Intelligent network connection unmanned equipment digital supervision system

Abstract

The application discloses an intelligent networked unmanned equipment digital supervision system.In the application, an entity filing system is constructed through visual fingerprint and trace appearance comparison technology.Combined with time-space position mutual exclusion judgment logic, fraud behaviors of multiple vehicles sharing the same code are accurately identified, and vehicle identification code is used as the main key to associate the record information, running track and visual features, three-dimensional cross verification of the trace appearance is realized, a vehicle file holographic portrait is established for each vehicle on the road, and the problem of unclear bottom number caused by serious inversion between record vehicles and actual running vehicles is solved.Combined with monocular depth estimation, vehicle length, width and height physical parameters are calculated in real time and compared with the record size automatically, and the GPS track is matched to the high-precision map lane layer by using a hidden Markov model to accurately determine whether the vehicle is driving in a non-motor vehicle lane, and the size compliance review is moved from the laboratory to the road operating environment, and the discovery rate of size violation behaviors and the accuracy of road right supervision are improved.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically a digital monitoring system for intelligent connected autonomous driving equipment. Background Technology

[0002] With the rapid development of autonomous driving technology, autonomous driving equipment has moved from the road testing phase to large-scale commercial operation. Taking Wuxi City as an example, since the establishment of the autonomous driving equipment registration and access system in 2023, a total of 273 test demonstration codes have been issued as of June 2025. Leading companies such as Neolix and Jiushi have successively launched their systems, and autonomous driving equipment has entered the stage of large-scale commercial operation. However, as companies shift from the testing phase to large-scale operation, the mismatch between industry growth and technological maturity has become increasingly serious, leading to a concentrated outbreak of regulatory conflicts. According to relevant statistics, in 2025, a certain city saw 248 police incidents involving autonomous driving equipment, an increase of 153.06% year-on-year; 125 accidents, an increase of 115.52% year-on-year; and 47 public complaints, an increase of 104.35% year-on-year. All three types of data showed a doubling growth.

[0003] However, existing technologies for regulating autonomous driving equipment suffer from the following shortcomings: First, there is a lack of effective means to identify unregistered vehicles on the road. Current regulations rely heavily on the administrative registration coding system, which is insufficient to address covert violations such as "one license plate for multiple vehicles." Furthermore, video surveillance networks lack dedicated recognition algorithms for the appearance features of autonomous driving equipment, making it impossible to achieve proactive monitoring across the entire road network around the clock. Second, there is a lack of real-time detection capabilities for vehicle dimensions. Dimensional compliance reviews remain at the pre-registration stage, failing to detect violations such as companies circumventing size restrictions by adding enclosures or replacing shells during operation. Third, there is a lack of real-time early warning mechanisms for right-of-way abuse and malfunctions. Alarms cannot be automatically triggered when vehicles occupy motor vehicle lanes or enter elevated tunnels or other high-speed scenarios, and there is no effective distinction between normal waiting and substantial obstruction. Fourth, there is a lack of penetrating regulatory means for enterprises. Individual vehicle violation data cannot be aggregated to the corresponding operating company, resulting in the ineffective transmission of violation costs to the responsible party.

[0004] In summary, existing technologies have not yet provided a systematic solution that can simultaneously address the multi-dimensional regulatory needs of unmanned equipment, such as identity recognition, size detection, behavior warning, and enterprise credit evaluation. Summary of the Invention

[0005] The purpose of this invention is to provide a digital monitoring system for intelligent connected unmanned driving equipment in order to solve the problems mentioned above.

[0006] The technical solution adopted in this invention is as follows: a digital supervision system for intelligent connected unmanned driving equipment, comprising: a data access layer, a data governance layer, an algorithm model layer, and an application presentation layer;

[0007] The algorithm model layer includes: a visual measurement engine, a spatial matching engine, and a rule reasoning engine;

[0008] The API gateway output of the data access layer is connected to the input of the ID-Mapping mechanism of the data governance layer, and the output of the ID-Mapping mechanism is connected to the holographic profile library.

[0009] The output of the image database is connected to the spatial matching engine, visual measurement engine and rule reasoning engine of the algorithm model layer respectively; the output of the spatial matching engine is connected to the visual command screen and the rule reasoning engine; the output of the visual measurement engine is connected to the rule reasoning engine and the mobile law enforcement APP; and the output of the rule reasoning engine is connected to the mobile law enforcement APP, the enterprise credit evaluation report and the visual command screen.

[0010] The query interface for the overall overview in the application presentation layer is connected to vehicle tracing, which in turn is connected to enterprise profiles. All three are connected back to the output interface of the profile library and the rule reasoning engine.

[0011] In a preferred embodiment, the visual measurement engine, taking into account the characteristics of unmanned driving equipment and unmanned drivers that are difficult to distinguish visually, utilizes ResNet-50 as the backbone network to extract tamper-proof, fine-grained appearance features such as vehicle logos, decal textures, and damage, generating a 512-dimensional feature vector. The algorithm constructs a Siamese neural network to calculate the cosine similarity of different capture feature vectors under the same encoding. If the similarity is less than 0.7 and the two spatiotemporal locations are mutually exclusive (i.e., the distance between the two points is greater than 5 kilometers), it is confirmed as multiple vehicles with the same license plate. The formula involved in this algorithm is the cosine similarity calculation formula:

[0012]

[0013] in and These represent the 512-dimensional visual feature vectors extracted from two different snapshots under the same encoding. The dot product of two vectors, denominator Sim is the product of the magnitudes of two vectors. The value of Sim ranges from 0 to 1. The closer it is to 1, the more similar the two vectors are.

[0014] In a preferred embodiment, the visual measurement engine is based on computer vision for exterior dimension recognition and out-of-specification verification. The algorithm utilizes pixel-level analysis of images captured at checkpoints, calibrates camera intrinsic parameters by identifying known reference objects in the image, such as the standard lane width of 3.5 meters and traffic sign dimensions, and uses monocular depth estimation technology to infer the vehicle's spatial position, achieving real-time reconstruction of the vehicle's three-dimensional exterior. It automatically outputs the vehicle's length, width, and height physical parameters, and establishes constraints based on the hard indicator of width ≤ 1.4 meters in the work guidelines. Simultaneously, it compares the measured width with the width registered in the database to identify instances of vehicles exceeding their registered dimensions. The formula involved in this algorithm is the out-of-specification judgment formula:

[0015]

[0016] in The representative algorithm outputs the physical parameters of the vehicle width in meters after monocular depth estimation and calibration with a reference object. The value is for exceeding the standard. A value of 1 indicates that the vehicle width exceeds the provincial standard limit, and a value of 0 indicates compliance.

[0017] In a preferred embodiment, the spatial matching engine performs trajectory-map matching and right-of-way analysis. This algorithm utilizes high-precision maps for lane-level monitoring. It uses a Hidden Markov Model to match GPS trajectory point sequences with the high-precision map road network to solve for the optimal path. It reads the attributes of the matched road segments; if the road segment attribute is a motor vehicle lane and the vehicle's continuous travel distance exceeds 50 meters, a lane occupancy alarm is triggered. Simultaneously, it uses geofence collision detection to monitor in real time whether vehicles exceed the permitted driving area to identify unauthorized driving behavior. The formula involved in this algorithm is the optimal path solution formula:

[0018]

[0019] Where Z represents the GPS trajectory point sequence. S represents a candidate path in the high-precision map road network GG. Let P(Z|S) represent the optimal matching path, where P(Z|S) is the emission probability of the observed trajectory ZZ given path S, and P(S) is the prior probability of path SS itself. This means selecting the path that maximizes the product from all candidate paths.

[0020] In a preferred embodiment, the rule-based reasoning engine addresses the pain point of congestion caused by fault-related delays by employing the ViBe algorithm to detect background differences and count the number of vehicles behind them. Simultaneously, it calculates the standard deviation of trajectory points to determine whether a vehicle is physically stationary. The visual queue analysis results and trajectory dwell analysis results are weighted and fused to construct a delay confidence score to distinguish between brief waiting and substantial obstruction. When the confidence score exceeds a threshold, a fault-related delay warning is triggered. The formula involved in this algorithm is the delay confidence score calculation formula:

[0021]

[0022] in Represents the confidence level of the blockade. The number of vehicles stranded behind the vehicle as counted by the ViBe algorithm. The indicator function takes the value 1 if there are more than 5 vehicles in the queue behind it, and 0 otherwise. Let the standard deviation of the trajectory points be . The threshold for determining stillness. The indicator function is set to 1 when the standard deviation of the trajectory points is less than the threshold, i.e., the vehicle is physically stationary, otherwise it is set to 0. α and β are the weight coefficients of the visual queue feature and the trajectory dwell feature, respectively.

[0023] In a preferred embodiment, the spatial matching engine establishes a highest-priority protection mechanism for key areas with speed limits of 80 km / h, such as elevated roads, tunnels, and bridges. It presets absolute no-entry electronic fences in a high-precision map. Simultaneously, in areas with speed limits of 60 km / h or higher, it detects whether the speed of the autonomous vehicle is below 40 km / h and marks it as an abnormally low-speed obstacle. A dedicated recognition algorithm is deployed using entrance checkpoint cameras; once an intrusion is confirmed, a P0-level highest alarm is immediately triggered to guide police forces for rapid interception. This algorithm uses both spatial red lines and speed red lines as dual judgment criteria.

[0024] In a preferred embodiment, the rule-based reasoning engine, considering the legal characteristic of unmanned vehicles operating without a driver on-site, shifts the regulatory focus back to the operating company. It automatically aggregates single-vehicle violation data to the corresponding company using OCR-recognized contact numbers, vehicle sticker brands, and other data. A weighted scoring model is constructed to assign weighted scores to various violations committed by the company within a given period. Simultaneously, an exponential decay function is introduced to reduce the weight of historical, outdated data, highlighting recent compliance. Based on the final credit score, a three-color regulatory level (red, yellow, green) is assigned. Companies with a score below 70 (red code) automatically trigger joint interviews and suspension of operations. The algorithm involves the weighted scoring model and the time decay function:

[0025]

[0026] in This represents the final credit score of company E within period T. Here, n represents the company's baseline score, n represents the total number of different types of violations committed by the company during the period, and i represents the violation type number. Here are the weighting coefficients for the i-th type of illegal behavior: w=4 for entering a restricted area, w=3 for exceeding regulations, and w=2 for occupying a road. The number of times the enterprise committed the i-th type of violation during the period is given. The total deduction is the sum of the products of the weights of each type and the number of violations. f(t) is the time decay function, where t is the time interval from the current assessment time, λ is the decay rate constant, and e is the base of the natural logarithm. The value of this function is multiplied by the original weight of the corresponding historical violation to achieve exponential decay of the weight of the old data.

[0027] In a preferred embodiment, the data access layer connects to four types of external data sources through an API gateway. The first type is the Ministry of Industry and Information Technology (MIIT) filing database, providing static attribute data for vehicles, including basic file information such as vehicle model parameters, external dimensions, autonomous driving test reports, and enterprise credit codes. The second type is a dedicated video network, providing unstructured images and video streams captured by city-wide electronic police, checkpoint electronic police, and roadside surveillance cameras. The third type is the traffic management integrated platform, providing historical and real-time business data such as violation records, accident information, and public complaints. The fourth type is the enterprise operation platform, providing real-time dynamic trajectory and status data of vehicles in operation, including sub-meter accuracy positioning coordinates, instantaneous speed, and timestamps. After protocol conversion and unified aggregation through the API gateway, all of the above four types of data are transmitted to the data governance layer for further processing.

[0028] In a preferred embodiment, the data governance layer uses an ID-Mapping unified identifier mapping mechanism as its core module. This module uses the Vehicle Identification Number (VID) as the primary key to associate and fuse four types of heterogeneous data from the data access layer. Specifically, using the VID as the anchor point, it performs cross-source association between the VIN code in the registration database and vehicle basic parameters, the 512-dimensional appearance feature vector extracted by the visual fingerprint algorithm in the video private network, violation accident records in the six-in-one platform, and continuous trajectory point sequences in the enterprise platform. This mapping mechanism eliminates the silo effect between different data sources, forming a wide-form holographic profile database with the VID as the primary key. After data cleaning and standardization, this profile database is output to the algorithm model layer as input to three analysis engines, while retaining a query interface for direct use by the application presentation layer.

[0029] In a preferred embodiment, the application presentation layer comprises two main components: visualization and business applications. The visualization component, relying on a comprehensive overview module, displays a heatmap of the city's unmanned driving equipment operation, high-violation-rate road sections, and real-time alarm locations based on a GIS map. It also provides a vehicle traceability module, allowing users to retrieve the holographic file of any vehicle using visual fingerprints or registration codes. The file includes historical trajectory playback and a heatmap of violation frequency. Additionally, a corporate profile module dynamically generates corporate credit scoring reports and displays red, yellow, and green regulatory levels. The business applications component provides differentiated services to two types of users: grassroots traffic police brigades receive P0-level, second-level alarms and interception commands via a mobile enforcement app and can provide on-site feedback on handling results; the Ministry of Industry and Information Technology views industry monitoring data and corporate credit evaluation results through a regulatory portal to trigger joint interview processes. The underlying data for both user views is connected back to the holographic profile library in the data governance layer and the rule reasoning engine output interface in the algorithm model layer, ensuring real-time synchronization between front-end display and back-end analysis.

[0030] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0031] 1. This invention constructs a physical filing system through visual fingerprint and document appearance comparison technology. Traditional supervision relies heavily on administrative filing codes, which cannot address hidden violations such as multiple vehicles with the same license plate or unregistered vehicles on the road. This invention extracts tamper-proof, fine-grained appearance features such as graffiti, sticker textures, and scratches to generate vehicle visual fingerprints. Combined with spatiotemporal location mutual exclusion judgment logic, it accurately identifies the behavior of multiple vehicles sharing the same code. Using the vehicle identification code as the primary key, it links filing information, operating trajectory, and visual features to achieve three-dimensional cross-verification of document appearance, creating a unique holographic profile for each vehicle on the road, improving the accuracy and completeness of regulatory agencies' control over the number of vehicles in operation.

[0032] 2. In this invention, known reference objects such as lane lines and traffic signs in the checkpoint image are used to complete camera calibration. The vehicle's length, width and height physical parameters are calculated in real time by combining monocular depth estimation and automatically compared with the registered dimensions. At the same time, a hidden Markov model is used to match the GPS trajectory to the lane layer of the high-precision map to accurately determine whether the vehicle is driving in the non-motorized vehicle lane. The size compliance review is moved from the laboratory to the road operating environment, which improves the detection rate of size violations and the accuracy of right-of-way supervision.

[0033] 3. This invention employs an active early warning mechanism that combines multimodal fusion with key area protection. On one hand, visual algorithms are used to detect the number of vehicles queuing behind, and on the other hand, the standard deviation of trajectory points is calculated to determine the physical stationary state of the vehicles. The results of these two analyses are fused to accurately distinguish between normal waiting and actual obstruction. Simultaneously, electronic fences are preset in key areas such as elevated roads and tunnels on a high-precision map. Once an intrusion or abnormally low-speed driving is detected, the highest level alarm is triggered, reducing the time delay from the occurrence of an event to the response and lowering the risk of rear-end collisions and secondary accidents.

[0034] 4. In this invention, contact numbers and brand logos on the vehicle are identified using OCR. Combined with visual fingerprints and a registration database, each violation is automatically attributed to the corresponding company. Based on this, a weighted scoring model is constructed, and a time decay function is introduced to dynamically calculate the company's real-time credit score, categorizing it into red, yellow, and green levels. Red-level companies automatically trigger joint interviews and suspension of new route approvals, among other punitive measures. This system extends supervision from individual vehicles to the operating entity, improving the efficiency of transmitting the cost of violations to the responsible party and reducing the probability of companies illegally deploying vehicles and neglecting safety management. Attached Figure Description

[0035] Figure 1 This is an overall system block diagram of the present invention;

[0036] Figure 2 This is a flowchart of the system operation process in this invention. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] Example: Refer to Figure 1-2 The intelligent connected unmanned driving equipment digital supervision system includes: data access layer, data governance layer, algorithm model layer and application presentation layer;

[0039] The algorithm model layer includes: a visual measurement engine, a spatial matching engine, and a rule-based reasoning engine;

[0040] The API gateway output of the data access layer is connected to the input of the ID-Mapping mechanism of the data governance layer, and the output of the ID-Mapping mechanism is connected to the holographic profile library.

[0041] The output of the image database is connected to the spatial matching engine, visual measurement engine, and rule reasoning engine of the algorithm model layer, respectively; the output of the spatial matching engine is connected to the visualization command screen and the rule reasoning engine; the output of the visual measurement engine is connected to the rule reasoning engine and the mobile law enforcement APP; and the output of the rule reasoning engine is connected to the mobile law enforcement APP, the enterprise credit evaluation report, and the visualization command screen.

[0042] The query interface for a single overview within the application presentation layer connects to vehicle tracing, which in turn connects to enterprise profiling. All three links back to the output interface of the profiling library and the rule reasoning engine.

[0043] The visual measurement engine addresses the challenges of distinguishing between autonomous driving equipment and drivers, which are difficult to differentiate visually. It utilizes ResNet-50 as the backbone network to extract tamper-proof, fine-grained appearance features such as vehicle logos, decal textures, and damage, generating a 512-dimensional feature vector. The algorithm constructs a Siamese neural network to calculate the cosine similarity of different capture feature vectors under the same encoding. If the similarity is less than 0.7 and the two spatiotemporal locations are mutually exclusive (i.e., the distance between the two points is greater than 5 kilometers), it is confirmed as multiple vehicles with the same license plate. The formula involved in this algorithm is the cosine similarity calculation formula:

[0044]

[0045] in and These represent the 512-dimensional visual feature vectors extracted from two different snapshots under the same encoding. The dot product of two vectors, denominator Sim is the product of the magnitudes of two vectors. The value of Sim ranges from 0 to 1. The closer it is to 1, the more similar the two vectors are.

[0046] The visual measurement engine is based on computer vision for exterior dimension recognition and out-of-specification verification. This algorithm utilizes pixel-level analysis of images captured at checkpoints, calibrates camera intrinsic parameters by identifying known reference points in the images, such as the standard lane width of 3.5 meters and traffic sign dimensions, and uses monocular depth estimation technology to infer the vehicle's spatial position, achieving real-time reconstruction of the vehicle's 3D exterior. It automatically outputs the vehicle's length, width, and height physical parameters, and establishes constraints based on the hard requirement of a width ≤ 1.4 meters in the work guidelines. Simultaneously, it compares the measured width with the width registered in the database to identify instances of vehicles exceeding their registered dimensions. The formula involved in this algorithm is the out-of-specification judgment formula:

[0047]

[0048] in The representative algorithm outputs the physical parameters of the vehicle width in meters after monocular depth estimation and calibration with a reference object. The value is for exceeding the standard. A value of 1 indicates that the vehicle width exceeds the provincial standard limit, and a value of 0 indicates compliance.

[0049] Spatial Matching Engine Trajectory - Map Matching and Right-of-Way Analysis. This algorithm utilizes high-precision maps for lane-level monitoring. It uses a Hidden Markov Model to match GPS trajectory point sequences with the high-precision map road network to solve for the optimal path. It reads the attributes of the matched road segments; if the road segment attribute is a motor vehicle lane and the vehicle's continuous travel distance exceeds 50 meters, a lane occupancy alarm is triggered. Simultaneously, it uses geofence collision detection to monitor in real time whether vehicles have exceeded the permitted driving area to identify unauthorized driving behavior. The formula involved in this algorithm is the optimal path solution formula:

[0050]

[0051] Where Z represents the GPS trajectory point sequence. S represents a candidate path in the high-precision map road network GG. Let P(Z|S) represent the optimal matching path, where P(Z|S) is the emission probability of the observed trajectory ZZ given path S, and P(S) is the prior probability of path SS itself. This means selecting the path that maximizes the product from all candidate paths.

[0052] The rule-based reasoning engine addresses the pain point of congestion caused by fault-related delays by employing the ViBe algorithm to detect background differences and count the number of vehicles behind them. Simultaneously, it calculates the standard deviation of trajectory points to determine whether a vehicle is physically stationary. The visual queue analysis results and trajectory dwell analysis results are weighted and fused to construct a delay confidence score to distinguish between brief waits and substantial blockages. When the confidence score exceeds a threshold, a fault-related delay warning is triggered. The formula involved in this algorithm is the delay confidence score calculation formula:

[0053]

[0054] in Represents the confidence level of the blockade. The number of vehicles stranded behind the vehicle as counted by the ViBe algorithm. The indicator function takes the value 1 if there are more than 5 vehicles in the queue behind it, and 0 otherwise. Let the standard deviation of the trajectory points be . The threshold for determining stillness. The indicator function is set to 1 when the standard deviation of the trajectory points is less than the threshold, i.e., the vehicle is physically stationary, otherwise it is set to 0. α and β are the weight coefficients of the visual queue feature and the trajectory dwell feature, respectively.

[0055] The spatial matching engine establishes the highest priority protection mechanism for key areas with speed limits of 80 km / h, such as elevated roads, tunnels, and bridges. It pre-defines absolute no-entry electronic fences on high-precision maps. Simultaneously, in areas with speed limits of 60 km / h or higher, it detects whether the speed of autonomous vehicles is below 40 km / h and marks them as abnormally low-speed obstacles. Dedicated recognition algorithms are deployed using entrance checkpoint cameras; once an intrusion is confirmed, a P0-level highest alarm is immediately triggered to guide police for rapid interception. This algorithm uses both spatial and speed red lines as dual judgment criteria.

[0056] The rule-based reasoning engine, addressing the legal characteristics of unmanned vehicles without on-site drivers, shifts the regulatory focus back to the operating companies. By using OCR-recognized contact numbers, vehicle sticker brands, and other data, it automatically aggregates single-vehicle violation data to the corresponding companies. A weighted scoring model is constructed to assign weighted scores to various violations committed by companies within a given period. Simultaneously, an exponential decay function is introduced to reduce the weight of historical, outdated data, highlighting recent compliance. Based on the final credit score, regulatory levels are divided into red, yellow, and green categories. Companies with a score below 70 (red code) automatically trigger joint interviews and suspension of operations. The algorithm involves the weighted scoring model and the time decay function:

[0057]

[0058] in This represents the final credit score of company E within period T. Here, n represents the company's baseline score, n represents the total number of different types of violations committed by the company during the period, and i represents the violation type number. Here are the weighting coefficients for the i-th type of illegal behavior: w=4 for entering a restricted area, w=3 for exceeding regulations, and w=2 for occupying a road. The number of times the enterprise committed the i-th type of violation during the period is given. The total deduction is the sum of the products of the weights of each type and the number of violations. f(t) is the time decay function, where t is the time interval from the current assessment time, λ is the decay rate constant, and e is the base of the natural logarithm. The value of this function is multiplied by the original weight of the corresponding historical violation to achieve exponential decay of the weight of the old data.

[0059] The data access layer connects to four types of external data sources through an API gateway. The first type is the Ministry of Industry and Information Technology (MIIT) filing database, providing static attribute data for vehicles, including basic file information such as vehicle model parameters, external dimensions, autonomous driving test reports, and enterprise credit codes. The second type is the video private network, providing unstructured images and video streams captured by city-wide electronic police, checkpoint electronic police, and roadside surveillance cameras. The third type is the traffic management integrated platform, providing historical and real-time business data such as violation records, accident information, and public complaints. The fourth type is the enterprise operation platform, providing real-time dynamic trajectory and status data of vehicles in operation, including sub-meter accuracy positioning coordinates, instantaneous speed, and timestamps. After protocol conversion and unified aggregation through the API gateway, all of these four types of data are sent to the data governance layer for further processing.

[0060] The data governance layer uses an ID-Mapping unified identifier mapping mechanism as its core module. This module uses the Vehicle Identification Number (VID) as the primary key to link and fuse four types of heterogeneous data from the data access layer. Specifically, using the VID as the anchor point, it performs cross-source association between the VIN code in the registration database and basic vehicle parameters, the 512-dimensional appearance feature vector extracted by the visual fingerprint algorithm from the video private network, violation and accident records from the six-in-one platform, and continuous trajectory point sequences from the enterprise platform. This mapping mechanism eliminates the silo effect between different data sources, forming a wide-form holographic profile library with the VID as the primary key. After data cleaning and standardization, this profile library is output to the algorithm model layer as input to three analysis engines, while retaining a query interface for direct use by the application presentation layer.

[0061] The application presentation layer comprises two main components: visualization and business applications. The visualization module, based on a GIS map, displays a heatmap of the city's unmanned driving equipment operations, high-violation-rate road sections, and real-time alarm locations. It also provides a vehicle traceability module, allowing users to retrieve a vehicle's holographic profile using visual fingerprints or registration codes. This profile includes historical trajectory playback and a heatmap of violation frequencies. Additionally, a corporate profile module dynamically generates corporate credit scoring reports and displays red, yellow, and green regulatory levels. The business applications provide differentiated services to two types of users: grassroots traffic police brigades receive P0-level, second-level alarms and interception commands via a mobile enforcement app and can provide on-site feedback on handling results; the Ministry of Industry and Information Technology views industry monitoring data and corporate credit evaluation results through a regulatory portal to trigger joint interview processes. The underlying data for both user views is connected to the holographic profile library in the data governance layer and the rule reasoning engine output interface in the algorithm model layer, ensuring real-time synchronization between front-end display and back-end analysis.

[0062] The operation method of the digital supervision system for intelligent connected unmanned equipment includes the following steps:

[0063] S1: The data access layer accesses vehicle static attribute data from the Ministry of Industry and Information Technology filing database, traffic camera images and surveillance video streams from the video private network, illegal accidents and police incident data from the six-in-one platform, and real-time dynamic trajectory and status data of vehicles from the enterprise platform through the API gateway, thereby completing the unified aggregation and protocol conversion of multi-source heterogeneous data.

[0064] S2: After receiving the raw data from the data access layer, the data governance layer uses the vehicle identification code as the primary key and associates the VIN code, visual fingerprint features, and enterprise credit code through the ID-Mapping unified identifier mapping mechanism to eliminate the silo effect between multi-source data. After cleaning and standardizing, the data is written into a one-vehicle-one-file holographic profile library.

[0065] S3: The data governance layer outputs the vehicle basic files and filing size parameters from the holographic image library to the visual measurement engine of the algorithm model layer, and outputs the trajectory point sequence and high-precision map data to the spatial matching engine to the rule inference engine.

[0066] S4: The visual measurement engine of the algorithm model layer uses ResNet-50 to extract fine-grained features of the vehicle body to generate a visual fingerprint. It calculates the cosine similarity of feature vectors captured under the same encoding to identify multiple vehicles with the same license plate. It also calculates the vehicle's external dimensions in real time through monocular depth estimation and reference object calibration to determine whether it exceeds the regulations or is a case of a vehicle with a small license plate but a large exterior. The spatial matching engine uses a hidden Markov model to match GPS trajectories to the lane layer of a high-precision map, reads road segment attributes to determine right-of-way compliance, and uses geofencing to detect intrusions into key areas. The rule reasoning engine receives the judgment results from the visual measurement engine and the spatial matching engine, uses DS evidence theory to fuse the visual queue analysis results and trajectory dwell analysis results to calculate the hindrance confidence, and performs enterprise credit weighted deduction and time decay calculation.

[0067] S5: The rule reasoning engine will integrate the generated P0-level highest alarm command, various early warning signals, enterprise credit scores and red, yellow and green three-color levels, as well as violation lists and heat map data, and output them to the data layers of the mobile law enforcement APP, enterprise credit evaluation report module and visual command screen in the application presentation layer.

[0068] S6: The application presentation layer's overview module receives high-precision map output from the spatial matching engine and alarm heat data from the rule reasoning engine based on the GIS map and displays them visually. The vehicle traceability module receives the vehicle identification code passed in by the user through clicks or searches on the interface, initiates a query request to the holographic profile library, and displays the vehicle's holographic profile. The enterprise profile module receives the enterprise credit score output by the rule reasoning engine and automatically collects all single-vehicle violation data under the enterprise's name in the vehicle traceability module, dynamically generating an enterprise credit evaluation report.

[0069] S7: The application presentation layer feeds back the on-site handling results submitted by the grassroots traffic management brigades through the mobile law enforcement APP to the rule reasoning engine, and synchronously writes the rectification and review information back to the holographic profile library of the data governance layer for profile update, completing the complete closed-loop operation from perception to handling to dynamic profile maintenance.

[0070] From the above, we can conclude that:

[0071] To address the challenges of unmanned vehicles, such as lack of drivers, difficulty in tracing their origins, and unclear inventory, this invention constructs a physical record-keeping system using visual fingerprint and identification feature comparison technology. Traditional supervision relies heavily on administrative registration codes, which cannot address hidden violations such as multiple vehicles sharing the same license plate or unregistered vehicles on the road. This invention extracts tamper-proof, fine-grained appearance features such as graffiti, sticker textures, and scratches to generate vehicle visual fingerprints. Combined with spatiotemporal location mutual exclusion judgment logic, it accurately identifies the behavior of multiple vehicles sharing the same code. Using the vehicle identification code as the primary key, it links registration information, operating trajectory, and visual features to achieve three-dimensional cross-verification of identification features, creating a unique holographic profile for each vehicle on the road. This improves the accuracy and completeness of regulatory agencies' control over the number of vehicles in operation.

[0072] This invention addresses the issue of companies circumventing size restrictions by adding enclosures or changing shells. It constructs a compliance verification model that combines computer vision-based external dimension recognition with high-precision map matching. Current regulations completely lack the capability to detect vehicle dimensions during operation, with compliance reviews limited to the registration stage. This invention utilizes known references such as lane lines and traffic signs in checkpoint images for camera calibration. It combines monocular depth estimation to calculate the vehicle's length, width, and height in real time and automatically compares this with the registered dimensions. Simultaneously, it employs a Hidden Markov Model to match GPS trajectories to the lane layer of a high-precision map, accurately determining whether a vehicle is traveling in a non-motorized vehicle lane. This moves size compliance review from the laboratory to the road operating environment, improving the detection rate of size violations and the accuracy of right-of-way supervision.

[0073] In this invention, addressing the frequent malfunctions and breakdowns of unmanned driving equipment, as well as safety hazards such as illegal lane occupation and intrusion into highways, a proactive early warning mechanism combining multimodal fusion and key area protection has been developed. This invention utilizes visual algorithms to detect the number of vehicles queuing behind, and calculates the standard deviation of trajectory points to determine the physical stationary state of vehicles. By fusing the results of these two analyses, it accurately distinguishes between normal waiting and substantial obstruction. Simultaneously, electronic fences are preset in high-precision maps for key areas such as elevated roads and tunnels. Once an intrusion or abnormally low-speed driving is detected, the highest level alarm is triggered, reducing the time delay from event occurrence to response and lowering the risk of rear-end collisions and secondary accidents.

[0074] This invention addresses the systemic flaw of unmanned vehicles lacking on-site drivers, making it difficult to hold drivers accountable for violations. It establishes a closed-loop governance system that integrates vehicle manufacturer responsibility penetration and dynamic credit scoring. The invention uses OCR to identify contact numbers and brand logos on the vehicle, combining visual fingerprints with a registration database to automatically categorize each individual vehicle violation under the corresponding company. Based on this, a weighted scoring model is constructed, incorporating a time decay function to dynamically calculate the company's real-time credit score and classify it into red, yellow, and green levels. Red-level companies automatically trigger punitive measures such as joint interviews and suspension of new route approvals. This system extends supervision from individual vehicles to the operating entity, improving the efficiency of transmitting violation costs to responsible parties and reducing the probability of unauthorized deployment and safety management deficiencies.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "comprising" or any other variations thereof is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0076] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital monitoring system for intelligent connected unmanned driving equipment, characterized in that: include: Data access layer, data governance layer, algorithm model layer, and application presentation layer; The algorithm model layer includes: a visual measurement engine, a spatial matching engine, and a rule reasoning engine; The API gateway output of the data access layer is connected to the input of the ID-Mapping mechanism of the data governance layer, and the output of the ID-Mapping mechanism is connected to the holographic profile library. The output of the image database is connected to the spatial matching engine, visual measurement engine and rule reasoning engine of the algorithm model layer respectively; the output of the spatial matching engine is connected to the visual command screen and the rule reasoning engine; the output of the visual measurement engine is connected to the rule reasoning engine and the mobile law enforcement APP; and the output of the rule reasoning engine is connected to the mobile law enforcement APP, the enterprise credit evaluation report and the visual command screen. The query interface for the overall view within the application presentation layer connects to vehicle tracing, which in turn connects to enterprise profiles. All three are linked back to the output interface of the profile library and the rule reasoning engine.

2. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The visual measurement engine addresses the characteristics of unmanned vehicles and drivers that are difficult to distinguish visually. It utilizes ResNet-50 as the backbone network to extract tamper-proof, fine-grained appearance features such as vehicle logos, decal textures, and damage, generating a 512-dimensional feature vector. The algorithm constructs a Siamese neural network to calculate the cosine similarity of different capture feature vectors under the same encoding. If the similarity is less than 0.7 and the two spatiotemporal locations are mutually exclusive (i.e., the distance between the two points is greater than 5 kilometers), it is confirmed as multiple vehicles with the same license plate. The cosine similarity calculation formula of this algorithm is: in and These represent the 512-dimensional visual feature vectors extracted from two different snapshots under the same encoding. The dot product of two vectors, denominator Sim is the product of the magnitudes of two vectors. The value of Sim ranges from 0 to 1. The closer it is to 1, the more similar the two vectors are.

3. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The visual measurement engine analyzes images captured by the checkpoint, calibrates camera intrinsic parameters by identifying known reference objects in the images, and uses monocular depth estimation technology to infer the vehicle's spatial position, achieving real-time reconstruction of the vehicle's three-dimensional outline. It automatically outputs the vehicle's length, width, and height physical parameters, and establishes constraints based on the hard requirement of a width ≤ 1.4 meters in the work guidelines. Simultaneously, it compares the measured width with the width registered in the database to identify instances of vehicles exceeding specifications. The formula for determining whether a vehicle exceeds the standard is: in The representative algorithm outputs the physical parameters of the vehicle width in meters after monocular depth estimation and calibration with a reference object. The value is for exceeding the standard. A value of 1 indicates that the vehicle width exceeds the provincial standard limit, and a value of 0 indicates compliance.

4. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The spatial matching engine employs trajectory-map matching and right-of-way analysis. This algorithm utilizes high-precision maps for lane-level monitoring. It uses a Hidden Markov Model to match GPS trajectory point sequences with the high-precision map road network to solve for the optimal path. It reads the attributes of the matched road segments; if the road segment attribute is a motor vehicle lane and the vehicle's continuous travel distance exceeds 50 meters, a lane occupancy alarm is triggered. Simultaneously, it uses geofence collision detection to monitor in real time whether vehicles exceed the permitted driving area to identify unauthorized driving behavior. The optimal path solution formula is: Where Z represents the GPS trajectory point sequence. S represents a candidate path in the high-precision map road network GG. Let P(Z|S) represent the optimal matching path, where P(Z|S) is the emission probability of the observed trajectory ZZ given path S, and P(S) is the prior probability of path SS itself. This means selecting the path that maximizes the product from all candidate paths.

5. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The rule-based reasoning engine addresses the pain point of congestion caused by fault-related delays by employing the ViBe algorithm to detect background differences and count the number of vehicles behind them. Simultaneously, it calculates the standard deviation of trajectory points to determine whether a vehicle is physically stationary. The visual queue analysis results and trajectory dwell analysis results are weighted and fused to construct a delay confidence score to distinguish between brief waiting and substantial obstruction. When the confidence score exceeds a threshold, a fault-related delay warning is triggered. The delay confidence score calculation formula is: in Represents the confidence level of the blockade. The number of vehicles stranded behind the vehicle as counted by the ViBe algorithm. The indicator function takes the value 1 if there are more than 5 vehicles in the queue behind it, and 0 otherwise. Let the standard deviation of the trajectory points be . The threshold for determining stillness. The indicator function is set to 1 when the standard deviation of the trajectory points is less than the threshold, i.e., the vehicle is physically stationary, otherwise it is set to 0. α and β are the weight coefficients of the visual queue feature and the trajectory dwell feature, respectively.

6. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The spatial matching engine establishes the highest priority protection mechanism for key areas with a speed limit of 80 km / h, such as elevated roads, tunnels, and bridges. It presets an absolute no-entry electronic fence in the high-precision map. At the same time, in areas with a speed limit of 60 km / h or higher, it detects whether the speed of the unmanned equipment is lower than 40 km / h and marks it as an abnormally low-speed obstacle. It uses entrance checkpoint cameras to deploy a dedicated recognition algorithm. Once an intrusion is confirmed, it immediately triggers the highest P0 level alarm to guide the police force to quickly intercept it. The algorithm uses both the spatial redline and the velocity redline as the dual criteria for judgment.

7. The intelligent connected unmanned driving equipment digital supervision system according to claim 1, characterized in that: The rule-based reasoning engine, considering the legal characteristics of unmanned vehicles without on-site drivers, shifts the regulatory focus back to the operating companies. It automatically aggregates single-vehicle violation data to the corresponding companies through OCR-recognized contact numbers, vehicle sticker brands, and other data. A weighted scoring model is constructed to assign weighted scores to various violations committed by companies within a given period. Simultaneously, an exponential decay function is introduced to reduce the weight of historical, outdated data, highlighting recent compliance. Based on the final credit score, regulatory levels are divided into red, yellow, and green categories. Companies with a score below 70 (red code) automatically trigger joint interviews and suspension of operations. The weighted scoring model and time decay function are as follows: in This represents the final credit score of company E within period T. Here, n represents the company's baseline score, n represents the total number of different types of violations committed by the company during the period, and i represents the violation type number. Here are the weighting coefficients for the i-th type of illegal behavior: w=4 for entering a restricted area, w=3 for exceeding regulations, and w=2 for occupying a road. The number of times the enterprise committed the i-th type of violation during the period is given. The total deduction is the sum of the products of the weights of each type and the number of violations. f(t) is the time decay function, where t is the time interval from the current assessment time, λ is the decay rate constant, and e is the base of the natural logarithm. The value of this function is multiplied by the original weight of the corresponding historical violation to achieve exponential decay of the weight of the old data.

8. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The data access layer connects to four types of external data sources through an API gateway, specifically including: The first category is the Ministry of Industry and Information Technology filing database, which provides static attribute data of vehicles, including basic file information such as vehicle model parameters, external dimensions, autonomous driving test reports and enterprise credit codes; The second category is dedicated video networks, which provide unstructured images and video streams captured by electronic police, checkpoint electronic police, and road surveillance cameras throughout the city; The third category is the traffic management six-in-one platform, which provides historical and real-time business data such as violation records, accident information, and public complaints and police reports; The fourth category is enterprise operation platforms, which provide real-time dynamic trajectory and status data of vehicles on the road, including sub-meter level positioning coordinates, instantaneous speed and timestamps.

9. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The data governance layer uses the ID-Mapping unified identifier mapping mechanism as its core module. This module uses the vehicle identification code (VID) as the primary key to associate and fuse the four types of heterogeneous data sent from the data access layer.

10. The digital monitoring system for intelligent connected unmanned driving equipment according to claim 1, characterized in that: The application presentation layer comprises two main components: visualization and business applications. The visualization part relies on the overview module to display a heat map of the city's unmanned driving equipment operation, high-violation road sections, and real-time alarm points based on a GIS map. It also provides a vehicle traceability module that allows users to retrieve the holographic file of any vehicle through visual fingerprints or registration codes. The file contains historical trajectory playback and a heat map of violation frequency. In addition, there is a corporate profile module that dynamically generates corporate credit scoring reports and displays red, yellow, and green regulatory levels.

Images