the real-time occlusion angle ranges of all detected vehicles from the initial sensing sector area of ​​the sensing vehicle, and obtain the effective sensing sector area of ​​the sensing vehicle at the current moment.

[0021] S106: Accumulate the effective sensing sector area of ​​the sensing vehicle at each moment along the time dimension to form the spatiotemporal sensing map of the sensing vehicle.

[0022] Preferably, in step S105, the effective sensing sector area of ​​the sensing vehicle is generated through the following steps:

[0023] S1051: In At any moment, for sensing vehicles Every surrounding vehicle detected Using camera geometry to capture surrounding vehicles The four corner points of the detection frame are projected onto the ground coordinate system, and the azimuth angles of the lines connecting the four corner points of the detection frame to the camera center are calculated. Among them, the camera's internal and external parameters of the sensing vehicle and the detection box information of the surrounding vehicles are fused, and the four corner points of the detection box are projected onto the ground coordinate system (bird's-eye view coordinate system) through inverse perspective transformation.

[0024] S1052: Regarding surrounding vehicles The minimum and maximum values ​​of the azimuth angle are taken to form the range of the occlusion angle caused by it. ;

[0025] The formula is described as follows:

[0026] ;

[0027] In the formula: Indicates surrounding vehicles The resulting obstruction area; Indicates surrounding vehicles No. The azimuth angle of the line connecting the corner point and the center of the camera;

[0028] S1053: All The total occlusion range is obtained by merging the occlusion areas caused by surrounding vehicles. ;

[0029] The formula is described as follows:

[0030] ;

[0031] In the formula: Indicates surrounding vehicles The resulting obstruction area; Indicates the total occlusion area;

[0032] S1054: From sensing vehicles Subtract the total occlusion area from the initial perceived sector region The vehicle was obtained Effective sensing sector area at the current moment ;

[0033] The formula is described as follows:

[0034] ;

[0035] In the formula: Indicates the perception of vehicles The initial perceptual sector area; To sense vehicles Effective perception of the sector area.

[0036] Preferably, in step S2, the road environment reliability score of the target vehicle is calculated through the following steps:

[0037] S201: Calculate the target vehicle The actual rectangular area corresponding to the detection box and the corresponding sensing vehicle The ratio of the intersection area between spatiotemporally perceived maps to their own area yields a base score calculated based on environmental occlusion. ;

[0038] The formula is described as follows:

[0039] ;

[0040] In the formula: Indicates the perception of vehicles For the target vehicle The base score based on ambient occlusion; Indicates the target vehicle The rectangular area it occupies; Indicates the perception of vehicles Spatiotemporal perception map; express and The area of ​​their intersection; express The area; This is the scaling factor;

[0041] S202: Acquiring Sensing Vehicles For the target vehicle The average confidence score output during target tracking ;

[0042] S203: The road environment reliability score is obtained by linearly weighting the base score calculated based on environmental shading and the average confidence score;

[0043] The formula is described as follows:

[0044] ;

[0045] In the formula: Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the perception of vehicles For the target vehicle The base score based on ambient occlusion; Indicates the perception of vehicles For the target vehicle The average confidence score during target tracking; and These represent adjustable weights that balance the contributions of environmental quality and algorithm confidence, respectively.

[0046] Preferably, in step S3, the local detection results of the perceived vehicle include standard information such as the target ID, type, location coordinates and timestamp of all target vehicles, as well as the road environment reliability score and the basic score calculated based on environmental occlusion of all target vehicles.

[0047] Preferably, in step S4, the multi-vehicle-multi-target detection matrix and reliability matrix are constructed through the following steps:

[0048] S401: Use a vehicle re-identification algorithm to match the same target vehicle detected by different sensing vehicles;

[0049] S402: Construct the detection matrix Among them, the detection matrix In the detection matrix, rows represent different target vehicles, and columns represent different sensing vehicles; elements in Indicates the perception of vehicles The report detected the target vehicle. ,element This indicates that nothing was detected.

[0050] S403: Constructing a Road Environment Reliability Score Matrix :like Then the road environment reliability score matrix elements in ;like Then the road environment reliability score matrix elements in .

[0051] Preferably, in step S5, the inherent reliability parameters of the perceived vehicle, namely the true positive rate and the true negative rate, are introduced, and a probabilistic generation model that integrates the environment and the inherent reliability of the vehicle is established by combining the detection matrix of multiple vehicles and multiple targets and the road environment reliability score matrix.

[0052] The formula for the probabilistic generative model is described as follows:

[0053] ;

[0054] In the formula: express and The probability of taking different values; Indicates the perception of vehicles For the target vehicle The test results; For the target vehicle The actual parking status; Indicates the perception of vehicles The true positive rate of the test; Indicates the perception of vehicles The true negative rate of the test; Indicates the perception of vehicles For the target vehicle The road environment reliability score.

[0055] Preferably, in step S6, the processing step of performing joint inference on the probability generation model using the expectation-maximization algorithm includes:

[0056] S601: Initialize the inherent reliability parameters of the sensing vehicle, namely the true positive rate and the true negative rate;

[0057] S602: In the E-step of the expectation-maximization algorithm, fix the inherent reliability parameters of all currently perceived vehicles, and calculate the posterior probability of each target vehicle having engaged in illegal parking behavior based on Bayes' theorem. ;

[0058] S603: In the M-step of the expectation-maximization algorithm, the current posterior probability estimate is fixed, and the inherent reliability parameters of all perceived vehicles are updated by weighted counting, using the complement of the road environment reliability score as the weight.

[0059] S604: Alternately execute steps S502 and S503 until the probabilistic generation model converges or the maximum number of iterations is reached;

[0060] S605: Outputs high-confidence determination results of illegal parking status for all target vehicles, and evaluation results of the sensing reliability of all sensing vehicles; among which, the determination results of illegal parking status for target vehicles... By analyzing the posterior probability Binarization is performed using a threshold; the evaluation results of the vehicle sensing reliability include the detection true positive rate. and detection of true negative rate The estimated value.

[0061] Preferably, in step S602, the inherent reliability parameters of all currently sensing vehicles are fixed. Based on Bayes' theorem and prior probability, the posterior probability of the target vehicle illegally parking is calculated.

[0062] The formula is described as follows:

[0063] ;

[0064] In the formula: Indicates the target vehicle The posterior probability of illegal parking; Let represent the prior global violation rate parameter, whose initial value is a random variable and follows a Beta prior distribution; and These represent the likelihood values ​​under conditions of illegal parking and no illegal parking, respectively.

[0065] Preferably, in step S603, the specific steps for updating the inherent reliability parameters of the perceived vehicle include:

[0066] S6031: Update prior global illegal parking rate parameters ;

[0067] The public notice indicates that:

[0068] ;

[0069] In the formula: This represents the prior global illegal parking rate parameter; and These are hyperparameters; Indicates the total number of target vehicles; Indicates the first step;

[0070] S6032: Update True Yang Rate ;

[0071] The formula is described as follows:

[0072] ;

[0073] In the formula: Indicates the perception of vehicles The true positive rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the target vehicle The posterior probability of illegal parking; To sense vehicles For the target vehicle The test results;

[0074] S6033: Update True Negative Rate ;

[0075] The formula is described as follows:

[0076] ;

[0077] In the formula: Indicates the perception of vehicles The true negative rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the target vehicle The posterior probability of illegal parking; To sense vehicles For the target vehicle The test results.

[0078] Compared with existing technologies, the group vehicle cooperative illegal parking detection method based on vehicle-road multivariate reliability modeling in this invention has the following advantages:

[0079] This invention first collects sensor data from sensing vehicles, then dynamically evaluates environmental occlusion and couples it with the sensing vehicles' own detection characteristics to generate local detection results including a road environment reliability score. The server-side constructs a probabilistic generation model, aggregates the local detection results uploaded by each sensing vehicle, and simultaneously optimizes the illegal parking event judgment and sensor reliability parameters through iterative reasoning, ultimately achieving reliable illegal parking vehicle detection. Traditional methods often ignore the differences in sensing reliability among different vehicles in different environments, leading to the amplification of invalid or erroneous information during the aggregation process. This invention, however, achieves intelligent filtering and weighted fusion of multi-source sensing data by dynamically evaluating the sensing capabilities of each vehicle in a specific time and space and fusing its inherent detection reliability. By introducing a dual reliability mechanism of environment and vehicle inherent reliability, it solves the practical problem in collaborative sensing of multiple vehicles where the more vehicles involved, the worse the detection results become. This significantly improves the accuracy and robustness of illegal parking detection.

[0080] This invention integrates vehicle-mounted camera parameters, the vehicle's own pose data, and real-time detected geometric information of surrounding vehicles. Through inverse perspective transformation and occlusion analysis, it maps vehicle detection results in image space to a ground (bird's-eye view) coordinate system. Based on the motion state of surrounding vehicles, it calculates the effective perception sector in real time, dynamically estimating the effective perception range of the vehicle in complex traffic environments and constructing a real-time perception map. This method does not rely on high-precision maps or fixed sensor deployments. It eliminates the need for high-precision prior maps, achieving accurate quantification of line-of-sight occlusion and dynamic modeling of line-of-sight occlusion in complex traffic environments using only vehicle-mounted cameras and positioning data. This significantly improves the system's environmental adaptability and perception robustness in complex traffic flows (such as congestion, intersections, and curves), laying the foundation for subsequent objective assessment of road environment reliability.

[0081] This invention proposes a dual-reliability probabilistic inference model based on the Expectation-Maximization (EM) algorithm. By decomposing road environment reliability into external factors influenced by environmental dynamics and internal intrinsic factors reflecting the vehicle's own perception capabilities, an exponentially weighted likelihood function, weighted by road environment reliability, is designed and coupled with each vehicle's inherent true positive / true negative ratio. This innovatively unifies two types of uncertainty sources into a single probabilistic generation model and iteratively optimizes the simultaneous inference of illegal parking status and vehicle reliability parameters. The advantages of this design lie in its ability to simultaneously handle external dynamic traffic occlusion and internal unknown equipment bias. The EM algorithm enables joint optimization of violation status inference and reliability parameter learning, iteratively optimizing high-confidence violation event judgments and quantitative assessments of the sensing reliability of each vehicle. This not only significantly improves the accuracy and robustness of the fusion results but also outputs a quantitative assessment of the long-term performance of each sensing vehicle.

[0082] This invention can further combine edge-cloud collaboration technology to deploy lightweight reliability assessment on the vehicle side and run complex fusion inference in the cloud, so as to realize large-scale, low-cost and highly reliable detection of illegal parking on urban roads, which greatly protects urban road traffic safety. It also takes advantage of the group vehicle perception without the need to deploy additional dedicated equipment, making it easy to deploy and apply in actual urban areas. Attached Figure Description

[0083] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0084] Figure 1 This is the overall system architecture diagram of a group vehicle cooperative illegal parking detection method based on vehicle-road multi-dimensional reliability modeling.

[0085] Figure 2 This is a schematic diagram of the dynamic perception map construction principle; (a) constructing a perception map based on camera metadata and geometric relationships with nearby vehicles, and (b) projecting image key points onto the ground coordinate system.

[0086] Figure 3 This is a performance comparison curve between the method of this invention and four baseline methods under different traffic densities (congested, normal, and open).

[0087] Figure 4 This is a performance comparison curve between the method of this invention and four baseline methods under different vehicle participation rates.

[0088] Figure 5 This is a robustness comparison curve of the performance of each method under different daytime weather conditions (sunny, rainy, and foggy).

[0089] Figure 6This is a robustness comparison curve of the performance of each method under different weather conditions at night (sunny, rainy, and foggy).

[0090] Figure 7 This is a robustness comparison curve of the performance of various methods under different lane positions (curve, near lane, middle lane, far lane) provided by the embodiments of the present invention.

[0091] Figure 8 This is a schematic diagram of the actual deployment of the prototype system. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0093] The following detailed explanation illustrates the specific implementation methods:

[0094] Example:

[0095] This embodiment discloses a group vehicle cooperative illegal parking detection method based on vehicle-road multivariate reliability modeling.

[0096] like Figure 1 As shown, the group vehicle cooperative illegal parking detection method based on vehicle-road multivariate reliability modeling includes:

[0097] S1: Acquire sensor data through several sensing vehicles, and generate a spatiotemporal perception map for each sensing vehicle that includes the target vehicles it detects based on the sensor data.

[0098] In this embodiment, each sensing vehicle (such as a taxi or ride-hailing vehicle) uses forward-facing cameras, GPS / IMU modules, and other sensing devices to continuously collect data such as video of the road ahead, its own geographical location, speed, and attitude angles during operation. Data collection can be achieved by deploying a smartphone app integrated with the above functions on the sensing vehicle.

[0099] S2: Calculate the road environment reliability score of the detected target vehicle based on the spatiotemporal perception map of the perceived vehicle.

[0100] S3: Obtain local detection results of all perceived vehicles, including the road environment reliability scores of the target vehicles they detect, through the server;

[0101] S4: Construct a multi-vehicle-multi-target detection matrix and a road environment reliability score matrix based on the local detection results of all perceived vehicles through the server;

[0102] S5: Construct a probabilistic generative model based on the detection matrix of multiple vehicles and multiple targets and the road environment reliability score matrix;

[0103] S6: The probability generation model is jointly inferred using the expectation-maximization algorithm to output the illegal parking status judgment results for all target vehicles and the sensor reliability evaluation results for all sensing vehicles. The reliability parameters of all sensing vehicles can be used not only for this inference but also as historical knowledge for subsequent batch detection tasks, continuously optimizing system performance.

[0104] This invention first collects sensor data from sensing vehicles, then dynamically evaluates environmental occlusion and couples it with the sensing vehicles' own detection characteristics to generate local detection results including a road environment reliability score. The server-side constructs a probabilistic generation model, aggregates the local detection results uploaded by each sensing vehicle, and simultaneously optimizes the illegal parking event judgment and sensor reliability parameters through iterative reasoning, ultimately achieving reliable illegal parking vehicle detection. Traditional methods often ignore the differences in sensing reliability among different vehicles in different environments, leading to the amplification of invalid or erroneous information during the aggregation process. This invention, however, achieves intelligent filtering and weighted fusion of multi-source sensing data by dynamically evaluating the sensing capabilities of each vehicle in a specific time and space and fusing its inherent detection reliability. By introducing a dual reliability mechanism of environment and vehicle inherent reliability, it solves the practical problem in collaborative sensing of multiple vehicles where the more vehicles involved, the worse the detection results become. This significantly improves the accuracy and robustness of illegal parking detection.

[0105] This invention can further combine edge-cloud collaboration technology to deploy lightweight reliability assessment on the vehicle side and run complex fusion inference in the cloud, so as to realize large-scale, low-cost and highly reliable detection of illegal parking on urban roads, which greatly protects urban road traffic safety. It also takes advantage of the group vehicle perception without the need to deploy additional dedicated equipment, making it easy to deploy and apply in actual urban areas.

[0106] To better illustrate the technical solution of the present invention, this embodiment is described in the following parts.

[0107] I. Spatiotemporal Perception Map

[0108] When a vehicle is traveling on the road, the sensing range of its onboard camera is easily affected by obstacles. The scene captured by the onboard video is very complex, with vehicles occluding each other and the sensing range changing constantly. To address this, this invention constructs a sensing map to dynamically simulate line-of-sight occlusion from the sensing vehicle to the target.

[0109] Specifically, a spatiotemporal perception map is generated for the sensing vehicle through the following steps:

[0110] S101: Acquire sensor data, including in-vehicle video, camera parameters, GPS and IMU data (used to acquire its own pose data), by sensing the vehicle;

[0111] S102: Determine the initial sensing sector area based on the camera metadata of the sensing vehicle;

[0112] S103: Perform vehicle target detection and tracking on each frame of the in-vehicle video of the sensing vehicle, and generate detection boxes for surrounding vehicles.

[0113] In this embodiment, vehicle detection and tracking in a road video stream are achieved using YOLOv8n object detection and SORT object tracking. YOLOv8n is an object detection algorithm that can detect vehicles, while the SORT algorithm can determine whether objects detected in consecutive frames are the same target, thus enabling target tracking. Both are lightweight object detection and tracking algorithms, making them easy to deploy on edge devices such as mobile phones.

[0114] S104: Based on the changes in the detection boxes of surrounding vehicles, combined with camera parameters, GPS and IMU data, determine their motion state, and identify stationary surrounding vehicles as target vehicles.

[0115] In this embodiment, to initially determine whether a target vehicle is illegally parked, it is necessary to identify its movement state. The local computing unit performs the following calculations for each tracked target vehicle:

[0116] First, based on the pixel area change of the target detection box in consecutive video frames, the camera's intrinsic parameters (focal length, principal point) and extrinsic parameters (installation height, pitch angle), and combined with prior knowledge of the actual physical size of this type of vehicle, the distance change sequence of the target vehicle relative to the vehicle camera is calculated through the camera imaging geometric model.

[0117] Secondly, based on the distance change sequence and the inter-frame time interval, the radial relative velocity of the target vehicle relative to the vehicle camera is calculated.

[0118] Finally, the vehicle's own velocity vector information obtained through the GPS / IMU module is integrated to convert the relative velocity into an estimate of the target vehicle's absolute velocity in the world coordinate system. If the magnitude of this absolute velocity is lower than a preset threshold (0.5 m / s) for a continuous period of time (5 seconds), the target vehicle is determined to be stationary and marked as a potential illegally parked target, thereby triggering the subsequent road environment reliability assessment and result reporting process.

[0119] S105: Calculate the real-time occlusion angle range caused by the geometric contours of surrounding vehicles, subtract the real-time occlusion angle ranges of all detected vehicles from the initial sensing sector area of ​​the sensing vehicle, and obtain the effective sensing sector area of ​​the sensing vehicle at the current moment.

[0120] S106: Accumulate the effective sensing sector area of ​​the sensing vehicle at each moment along the time dimension to form the spatiotemporal sensing map of the sensing vehicle.

[0121] To construct an effective sensing fan-shaped region, the camera first needs to determine its initial sensing range. Then, for each surrounding vehicle detected in the image, its occlusion range is calculated. The occluded portion is subtracted from the initial sensing range, resulting in the effective sensing fan-shaped region. Specifically, the effective sensing fan-shaped region for sensing vehicles is generated through the following steps:

[0122] S1051: In At any moment, for sensing vehicles Every surrounding vehicle detected Using camera geometry to capture surrounding vehicles The four corner points of the detection frame are projected onto the ground coordinate system, and the azimuth angles of the lines connecting the four corner points of the detection frame to the camera center are calculated. Among them, the camera's internal and external parameters of the sensing vehicle and the detection box information of the surrounding vehicles are fused, and the four corner points of the detection box are projected onto the ground coordinate system (bird's-eye view coordinate system) through inverse perspective transformation.

[0123] In this embodiment, as Figure 2 (a) and Figure 2 As shown in (b), for time... Extract the lower midpoint of the bounding box for surrounding vehicles in the frame. Project it onto the ground coordinate system ;

[0124] Formula description:

[0125] ; ;

[0126] In the formula: , , , These are internal camera parameters. This refers to the camera's height above the ground.

[0127] Based on the actual length and width of the surrounding vehicles, the coordinates of the four corner points of the surrounding vehicles in the ground coordinate system can be calculated.

[0128] S1052: Regarding surrounding vehicles The minimum and maximum values ​​of the azimuth angle are taken to form the range of the occlusion angle caused by it. ;

[0129] The formula is described as follows:

[0130] ;

[0131] In the formula: Indicates surrounding vehicles The resulting obstruction area; Indicates surrounding vehicles No. The azimuth angle of the line connecting the corner point and the center of the camera;

[0132] S1053: All The total occlusion range is obtained by merging the occlusion areas caused by surrounding vehicles. ;

[0133] The formula is described as follows:

[0134] ;

[0135] In the formula: Indicates surrounding vehicles The resulting obstruction area; This represents the total occlusion area.

[0136] S1054: From sensing vehicles Subtract the total occlusion area from the initial perceived sector region The vehicle was obtained The real, effective, and efficient perception of the sector area at the present moment. .

[0137] The formula is described as follows:

[0138] ;

[0139] In the formula: Indicates the perception of vehicles The initial perceptual sector area; To sense vehicles Effective perception of the sector area.

[0140] This invention integrates vehicle-mounted camera parameters, the vehicle's own pose data, and real-time detected geometric information of surrounding vehicles. Through inverse perspective transformation and occlusion analysis, it maps vehicle detection results in image space to a ground (bird's-eye view) coordinate system. Based on the motion state of surrounding vehicles, it calculates the effective perception sector in real time, dynamically estimating the effective perception range of the vehicle in complex traffic environments, and thus constructing a real-time perception map. This method does not rely on high-precision maps or fixed sensor deployments. It eliminates the need for high-precision prior maps, achieving accurate quantification of line-of-sight occlusion and dynamic modeling of line-of-sight occlusion in complex traffic environments using only vehicle-mounted cameras and positioning data. This significantly improves the system's environmental adaptability and perception robustness in complex traffic flows (such as congestion, intersections, and curves), laying the foundation for subsequent objective assessment of road environment reliability.

[0141] II. Road Environment Reliability Score

[0142] This invention requires the sensing vehicle to assess a road environment reliability score for each potential illegally parked target. This assessment consists of two levels: first, based on the generated spatiotemporal perception map, the target... The intersection area between the occupied rectangular region and the map is calculated using an efficient geometric polygon clipping algorithm (such as the Greiner-Hormann algorithm) to ensure accuracy. Next, the algorithm confidence score is obtained, which is the average confidence score of the perceptual vehicle during target tracking obtained from object detectors such as YOLO, reflecting the determinism at the algorithm level. Finally, these two scores are weighted and fused to obtain a comprehensive road environment reliability score.

[0143] Specifically, the road environment reliability score of the target vehicle is calculated through the following steps:

[0144] S201: Calculate the target vehicle The rectangular area occupied and the corresponding sensing vehicle The ratio of the intersection area between spatiotemporally perceived maps to their own area yields a base score calculated based on environmental occlusion. ;

[0145] The formula is described as follows:

[0146] ;

[0147] In the formula: Indicates the perception of vehicles For the target vehicle The base score based on ambient occlusion; Indicates the target vehicle The area of ​​the rectangular region it occupies; Indicates the perception of vehicles The area of ​​the spatiotemporal perception map; This is the scaling factor;

[0148] S202: Acquiring Sensing Vehicles (Through target detection and tracking algorithms) Target vehicle The average confidence score output during target tracking ;

[0149] In this embodiment, the confidence score of the detected vehicle is recorded by the YOLOv8n target detection algorithm in the video frame, and then the SORT algorithm is used to track the same target, so that the average confidence score of the target during the tracking process can be calculated.

[0150] S203: The road environment reliability score is obtained by linearly weighting the base score calculated based on environmental shading and the average confidence score;

[0151] The formula is described as follows:

[0152] ;

[0153] In the formula: Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the perception of vehicles For the target vehicle The base score based on ambient occlusion; Indicates the perception of vehicles For the target vehicle The average confidence score during target tracking; and These represent adjustable weights that balance the contributions of environmental quality and algorithm confidence, respectively.

[0154] III. Local test results

[0155] In this embodiment, the local detection results of the sensing vehicles include standard information such as the target ID, type, location coordinates, and timestamp of all target vehicles, as well as the road environment reliability score and the base score calculated based on environmental occlusion for all target vehicles. The server aggregates the data uploaded by all sensing vehicles, matches the same target vehicles reported by different sensing vehicles, and constructs a detection matrix and a road environment reliability score matrix for subsequent model inference.

[0156] IV. Probabilistic Generation Model

[0157] In this embodiment, the detection matrix and reliability matrix for multiple vehicles and multiple targets are constructed through the following steps:

[0158] S401: Use a vehicle re-identification algorithm to match the same target vehicle detected by different sensing vehicles;

[0159] S402: Construct a binary detection matrix with M rows (target vehicle) and S columns (sensing vehicle). Detection matrix In the detection matrix, rows represent different target vehicles, and columns represent different sensing vehicles; elements in Indicates the perception of vehicles The report detected the target vehicle. ,element This indicates that nothing was detected.

[0160] S403: Constructing a Road Environment Reliability Score Matrix :like Then the road environment reliability score matrix elements in ;like Then the road environment reliability score matrix elements in .

[0161] Specifically, the inherent reliability parameters of the sensing vehicle, namely the true positive rate and the true negative rate, are introduced. A probabilistic generation model that integrates the inherent reliability of the environment and the vehicle is established by combining the detection matrix of multiple vehicles and multiple targets and the road environment reliability score matrix. The true positive rate is the probability that the sensing vehicle correctly reports "detected" when the target is actually illegally parked, and the true negative rate is the probability that the sensing vehicle correctly reports "not detected" when the target is not illegally parked.

[0162] By coupling environmental factors and perceived vehicle inherent reliability factors into a unified likelihood function through an exponential weighting relationship, a probabilistic generative model is obtained.

[0163] The formula is described as follows:

[0164] ;

[0165] In the formula: express and The probability of taking different values; Indicates the perception of vehicles For the target vehicle The test results; For the target vehicle The actual parking status; Indicates the perception of vehicles The true positive rate of the test; Indicates the perception of vehicles The true negative rate of the test; Indicates the perception of vehicles For the target vehicle The road environment reliability score.

[0166] V. Expectation-Maximization Algorithm

[0167] This invention establishes a probabilistic generation model that integrates road environment reliability and vehicle inherent reliability. The model takes the detection matrix and environment matrix as input, initializes relevant parameters, uses the expectation-maximization algorithm for joint inference, iteratively executes the probabilistic inference and parameter update process until the model converges, and finally outputs a high-confidence illegal parking status judgment result and quantitative evaluation parameters of the sensing reliability of each sensing vehicle.

[0168] In this embodiment, the processing steps for joint inference of the probability generation model using the expectation-maximization algorithm include:

[0169] S601: Initialize the inherent reliability parameters of the sensing vehicle, namely the true positive rate and the true negative rate;

[0170] In this embodiment, since the actual illegal parking status of the target is unknown, it is assumed that the illegal parking status of each target is a Bernoulli random variable;

[0171] The formula is described as follows: ;

[0172] In the formula: For the target vehicle The actual parking status, This is the global parking violation rate parameter.

[0173] Based on the hierarchical Bayesian model, we further assume this global illegal parking rate It is itself a random variable and follows a Beta distribution;

[0174] The formula is described as follows: ;

[0175] In the formula: and These are hyperparameters.

[0176] S602: In the E-step of the expectation-maximization algorithm, fix the inherent reliability parameters of all currently perceived vehicles, and calculate the posterior probability of each target vehicle having engaged in illegal parking behavior based on Bayes' theorem. ;

[0177] S603: In the M-step of the expectation-maximization algorithm, the current posterior probability estimate is fixed, and the inherent reliability parameters of all perceived vehicles are updated by weighted counting, using the complement of the road environment reliability score as the weight.

[0178] S604: Alternately execute steps S502 and S503 until the probabilistic generation model converges or the maximum number of iterations is reached;

[0179] S605: Outputs high-confidence determination results of illegal parking status for all target vehicles, and evaluation results of the sensing reliability of all sensing vehicles; among which, the determination results of illegal parking status for target vehicles... By analyzing the posterior probability Binarization is performed using a threshold; the evaluation results of the vehicle sensing reliability include the detection true positive rate. and detection of true negative rate The estimated value.

[0180] 1. Posterior probability

[0181] In this embodiment, for each target, the likelihood of all vehicle reports is calculated under the assumption that the parking was indeed illegal.

[0182] The formula is described as follows:

[0183] ;

[0184] In the formula: Let the likelihood value be... To sense vehicles Detecting the true positive rate, To sense vehicles For the target vehicle The road environment reliability score, To sense vehicles For the target vehicle The test results.

[0185] Similarly, assuming the target vehicle is not illegally parked, the likelihood of all vehicle reports is observed as follows:

[0186] The formula is described as follows:

[0187] ;

[0188] In the formula: Let the likelihood value be... To sense vehicles Detecting the true negative rate, To sense vehicles For the target vehicle The road environment reliability score, To sense vehicles For the target vehicle The test results.

[0189] Based on Bayes' theorem and prior probabilities, calculate each target... The posterior probability of illegal parking:

[0190] The formula is described as follows:

[0191] ;

[0192] In the formula: Indicates the target vehicle The posterior probability of illegal parking; Let represent the prior global violation rate parameter, whose initial value is a random variable and follows a Beta prior distribution: ; and These represent the likelihood values ​​under conditions of illegal parking and no illegal parking, respectively.

[0193] 2. Inherent reliability parameter update

[0194] In this embodiment, the specific steps for updating the inherent reliability parameters of the sensing vehicle include:

[0195] S6031: Update prior global illegal parking rate parameters ;

[0196] The public notice indicates that:

[0197] ;

[0198] In the formula: This represents the prior global illegal parking rate parameter; and These are hyperparameters; Indicates the total number of target vehicles; Indicates the first step;

[0199] S6032: Update True Yang Rate ;

[0200] The formula is described as follows:

[0201] ;

[0202] In the formula: Indicates the perception of vehicles The true positive rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the target vehicle The posterior probability of illegal parking; To sense vehicles For the target vehicle The test results;

[0203] S6033: Update True Negative Rate ;

[0204] The formula is described as follows:

[0205] ;

[0206] In the formula: Indicates the perception of vehicles The true negative rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the target vehicle The posterior probability of illegal parking; To sense vehicles For the target vehicle The test results.

[0207] This invention proposes a dual-reliability probabilistic inference model based on the Expectation-Maximization (EM) algorithm. By decomposing road environment reliability into external factors influenced by environmental dynamics and internal intrinsic factors reflecting the vehicle's own perception capabilities, an exponentially weighted likelihood function, weighted by road environment reliability, is designed and coupled with each vehicle's inherent true positive / true negative ratio. This innovatively unifies two types of uncertainty sources into a single probabilistic generation model and iteratively optimizes the simultaneous inference of illegal parking status and vehicle reliability parameters. The advantages of this design lie in its ability to simultaneously handle external dynamic traffic occlusion and internal unknown equipment bias. The EM algorithm enables joint optimization of violation status inference and reliability parameter learning, iteratively optimizing high-confidence violation event judgments and quantitative assessments of the sensing reliability of each vehicle. This not only significantly improves the accuracy and robustness of the fusion results but also outputs a quantitative assessment of the long-term performance of each sensing vehicle.

[0208] VI. Experimental Instructions

[0209] To make the objectives, technical solutions, and advantages of this invention clearer, the embodiments of this invention will be further described in detail below in conjunction with simulation and real road test experiments.

[0210] 1. Simulation Experiment Implementation Method:

[0211] To fully verify the effectiveness of the method of this invention, the CARLA simulation platform (version 0.9.14), widely recognized in the field of autonomous driving, was selected to construct the test environment. Experiments were conducted in two scenarios with complex urban road networks, "Town10HD" and "Town04," built into CARLA. The test road types comprehensively covered straight roads, curves, intersections, etc.; the number of lanes ranged from single lane to four lanes to simulate different scenarios from side roads to urban arterial roads. Furthermore, to fully test the system's environmental adaptability, this invention set various combinations of lighting and weather conditions: including sunny days (default), rainy days (slippery roads, precipitation obstructing visibility), foggy days (gradual changes in visibility), and nighttime (relying on streetlights).

[0212] In each test cycle, this invention statically places 15 to 40 vehicles as target vehicles for illegal parking on a designated road section. Simultaneously, to simulate real dynamic traffic flow, this invention uses CARLA's traffic flow generator to generate 5 to 50 ordinary vehicles, whose behavior is controlled by an AI controller, possessing capabilities such as following, lane changing, and obstacle avoidance. This constructs different traffic density scenarios ranging from "open" (minimum longitudinal distance between vehicles ≥ 8 meters), "normal" (distance 2.8-8 meters), to "congested" (distance ≤ 2.8 meters). All sensing vehicles are equipped with a virtual vehicle-mounted RGB camera, with its core parameters set as follows: horizontal field of view (FOV) of 100 degrees, maximum sensing range of 30 meters, installation height of 0.7 meters, and optical center axis parallel to the ground and strictly aligned with the vehicle's direction of travel.

[0213] The sensing vehicle continuously acquires RGB images of the road ahead at a resolution of 640×480 at a rate of 30 frames per second, while simultaneously recording its own high-frequency GPS trajectory and IMU data. All of this data is input in real time into the algorithm flow of this invention, which is simulated and deployed on the edge computing unit of the vehicle. The entire simulation experiment involved the deployment of over 400 sensing vehicles, and the collected driving data covered all preset environmental and traffic combinations, ensuring the effectiveness and significance of subsequent statistical analysis.

[0214] To conduct a fair and comprehensive performance evaluation, this invention was compared with several advanced baseline methods. These methods represent different technical approaches:

[0215] (1) Vehicle selection-based methods: including CS-MAX (a greedy algorithm designed to maximize the joint coverage area of ​​multi-vehicle perception regions) and CodeFilling (which generates binary indicator masks from image bounding boxes and selects vehicles to maximize mask coverage of the target lane region).

[0216] (2) Reliability estimation-based methods: including RATD (a reliability adaptive truth discovery method based on fast gradient descent, which can jointly optimize source reliability and truth estimation) and SpotE-TD (which uses block coordinate descent to minimize the difference between the binary label prediction and observation of each target, while estimating the reliability weight of each sensing vehicle).

[0217] (3) Variations of the present invention (for ablation studies): include removing the environmental dependence reliability assessment module, only updating vehicle dependence reliability and illegal parking inference alternately through the EM algorithm and removing the probabilistic inference model, using a predefined road environment reliability threshold to screen detection results and perform weighted voting.

[0218] The evaluation uses precision (the proportion of correctly detected illegal parking out of all reported violations) and recall (the proportion of correctly detected illegal parking out of all actual violations), both widely accepted metrics in information retrieval, as core performance indicators. Extensive simulation results demonstrate that the proposed solution consistently exhibits optimal and stable performance under various challenging scenarios, including different traffic densities, numbers of participating vehicles, weather and lighting conditions, and lane positions.

[0219] The specific experimental results are as follows:

[0220] Under different traffic densities (congested, normal, and empty), the average recall rate of this invention is 8.7% to 23.3% higher than that of four mainstream baseline methods (CS-MAX, CodeFilling, RATD, SpotE-TD), and the precision rate is 9.5% to 19.3% higher. The reason for this is that in congested road conditions, severe occlusion by adjacent vehicles leads to a large number of false negatives or false positives in the baseline methods. This invention, by explicitly modeling the reliability of environmental occlusion and incorporating probabilistic inference, effectively suppresses the negative impact of unreliable detection, thus maintaining a precision rate of 90% and a recall rate of 86%. This demonstrates its excellent anti-interference and environmental adaptability. The results are as follows: Figure 3 (a) and Figure 3 As shown in (b).

[0221] In tests with varying vehicle participation rates (gradually increasing from 10% to 70%), the detection performance of this invention showed a stable and monotonous upward trend with increasing vehicle numbers. When the participation rate reached 70%, the recall rate reached 97%, approaching saturation. More importantly, at different participation rates, the average recall and precision of this invention were 8.5% and 7.4% higher than the baseline method, respectively. This indicates that the dual reliability mechanism proposed in this invention can more intelligently and efficiently fuse multi-source information, achieving detection results with fewer vehicles compared to other methods that require more vehicles, thus improving the system's economy. The results are as follows... Figure 4 (a) and Figure 4 As shown in (b).

[0222] In system robustness testing, the average recall and precision of this invention were 81.7% and 88.7% under different daytime weather conditions (sunny, rainy, foggy), respectively. At night, due to insufficient light and a significant decrease in image quality, the average metrics decreased to 70.3% and 87.3%. Even so, the performance of this invention under all conditions significantly outperformed all baseline comparisons, as shown in the results. Figure 5 (a) and Figure 5 (b), and Figure 6 (a) and Figure 6 As shown in (b).

[0223] In detection tasks where the sensing vehicle is positioned in different lanes (the sensing vehicle is traveling in lanes 1, 2, and 3 of a four-lane straight road, while the target vehicle is located on the side of lane 4, i.e., the sensing vehicle is traveling in the lanes near, middle, and far from the target vehicle; and the sensing vehicle is traveling in the outer lane of a two-lane curve, while the target vehicle is located on the inner lane), the performance of all methods decreases as the lateral distance between the sensing vehicle and the roadside target increases, which is expected. However, it is noteworthy that even in the farthest lane (with the largest lateral distance), the recall and precision of this invention still outperform the best baseline method by 14.5% and 8%, respectively, highlighting the superiority of this method, as shown in the results. Figure 7 (a) and Figure 7 As shown in (b).

[0224] Ablation experiments strongly demonstrate the significant contributions of the two core modules of this invention: removing the road environment reliability module resulted in a sharp decrease in recall rate of 26.5%; while removing the probabilistic inference module and using only simple weighted voting also resulted in a decrease in recall rate of 19.5%. This fully demonstrates that the dual reliability model combining explicit road environment reliability assessment with implicit vehicle reliability learning, and the joint inference framework based on the EM algorithm, are indispensable and extremely effective in solving the problem of cooperative perception uncertainty.

[0225] 2. Real-world road test implementation method:

[0226] To further verify the real-world application value and deployment feasibility of this invention, a complete end-to-end prototype system was developed. This system consists of a custom-developed Android application (APK) and a central cloud server equipped with a high-performance NVIDIA RTX 3090 GPU. The prototype system is as follows... Figure 8 As shown.

[0227] The Android application (compatible with Android 10 and above) efficiently utilizes the smartphone's rear main camera and GPS / IMU module. During vehicle operation, the app continuously captures video streams from the front of the vehicle at 30 FPS and 1080p resolution, and records the vehicle's latitude, longitude, speed, heading angle, and acceleration pose data in real time. All data is processed on the mobile device using a lightweight illegal parking detection algorithm. This algorithm integrates the YOLOv8n model for vehicle target detection and employs the SORT algorithm for target tracking. After processing, the generated local structured detection results (including the target vehicle's cropped appearance features, 2D position coordinates projected onto a bird's-eye view, and a road environment reliability score calculated according to the method described in this patent) are uploaded to a central server via a 4G / 5G mobile network.

[0228] This invention underwent a two-week field deployment test in the urban area of ​​Chongqing, China. A total of six vehicles (including taxis and private cars) were equipped with the app, and collaborative sensing data was collected along more than 90 kilometers of urban roads (including areas with high rates of illegal parking such as commercial districts, residential areas, and school perimeters). All vehicles traveled naturally according to their daily operating routes, but a scheduling system ensured that they patrolled designated monitoring areas in a time-sharing and collaborative manner. This effectively controlled the time interval between any two vehicles passing the same location to within 10 minutes, thus constructing a low-cost, high-coverage dynamic collaborative sensing network.

[0229] Experimental results show that the present invention can effectively cope with complex occlusions (such as buses and green belts), lighting changes (tree shadows, bridge shadows), and uncertainties in the perception of vehicles in real-world environments. Quantitative analysis shows that compared with single-vehicle detection, 6-vehicle collaborative detection improves the recall rate by 10.3% (from 82.5% to 92.8%) and the precision rate by 8.8% (from 87.3% to 96.1%). This result is consistent with the simulation trend, successfully verifying the effectiveness and necessity of the present invention in collaborative scenarios.

[0230] In summary, numerous and comprehensive simulation experiments and real-world road experiments have consistently demonstrated that the technical solution combining "dynamic road environment reliability assessment" and "dual reliability probabilistic reasoning" provided by this invention can resolve the collaborative perception paradox of "the more vehicles involved, the worse the perception results," significantly improving the accuracy, robustness, and practical application value of illegal parking detection in a group vehicle collaborative environment, and providing a novel technical means for intelligent traffic management.

[0231] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A group vehicle cooperative illegal parking detection method based on vehicle-road multivariate reliability modeling, characterized in that, include: S1: Acquire sensor data through several sensing vehicles, and generate a spatiotemporal perception map for each sensing vehicle that includes the target vehicles it detects based on the sensor data. S2: Calculate the road environment reliability score of the detected target vehicle based on the spatiotemporal perception map of the perceived vehicle. S3: Obtain local detection results of all perceived vehicles, including the road environment reliability scores of the target vehicles they detect, through the server; S4: Construct a multi-vehicle-multi-target detection matrix and a road environment reliability score matrix based on the local detection results of all perceived vehicles through the server; S5: Construct a probabilistic generative model based on the detection matrix of multiple vehicles and multiple targets and the road environment reliability score matrix; Introducing the inherent reliability parameters of the sensing vehicle, namely the true positive rate and the true negative rate, and combining the multi-vehicle-multi-target detection matrix and the road environment reliability score matrix, a probabilistic generation model that integrates the inherent reliability of the environment and the vehicle is established. The formula for the probabilistic generative model is described as follows: ； In the formula: express and The probability of taking different values; Indicates the perception of vehicles For the target vehicle The test results; For the target vehicle The actual parking status; Indicates the sensing vehicle The true positive rate of the test; Indicates the perception of vehicles The true negative rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; S6: Perform joint inference on the probabilistic generation model using the expectation-maximization algorithm to output the illegal parking status judgment results for all target vehicles; Step S6, the processing steps for joint inference of the probabilistic generation model using the expectation-maximization algorithm, include: S601: Initialize the inherent reliability parameters of the sensing vehicle, namely the true positive rate and the true negative rate; S602: In the E-step of the expectation-maximization algorithm, fix the inherent reliability parameters of all currently perceived vehicles, and calculate the posterior probability of each target vehicle exhibiting illegal parking behavior based on Bayes' theorem. ; S603: In the M-step of the expectation-maximization algorithm, the current posterior probability estimate is fixed, and the inherent reliability parameters of all perceived vehicles are updated by weighted counting, using the complement of the road environment reliability score as the weight. S604: Alternately execute steps S602 and S603 until the probabilistic generation model converges or the maximum number of iterations is reached; S605: Outputs high-confidence determination results of illegal parking status for all target vehicles, and evaluation results of the sensing reliability of all sensing vehicles; among which, the determination results of illegal parking status for target vehicles... By analyzing the posterior probability Binarization is performed using a threshold; the evaluation results of the vehicle sensing reliability include the detection true positive rate. and detection of true negative rate The estimated value.

2. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 1, characterized in that: In step S1, a spatiotemporal perception map is generated for the sensing vehicle through the following steps: S101: Acquire sensor data, including in-vehicle video, camera parameters, and its own pose data, by sensing the vehicle. S102: Determine the initial sensing sector area based on the camera metadata of the sensing vehicle; S103: Perform vehicle target detection and tracking on each frame of the in-vehicle video of the sensing vehicle, and generate detection boxes for surrounding vehicles. S104: Based on the changes in the detection boxes of surrounding vehicles, combined with camera parameters and its own pose data, determine its motion state, and take the surrounding vehicles that are stationary as the target vehicles. S105: Calculate the real-time occlusion angle range caused by the geometric contours of surrounding vehicles, subtract the real-time occlusion angle ranges of all detected vehicles from the initial sensing sector area of ​​the sensing vehicle, and obtain the effective sensing sector area of ​​the sensing vehicle at the current moment. S106: Accumulate the effective sensing sector area of ​​the sensing vehicle at each moment along the time dimension to form the spatiotemporal sensing map of the sensing vehicle.

3. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 2, characterized in that: In step S105, the effective sensing sector area of ​​the sensing vehicle is generated through the following steps: S1051: In At any moment, for sensing vehicles Every surrounding vehicle detected Using camera geometry to capture surrounding vehicles The four corner points of the detection frame are projected onto the ground coordinate system, and the azimuth angles of the lines connecting the four corner points of the detection frame to the center of the camera are calculated. S1052: Regarding surrounding vehicles The minimum and maximum values ​​of the azimuth angle are taken to form the range of the occlusion angle caused by it; S1053: Combine the occlusion areas caused by all surrounding vehicles to obtain the total occlusion range; S1054: From sensing vehicles Subtracting the total occlusion range from the initial perceived sector area, we obtain the perceived vehicle. The effective sensing sector area at the current moment.

4. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 2, characterized in that: In step S2, the road environment reliability score of the target vehicle is calculated through the following steps: S201: Calculate the target vehicle The actual rectangular area corresponding to the detection box and the corresponding sensing vehicle The ratio of the intersection area between spatiotemporally perceived maps to their own area yields a base score calculated based on environmental occlusion. ; The formula is described as follows: ； In the formula: Indicates the perception of vehicles For the target vehicle The base score based on ambient occlusion; Indicates the target vehicle The rectangular area it occupies; Indicates the perception of vehicles Spatiotemporal perception map; express and The area of ​​their intersection; express The area; This is the scaling factor; S202: Acquiring Sensing Vehicles For the target vehicle The average confidence score output during target tracking ; S203: The road environment reliability score is obtained by linearly weighting the base score calculated based on environmental shading and the average confidence score; The formula is described as follows: ； In the formula: Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the perception of vehicles For the target vehicle The base score based on ambient occlusion; Indicates the perception of vehicles For the target vehicle The average confidence score during target tracking; and These represent adjustable weights.

5. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 4, characterized in that: In step S3, the local detection results of the perceived vehicle include standard information such as the target ID, type, location coordinates and timestamp of all target vehicles, as well as the road environment reliability score and the basic score calculated based on environmental occlusion of all target vehicles.

6. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 5, characterized in that: In step S4, the detection matrix and reliability matrix for multiple vehicles and multiple targets are constructed through the following steps: S401: Use a vehicle re-identification algorithm to match the same target vehicle detected by different sensing vehicles; S402: Construct the detection matrix Detection matrix In the detection matrix, rows represent different target vehicles, and columns represent different sensing vehicles; elements in Indicates the perception of vehicles The report detected the target vehicle. ,element This indicates that nothing was detected. S403: Constructing a Road Environment Reliability Score Matrix :like Then the road environment reliability score matrix elements in ;like Then the road environment reliability score matrix elements in .

7. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 1, characterized in that: In step S602, the inherent reliability parameters of all current sensing vehicles are fixed. Based on Bayes' theorem and prior probability, the posterior probability of the target vehicle illegally parking is calculated. The formula is described as follows: ； In the formula: Indicates the target vehicle The posterior probability of illegal parking; This represents the prior global illegal parking rate parameter; and These represent the likelihood values ​​under conditions of illegal parking and no illegal parking, respectively.

8. The method for detecting illegal parking by group vehicles based on vehicle-road multivariate reliability modeling as described in claim 7, characterized in that: In step S603, the specific steps for updating the inherent reliability parameters of the perceived vehicle include: S6031: Update prior global illegal parking rate parameters ; The public notice indicates that: ； In the formula: This represents the prior global illegal parking rate parameter; and These are hyperparameters; Indicates the total number of target vehicles; Indicates the first step; S6032: Update True Yang Rate ; The formula is described as follows: ； In the formula: Indicates the perception of vehicles The true positive rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the target vehicle The posterior probability of illegal parking; To sense vehicles For the target vehicle The test results; S6033: Update True Negative Rate ; The formula is described as follows: ； In the formula: Indicates the perception of vehicles The true negative rate of the test; Indicates the perception of vehicles For the target vehicle Road environment reliability score; Indicates the target vehicle The posterior probability of illegal parking; To sense vehicles For the target vehicle The test results.

Images