Non-inductive parking fee payment management method and platform

Abstract

The invention discloses a non-inductive parking fee payment management method and platform, and relates to the technical field of intelligent parking. The method comprises the following steps: acquiring vehicle videos shot by cameras inside and outside a field, and forming a visual sequence after frame extraction; analyzing the sequence by using a visual semantic analyzer to generate a semantic long description text; constructing an abnormal behavior query problem set according to a charging rule and a historical abnormal record; calling a multi-modal question and answer model, answering questions based on the semantic text, and outputting fine-grained answers and confidence; integrating question and answer results, and constructing a multi-node relation heterogeneous graph; and carrying out anomaly analysis on the graph, if an anomaly is found, sending out an early warning, and managing the cost according to the early warning. The technical problems that in an existing non-inductive payment parking scene, due to the fact that visual data understanding is shallow and multi-source information is isolated, abnormal behavior detection is inaccurate, and tracing is difficult are solved, and the technical effect that through deep semantic understanding and multi-relation graph analysis, the parking lot vehicle behavior recognition precision and parking fee management safety are improved is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent parking technology, specifically to a method and platform for managing parking fees through contactless payment. Background Technology

[0002] With the deepening of smart city construction, contactless payment parking has been widely adopted as an efficient and convenient service model. However, current mainstream platforms heavily rely on single license plate recognition technology, which faces severe challenges in complex real-world scenarios. For example, in poor lighting conditions, at tilted angles, or with damaged or deliberately obscured license plates, recognition accuracy drops significantly, leading to billing errors or payment failures. More problematic is the lack of effective deep perception and correlation analysis capabilities for platforms targeting malicious behaviors such as cloned license plates, tailgating, and cross-regional serial offenses. These platforms can only provide "recognition results" but cannot understand "vehicle behavior," resulting in a situation where visual data is abundant but semantic understanding is lacking, and multi-source information exists but correlation insights are insufficient. This not only causes direct economic losses but also triggers user disputes and hinders the healthy development of the industry. Summary of the Invention

[0003] This application provides a method and platform for managing parking fees through contactless payment, which solves the technical problems of inaccurate detection of abnormal behavior and difficulty in tracing the source of information caused by the superficial understanding of visual data and the isolation of multi-source information in existing contactless payment parking scenarios.

[0004] A first aspect of this application provides a method for managing parking fees via contactless payment, the method comprising: The system acquires a set of vehicle videos collected by cameras at the parking lot entrance, exit, and inside the lot. Frames are extracted from these videos to obtain a visual input information sequence for the target vehicle. A visual semantic analyzer is used to analyze this sequence, generating a long semantic description text sequence for the target vehicle. A set of query questions for abnormal vehicle behavior is constructed based on preset rules for contactless parking fee payment and a set of historical fee anomaly records. A multimodal question-answering model is invoked to answer the query questions based on the long semantic description text sequence, obtaining a fine-grained semantic question-answering result set and a question-answer confidence set. The fine-grained semantic question-answering result set and question-answer confidence set are analyzed to construct a multi-node heterogeneous relationship graph. Abnormal behavior analysis is performed on the multi-node heterogeneous relationship graph. If abnormal behavior is detected, an anomaly warning is obtained, and fee management is implemented based on the anomaly warning information.

[0005] A second aspect of this application provides a contactless payment parking fee management platform, the platform comprising: The system comprises the following modules: Video Acquisition Module: Acquires vehicle video sets from parking lot entrances, exits, and in-park cameras; extracts frames from these video sets to obtain a visual input information sequence for the target vehicle; Video Analysis Module: Analyzes the visual input information sequence using a visual semantic analyzer to generate a semantic long-description text sequence for the target vehicle; Query Question Construction Module: Constructs a set of query questions for abnormal vehicle behavior based on preset seamless parking fee payment rules and a set of historical fee anomaly records; Semantic Question Answering Module: Calls a multimodal question answering model to answer the query questions based on the semantic long-description text sequence, obtaining a fine-grained semantic question answering result set and a question answering confidence set; Question Answering Result Analysis Module: Combines the fine-grained semantic question answering result set and the question answering confidence set for analysis, constructing a multi-node heterogeneous graph; Fee Management Module: Performs abnormal behavior analysis on the multi-node heterogeneous graph. If abnormal behavior is found in the analysis results, an anomaly warning is obtained, and fee management is performed based on the anomaly warning information.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, video data of vehicles at the parking lot entrance, exit, and inside the lot is collected using camera equipment, and each frame is extracted to obtain a sequence of visual information for the target vehicles. Then, visual semantic analysis technology is used to process this image information, generating detailed semantic description text. Next, based on preset contactless payment rules and historical fee anomaly records, a set of query questions for abnormal vehicle behavior is constructed. These queries are then analyzed in depth using a multimodal question-answering model, yielding fine-grained question-answer results and corresponding confidence data. Finally, these fine-grained question-answer results are combined with the confidence data for further analysis, constructing a multi-node relationship graph to represent the correlations between vehicle behaviors. If abnormal behavior is detected through graph analysis, an alert is issued, and fee management is implemented based on the alert information, thereby achieving intelligent management and anomaly monitoring of parking lot fees. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is a schematic diagram of the contactless payment parking fee management method provided in an embodiment of this application.

[0009] Figure 2 This is a schematic diagram of the structure of the contactless payment parking fee management platform provided in this application embodiment.

[0010] Figure labeling: Video acquisition module 11, Video analysis module 12, Query question construction module 13, Semantic question answering module 14, Question answering result analysis module 15, Fee management module 16. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] Example 1, as Figure 1 As shown, this application provides a method for managing parking fees through contactless payment, the method including: The system acquires a set of vehicle videos collected by cameras at the parking lot entrance, exit, and inside the parking lot. Frames are extracted from the vehicle video set to obtain the visual input information sequence of the target vehicle.

[0013] In this embodiment, multiple cameras are installed in the parking lot to capture real-time images of vehicles entering and exiting the parking lot and their movement within it. These cameras include, but are not limited to, monitoring equipment at the entrance, exit, and around each parking space. By capturing all movement of the target vehicles in real-time, a vehicle video set is formed, containing vehicle image information at multiple time points. Subsequently, the collected vehicle video data is processed, employing frame extraction technology to extract each frame from the video stream. Each frame represents vehicle image information at a specific time point, and these images serve as the visual input sequence for subsequent analysis. This ensures that the movement trajectory and behavioral characteristics of each vehicle in the parking lot are fully captured, providing accurate information support for fee management and anomaly detection.

[0014] The visual input information sequence is analyzed using a visual semantic analyzer to generate a long semantic description text sequence of the target vehicle.

[0015] In one embodiment, after obtaining the visual input information sequence, this sequence is fed into a pre-built visual semantic analyzer. This analyzer utilizes a neural network model from deep learning to analyze and process the vehicle video frame sequence. Specifically, the visual semantic analyzer receives each frame of image data obtained from the frame extraction process and performs feature extraction and semantic understanding through a neural network model. This neural network model typically includes a convolutional neural network (CNN) or other neural network structures suitable for image processing, capable of image recognition and feature extraction for each frame, extracting important information from the image, such as the vehicle's appearance features, the visibility of the license plate, the license plate number, the vehicle's trajectory, speed, and relative position to other vehicles. Through multi-layer convolution and pooling operations of the convolutional neural network, the network can extract multi-level abstract features from the original image. Subsequently, the visual semantic analyzer performs semantic analysis on these extracted visual features through internal sequence processing models such as Long Short-Term Memory (LSTM) networks or Transformers, transforming the information in each frame of the image into corresponding semantic descriptions, and organizing them into a long semantic description text sequence in chronological order. This long semantic description text sequence includes the vehicle's appearance features, behavioral state, driving process, etc., providing a reliable semantic basis for subsequent abnormal behavior detection, cost management, and risk control analysis.

[0016] The convolutional neural network and long short-term memory network used were pre-trained using sample data. The training steps included forward propagation, loss calculation, backpropagation, and parameter optimization.

[0017] Furthermore, each semantically long descriptive text includes semantic features of vehicle appearance attributes, license plate visibility status, vehicle behavior actions, and vehicle location.

[0018] Preferably, each semantically long descriptive text includes semantic features of vehicle appearance attributes, license plate visibility status, vehicle behavior, and vehicle location. Vehicle appearance attribute semantic features refer to the external features of the vehicle identified by a visual semantic analyzer, including the vehicle's color, brand, model, logo, and specific shape characteristics. For example, it can identify and describe "red sedan" or "blue SUV." This information helps accurately distinguish different vehicles and provides visual recognition basis for vehicle behavior and status. License plate visibility status semantic features refer to the degree of recognition of the vehicle's license plate, typically including categories such as clearly visible, partially obscured, or unrecognizable. Vehicle behavior semantic features refer to the vehicle's behavior and dynamic state within the parking lot, including whether the vehicle is entering, exiting, parking, or driving within the lot. For example, a vehicle can be identified as "entering the entrance" or "exiting the parking lot." The semantic features of vehicle location describe the specific location of a vehicle within a parking lot and its surrounding environment. This includes the parking space number, the area where the vehicle is located (e.g., near the entrance, exit, or deeper within the parking lot), and its relative position to other vehicles. For example, "the vehicle is parked in parking space number 5 in area A" or "the vehicle is near the exit area." This location data not only helps with the spatial positioning of vehicles but also supports the flow management within the parking lot. By concatenating and combining these four parts of information in a predetermined order, a complete semantic long description text can be generated. This provides rich semantic information for subsequent data analysis and strong support for abnormal behavior identification, cost management, and risk control.

[0019] Based on the preset rules for contactless payment parking fees and the set of historical fee anomaly records, a set of questions for querying abnormal vehicle behavior is constructed.

[0020] In one embodiment, the system first reads preset seamless payment parking fee rules and retrieves a set of historical fee anomaly records. These preset rules are predefined standard rules, mainly including vehicle entry and exit rules, license plate recognition, and visibility rules. These rules provide a basic framework for behavioral judgment, helping to identify normal vehicle behavior. For example, they might stipulate that when a vehicle enters the parking lot, its license plate must be clearly recognized; the vehicle must be parked in a valid parking space; and if parking exceeds a predetermined time, additional fees will be automatically calculated according to the rules. This set of historical fee anomaly records contains all abnormal fee records that have occurred in previous parking lot management, such as unsuccessful deductions, duplicate deductions, incorrect parking time calculations, and abnormal vehicle behavior. These historical records serve as a reference to help identify similar abnormal behaviors. For example, if a vehicle's license plate was previously unrecognizable, resulting in incorrect fee settlement, it will be automatically marked as a potential anomaly when similar license plate recognition anomalies are detected subsequently. By aggregating historical fee anomaly records, similar anomalies can be grouped into a single set. This aggregation result is then used to supplement the basic set of vehicle anomaly behavior query questions determined according to preset seamless parking fee payment rules, constructing the final set of vehicle anomaly behavior query questions. This set aims to detect whether vehicle behavior deviates from normal patterns, potentially leading to fee anomalies. Query questions might include phrases such as "Did the vehicle fail to recognize its license plate upon entry?", "Was the vehicle parked in a non-designated parking space?", and "Did the vehicle exit within the specified time?". This set of query questions allows for precise querying and analysis of each target vehicle's behavior, determining the presence of anomalies. This provides data support for subsequent anomaly detection, fee correction, and risk control decisions, ensuring timely identification and handling of potential fee anomalies and behavioral deviations in actual operation.

[0021] Furthermore, the preset rules for contactless payment parking fees include rules for normal vehicle entry and exit, rules for license plate recognition and visibility, rules for parking space usage, and rules for fee deduction.

[0022] Preferably, the preset seamless parking fee payment rules are a series of standard rules set to ensure the standardization of vehicle behavior and the accuracy of fee calculation within the parking lot. These rules cover all aspects of vehicle entry and exit, license plate recognition, parking space usage, and fee deduction. Among them, the normal vehicle entry and exit rules define the behavioral norms that vehicles must follow when entering and exiting the parking lot. This includes ensuring that the license plate is successfully recognized upon entry, and that the license plate information matches the vehicle registration information in the parking lot. Upon exiting the parking lot, it is necessary to verify whether the vehicle has completed parking within the prescribed time limit and successfully settled the fee. If a vehicle is detected to have entered or exited improperly, or if the license plate cannot be recognized upon entry, it will be marked as abnormal and processed accordingly. The license plate recognition and visibility rules are crucial for ensuring accurate identification and verification of vehicle identity. These rules require the license plate recognition system in the parking lot to have good recognition capabilities, accurately reading vehicle license plate information. Simultaneously, the license plate must be clearly visible, without being obscured or blurred. If incomplete or unrecognizable license plate information is detected, supplementary measures will be taken according to the rules, such as re-recognition or triggering manual intervention prompts. Parking space usage rules are management rules for parking spaces, designed to ensure that each vehicle is parked in its designated space and that the use of parking spaces complies with the parking lot's planning and requirements. If a vehicle is parked in an undesignated space or occupies multiple spaces, it will be automatically flagged as abnormal and processed accordingly. Fee deduction rules involve calculating parking duration and settling fees. Based on factors such as parking duration, parking location, and the vehicle owner's membership status, parking fees are automatically calculated and deducted when the vehicle leaves. If a fee calculation error occurs, such as inaccurate parking duration calculation or failure to pay according to the rules, it will be corrected according to preset rules, triggering an exception handling procedure. This includes deducting the incorrect fee or recalculating the fee. Through these preset rules, automated identification, standardized management, and accurate fee settlement of vehicle behavior can be achieved without manual intervention, ensuring the efficiency of parking lot operations and providing car owners with a convenient, transparent, and seamless payment experience, ensuring the accuracy and fairness of parking fees.

[0023] Furthermore, based on the preset rules for contactless payment parking fees and the set of historical abnormal fee records, a set of questions for querying abnormal vehicle behavior is constructed, including: The historical fee anomaly record set is aggregated by category to obtain multiple clustered historical fee anomaly record sets; a basic vehicle anomaly behavior query question set is constructed based on the preset contactless payment parking fee rules; the basic vehicle anomaly behavior query question set is refined and expanded according to the multiple clustered historical fee anomaly record sets to obtain a vehicle anomaly behavior query question set.

[0024] Preferably, to more accurately identify and handle potential abnormal vehicle behavior in parking lots, historical fee anomaly records containing information such as incorrect fee calculations, duplicate charges, payment failures, and license plate recognition anomalies are clustered. That is, anomaly records with similar characteristics are grouped together. For example, all fee anomalies caused by license plate obstruction leading to recognition failure will be grouped together, as will all fee anomalies caused by incorrect parking duration calculations. Through clustering, different types of fee anomaly issues can be identified, and specific processing logic can be established for each type. After clustering, multiple clustered historical fee anomaly record sets will be obtained, each set containing anomaly events with similar characteristics. Subsequently, based on the preset rules for contactless payment parking fees, a basic set of query questions for abnormal vehicle behavior is constructed. This set of questions includes basic query questions for vehicle behavior in parking lot management, with the aim of identifying and detecting behaviors that may lead to abnormal fees. For example, the basic query question set may include questions such as "Did the vehicle enter the parking lot in the prescribed order and was the license plate successfully recognized?" and "Is the vehicle parked in a valid parking space?" These questions are mainly based on preset parking fee rules, such as entry and exit rules, license plate recognition rules, and parking space rules. Through these basic questions, most common abnormal vehicle behaviors can be identified, providing a foundation for further anomaly detection. Subsequently, based on the obtained sets of multiple clustered historical expense anomaly records, the basic query question set is refined and expanded. This process aims to supplement the deficiencies in the basic query question set, especially those questions that have not yet appeared or are too general, to ensure more accurate identification of various potential abnormal behaviors. For example, if clustering of historical expense anomaly records reveals that a certain type of problem, such as expense anomalies caused by license plate obstruction, occurs frequently, then refined questions targeting this problem will be added to the query question set, such as "Is the vehicle's license plate obstructed?", "Does the degree of obstruction affect recognition?", and "Is the license plate information blurred when the vehicle enters the parking lot?". Furthermore, some overly general questions will be refined. For example, the question "Is the vehicle parked in a valid parking space?" might be too general and unable to accurately determine abnormal behavior. After refinement and expansion, the question might become "Is the vehicle parked in a designated area?", "Does the vehicle occupy multiple parking spaces?", and "Is the vehicle parked in a no-parking area?". This refined expansion allows for the addition of more targeted and precise query questions that can accurately detect abnormal behavior, forming a set of abnormal vehicle behavior query questions. This improves the accuracy and effectiveness of abnormal behavior detection, ensures the accuracy of cost calculation, and reduces the risk of errors or missed detections.

[0025] A multimodal question-answering model is invoked to answer the question set of abnormal vehicle behavior queries based on the semantic long description text sequence, thereby obtaining a fine-grained semantic question-answering result set and a question-answer confidence set.

[0026] In one embodiment, after obtaining the semantically long descriptive text sequence, a multimodal question answering model is invoked to process this sequence. A multimodal question answering model is an intelligent model capable of simultaneously processing multiple input data types (such as text, images, audio, etc.). It is typically built upon convolutional neural networks, the Transformer architecture, and vision-language pre-trained models (such as VisualBERT, ViLBERT, LXMERT, etc.). The main input to the model is the previously generated semantically long descriptive text sequence. Based on this text data, the multimodal question answering model can answer a set of questions related to abnormal vehicle behavior queries, obtaining a fine-grained set of semantic question answering results and a set of associated semantically long descriptive fragments. For example, if the query set includes "whether the vehicle is parked in a valid parking space," the model will infer and answer based on information in the semantically long descriptive text, such as the vehicle's location and whether it is parked in a marked parking space. Subsequently, context iterative retrieval and confidence identification are performed on the obtained set of associated semantic long description fragments to obtain a question-and-answer confidence set. This question-and-answer confidence set reflects the credibility of the model for each question-and-answer result, providing a more accurate basis for subsequent abnormal behavior detection, ensuring more accurate abnormal behavior detection, and thus improving the efficiency and accuracy of parking fee management.

[0027] Furthermore, a multimodal question-answering model is invoked based on the semantically long descriptive text sequence to answer the question set regarding abnormal vehicle behavior, obtaining a fine-grained semantic question-answering result set and a question-answer confidence set, including: A multimodal question-answering model is invoked to answer the question set of abnormal vehicle behavior queries based on the semantically long descriptive text sequence, obtaining a fine-grained semantic question-answering result set and a set of associated semantically long descriptive fragments; based on the set of associated semantically long descriptive fragments, context iterative retrieval is performed on the semantically long descriptive text sequence to determine the iterative neighborhood set of associated semantically long descriptive fragments; based on the iterative neighborhood set of associated semantically long descriptive fragments, confidence is identified on the fine-grained semantic question-answering result set to obtain the question-answer confidence set.

[0028] Optionally, a multimodal question-answering model is first invoked to answer a set of questions about abnormal vehicle behavior based on a sequence of semantically long descriptive text. This sequence contains multi-dimensional information about the vehicle's appearance, behavior, and location, providing accurate behavioral context. The multimodal question-answering model combines this information to analyze each query and generate fine-grained answer results. For example, if the question is "Is the vehicle parked for the prescribed duration?", the model will answer based on the specific duration of the vehicle's parking. During this process, the model not only organizes the generated question-answer results into a fine-grained semantic question-answering result set but also identifies semantically long descriptive fragments related to the question-answer results and summarizes them into a set of associated semantically long descriptive fragments. These semantically long descriptive fragments refer to the parts of the semantically long descriptive text closely related to a particular query question; these parts provide supporting information sources for the answer. For example, if the question is related to license plate recognition, the model will associate semantic fragments related to license plate information. Subsequently, based on the obtained set of associated semantic long description fragments, contextual iterative retrieval is performed in the semantic long description text sequence to obtain semantic fragments closely related to the query question. For example, if a query question involves the location of a vehicle's parking space, all description fragments related to "parking location" are searched. Through contextual iterative retrieval, contextual information closely related to the target behavior is captured, forming a semantic long description fragment iterative neighborhood set. This set contains contextual information highly relevant to the query question, helping to more comprehensively understand vehicle behavior and providing support for further confidence identification and anomaly detection. Next, based on the associated semantic long description fragment iterative neighborhood set, confidence identification is performed on the obtained fine-grained semantic question-answering result set. Through analysis of positive consistency statements and negative conflict semantics, the confidence scores of all fine-grained semantic question-answering results are generated and summarized into a question-answer confidence set, which reflects the reliability of the model's answer to each question. If the confidence score of an answer to a question is high, the answer is considered reliable and can be used for further anomaly detection and cost management. Conversely, if the confidence score is low, it may trigger a manual review process or further data validation. In summary, by using a multimodal question-answering model to answer a set of questions about abnormal vehicle behavior, combined with contextual iterative retrieval and confidence level recognition, abnormal vehicle behavior can be identified comprehensively and accurately, improving the intelligence level of parking lot management, enhancing the accuracy of anomaly detection, and ensuring the fairness and transparency of fee settlement.

[0029] Furthermore, based on the iterative neighborhood set of the associated semantic long description fragment, confidence scores are identified for each of the fine-grained semantic question-answering result sets to obtain the question-answering confidence set, including: Extract the first related semantic long description fragment iterative neighborhood and the first fine-grained semantic question-answering result from the related semantic long description fragment iterative neighborhood set and the fine-grained semantic question-answering result set; traverse and compare the support of the first related semantic long description fragment iterative neighborhood to the first fine-grained semantic question-answering result to obtain the number of positively consistent statements and the set of semantic relevance scores, the number of reverse conflicting semantics and the set of semantic conflict degrees; perform confidence weight analysis according to the number of positively consistent statements and the number of reverse conflicting semantics, and combine the analysis results to perform weighted analysis on the set of semantic relevance scores and the set of semantic conflict degrees to obtain the first question-answer confidence score, and add the first question-answer confidence score to the question-answer confidence score set.

[0030] Optionally, a pair of data is first randomly selected from the iterative neighborhood set of the associated semantic long description fragment and the fine-grained semantic question-answering result set, defined as the first iterative neighborhood of the associated semantic long description fragment and the first fine-grained semantic question-answering result. Then, the relationship between the first iterative neighborhood of the associated semantic long description fragment and the first fine-grained semantic question-answering result is traversed and compared to analyze the support of the question-answering results. Support refers to the degree of matching between the question-answering result and the associated semantic long description fragment. During the comparison process, the number of positively consistent statements is counted. Positively consistent statements refer to statements in the question-answering result that are consistent with and support the semantic fragment. For example, if the query is "whether the vehicle is parked for the prescribed period," and the relevant semantic description fragment mentions "the vehicle has been parked for more than 24 hours," then this description is consistent with the question-answering result and belongs to positive consistency. A pre-trained word vector model is then used to convert each semantic description fragment and question-answering result into a vector representation. The semantic relevance score is measured by calculating the cosine similarity between the two text vectors and added to the semantic relevance score set. The closer the semantic relevance score is to 1, the more similar the two texts are semantically. Next, the number of reverse conflict semantics is counted. Reverse conflict semantics refers to statements in the question-and-answer result that contradict or conflict with the related semantic description fragment. For example, if the question-and-answer result is "the vehicle exhibits following behavior," while the related semantic description fragment mentions "the vehicle passes through the barrier alone, without closely following other vehicles," then there is a semantic conflict between the two, which will be marked as a reverse conflict. Then, the semantic description fragment and related keywords or phrases in the question-and-answer result are mapped to a vector space using the same method to obtain their respective word vectors. The cosine similarity between conflicting keywords is then calculated to measure the degree of semantic conflict and added to the semantic conflict degree set. The closer this semantic conflict degree is to -1, the greater the semantic contradiction between the two words or phrases. Then, confidence weight analysis is performed based on the number of positively consistent statements and the number of reverse conflict semantics. That is, the number of positively consistent statements is divided by the total number of both, and the number of reverse conflict semantics is divided by the total number of both, to obtain the corresponding confidence weights. These calculated confidence weights are then used in conjunction with the semantic relevance score set and the semantic conflict degree set for weighted analysis to obtain a comprehensive first question-and-answer confidence score. By adding the first question-and-answer confidence score to the question-and-answer confidence score set, the final confidence score data is formed, thereby providing accurate decision support for subsequent anomaly detection and cost management, ensuring the fairness of cost calculation and the efficiency of platform operation.

[0031] By combining the fine-grained semantic question-answering result set and the question-answering confidence set, a multi-node relationship heterogeneous graph is constructed.

[0032] In one embodiment, after obtaining the fine-grained semantic question-answering result set and the question-answering confidence set, the semantic elements in the fine-grained semantic question-answering result set are used as nodes, and the identification information in the question-answering confidence set is used to label these nodes. Then, topological connections are made based on the relationships between semantic elements to construct a multi-node heterogeneous graph. Each node in this multi-node heterogeneous graph represents a semantic element, such as important semantic features like vehicle behavior, vehicle status, parking space information, and fee information. These nodes are connected by relational edges to represent their semantic associations. For example, if a vehicle's license plate cannot be recognized, the "vehicle" node is connected to the "license plate recognition failed" node in the graph; if a vehicle is parked in an illegal parking space, the "vehicle parking behavior" node is connected to the "illegal parking space" node. This multi-node heterogeneous graph not only helps in understanding vehicle behavior but also provides accurate support for subsequent abnormal behavior detection and fee management, thereby improving the intelligence level of parking lot management.

[0033] Furthermore, by combining the aforementioned fine-grained semantic question-answering result set and question-answering confidence set for analysis, a multi-node heterogeneous relationship graph is constructed, including: Structured semantic elements are extracted from the fine-grained semantic question-answering result set, and each extracted semantic element is treated as a node to obtain a semantic element node set; the semantic element node set is identified using the question-answering confidence set to obtain an identified semantic element node set; based on the association between semantic elements, a relational topology is performed on the identified semantic element node set to obtain the multi-node relational heterogeneous graph.

[0034] Preferably, structured semantic elements are extracted from the fine-grained semantic question-answering result set. Fine-grained semantic question-answering results typically contain multi-level information, such as vehicle behavior, status, time, and location. By extracting this information, lexical units with clear semantics can be obtained and used as semantic elements. For example, assuming a question-answering result is "The vehicle was parked in parking space number 5 in area A for more than 2 hours," structured semantic elements such as vehicle, parking space, parking duration, parking area, and exceeding the time limit will be extracted. After these structured semantic elements are extracted, they are used as semantic element nodes and aggregated into a semantic element node set, where each node contains a specific description of the vehicle behavior. Subsequently, these nodes are labeled according to a question-answering confidence set, which contains a credibility score for each question-answering result, reflecting the reliability and accuracy of the result. Based on the correspondence between these question-answering confidence scores and the question-answering results, these confidence scores can be labeled onto the corresponding nodes, forming a set of labeled semantic element nodes that indicates their reliability. Next, a relational topology is constructed based on the relationships between semantic element nodes. These relationships refer to the semantic connections or logical relationships between different semantic elements. For example, there is a "parked in" relationship between "vehicle" and "parking space," and a "whether timeout occurred" relationship between "parking duration" and "overtime behavior." Based on these relationships, relational edges can be defined between semantic element nodes, representing logical connections between different nodes. Through these relational edges, all semantic element nodes can be linked together to form a multi-node heterogeneous relational graph. This multi-node heterogeneous relational graph comprehensively displays various information about vehicle behavior within the parking lot through semantic element nodes and their relationships, providing reliable data support for subsequent abnormal behavior detection and fee management.

[0035] Anomaly analysis is performed on the heterogeneous graph of the multi-node relationship. If the analysis results show abnormal behavior, anomaly warning information is obtained, and cost management is carried out based on the anomaly warning information.

[0036] In one embodiment, after constructing a heterogeneous graph of multi-node relationships, abnormal behavior identification is performed on each semantic element node in the graph. By analyzing information such as vehicle parking duration, license plate recognition, and parking location, it is determined whether these behaviors comply with the normal management rules of the parking lot. If a node displays an anomaly, it is marked as abnormal, and a heterogeneous subgraph of multi-node relationships related to the anomaly is extracted. This subgraph contains all nodes related to the abnormal behavior and their interrelationships. Subsequently, the extracted abnormal behavior subgraphs are integrated, merging all relevant abnormal information to form a preliminary abnormal behavior integration result. Then, based on the heterogeneous graph of multi-node relationships, the preliminary abnormal behavior integration result is corrected to obtain the final abnormal behavior analysis result, and an abnormal warning message is automatically generated. This warning message describes in detail the abnormal behavior that occurred and its potential impact on parking lot management. For example, it may indicate that a vehicle has been parked for more than the prescribed time, or that a vehicle's license plate cannot be recognized, potentially indicating license plate cloning or fraudulent behavior. Following this anomaly alert, fee management is implemented. If an anomaly is detected in a vehicle's parking fee calculation, the fee will be corrected based on the alert. For example, if a vehicle's parking fee is not charged on time due to license plate recognition failure, the fee will be recalculated and deducted correctly. For other fee issues arising from abnormal behavior, parking fees will be automatically adjusted, and additional fees will be charged to the offending vehicle according to the parking lot's charging rules. If a vehicle is found to frequently exceed the parking time limit or park in illegal areas, higher fees or no-parking measures can be implemented for that vehicle to ensure the reasonableness and accuracy of parking fees.

[0037] Furthermore, abnormal behavior analysis is performed on the aforementioned multi-node heterogeneous graph. If abnormal behavior is found in the analysis results, abnormal warning information is obtained, and cost management is performed based on the abnormal warning information, including: Abnormal behavior is identified by traversing each semantic element node in the heterogeneous multi-node relationship graph, and an abnormal behavior heterogeneous multi-node relationship subgraph is extracted. Abnormal behavior is integrated based on the abnormal behavior heterogeneous multi-node relationship subgraph to obtain an initial abnormal behavior integration result. The initial abnormal behavior integration result is corrected according to the heterogeneous multi-node relationship graph to obtain the analysis result.

[0038] Preferably, each semantic element node in the heterogeneous graph of multi-node relationships is first traversed. These nodes represent various semantic elements of vehicle behavior, such as the vehicle's parking status, parking location, and parking duration. Each node is also connected to other nodes through relational edges, representing the logical relationships between these semantic elements. For each traversed node, abnormal behavior identification is performed based on its surrounding relational edges. The goal of abnormal behavior identification is to identify behaviors that do not conform to predetermined parking rules or have potential problems. For example, if the vehicle behavior represented by a node exceeds the preset parking duration, or if the vehicle is parked in a no-parking area, the behavior will be marked as abnormal. After identifying abnormal behavior, a heterogeneous subgraph of multi-node relationships related to that abnormal behavior is extracted. This subgraph includes all semantic element nodes related to the abnormal behavior and the edges connecting them. For example, if overstaying parking is identified as an abnormal behavior, all nodes related to "overstaying parking," such as "parking duration" and "parking area," are extracted and connected to form a subgraph. This heterogeneous subgraph reflects the specific semantic and contextual information of the abnormal behavior. The nodes and edges in the subgraph also carry relevant question-answer confidence information, helping to determine the reliability of different nodes and edges in the subsequent abnormal behavior integration process. Then, the obtained heterogeneous subgraph of multi-node relationships is used for abnormal behavior integration. The purpose of integration is to combine the abnormal information between different nodes and edges to form a more accurate abnormal behavior identification result. During the abnormal behavior integration process, weighted integration is performed based on the question-answer confidence of the nodes. Specifically, the question-and-answer confidence score of each node and edge is used as a weighting factor. The influence of each node in the integration process is adjusted based on its confidence score. For example, if a node's anomalous behavior has a high confidence score (e.g., above 90%), that node has a greater impact on the final anomalous behavior integration result; conversely, if a node's confidence score is low (e.g., below 60%), its weight in the integration process will be smaller, or it may even be excluded from the integration result. This confidence-based weighted integration effectively improves the accuracy of anomalous behavior identification and avoids low-confidence anomalous behaviors affecting the final decision. After the integration process is completed, an initial anomalous behavior integration result is obtained, which contains information on various anomalous behaviors. Then, the initial abnormal behavior integration results are corrected based on the entire multi-node heterogeneous graph. The purpose of the correction is to ensure the accuracy and consistency of abnormal behavior identification. The correction process can be carried out through global semantic consistency verification and question-answer confidence of conflicting nodes, thereby obtaining an accurate analysis result. This analysis result includes the final identification results of all abnormal behaviors, as well as the confidence score of each abnormal behavior, providing an important basis for subsequent cost management, abnormal warning, etc.

[0039] Furthermore, the initial abnormal behavior integration results are corrected based on the aforementioned multi-node heterogeneous graph to obtain analysis results, including: Based on a multi-node heterogeneous graph, a global semantic consistency check is performed on the initial abnormal behavior integration result to obtain a semantic consistency score. If the semantic consistency score is greater than or equal to a preset threshold, no correction is made, and the initial abnormal behavior integration result is used as the analysis result. If the semantic consistency score is less than the preset threshold, conflict nodes are extracted, and the initial abnormal behavior integration result is corrected based on the question-answer confidence of the conflict nodes to obtain the analysis result.

[0040] Optionally, when correcting the initial abnormal behavior integration results, a global semantic consistency check is first performed on the initial abnormal behavior integration results based on the multi-node relationship heterogeneous graph. This global semantic consistency check includes cross-node consistency checks and cross-edge relationship checks. The purpose of the cross-node consistency check is to check whether the abnormal behavior conflicts with the semantics of multiple associated nodes; the purpose of the cross-edge relationship check is to check whether the initial abnormal behavior is consistent with the semantic direction of multiple relationship edges. When performing the cross-node consistency check, each node in the multi-node relationship heterogeneous graph is traversed, with particular attention paid to the context nodes related to the abnormal behavior. For example, if a vehicle's parking behavior is found to exceed the specified time, the relationship between this behavior and time nodes (such as parking duration), device nodes (such as license plate recognition device status), and location nodes (such as parking area) is checked. If the data of multiple context nodes are consistent with the abnormal behavior, then the abnormal behavior will be marked as "strongly supported" and assigned a corresponding score, meaning that the behavior is semantically consistent with other nodes. Conversely, if a semantic contradiction is found between the abnormal behavior and certain context nodes—for example, a mismatch between parking behavior and parking duration, or a conflict between device status and recognition results—the abnormal behavior will be marked as "conflict pending correction" and assigned a corresponding score, indicating a potential problem in the recognition of the behavior that requires further verification or correction. During cross-edge relationship verification, it is checked whether the initial integrated result of the abnormal behavior is consistent with the attributes of these edges. For example, if a vehicle's parking behavior is marked as overstaying, but it is found that the behavior does not match the device status of the "vehicle-device" edge (e.g., license plate recognition failure), or is inconsistent with the parking location of the "behavior-location" edge, then the confidence level of the abnormal behavior is considered low, and further correction may be needed. If the initial abnormal behavior is inconsistent with the attributes of multiple edges, the confidence level of the abnormal behavior will be reduced, and this information will be used to determine whether further correction of the abnormal behavior's recognition result is required. After completing the cross-node consistency check and cross-edge relationship verification, the node consistency score and edge consistency score are weighted and fused to obtain a semantic consistency score, representing the overall consistency of the abnormal behavior's recognition result. Next, the calculated semantic consistency score is compared with a preset threshold. If the semantic consistency score is greater than or equal to the preset threshold, it indicates that the anomalous behavior has high semantic consistency with multiple nodes and relationships, and the identification result of the anomalous behavior is considered reliable, requiring no correction. The initial integrated result of the anomalous behavior is directly used as the final analysis result. If the semantic consistency score is less than the preset threshold, it is considered that the anomalous behavior may have semantic contradictions or conflicts, requiring further correction. In this case, the conflict node extraction and correction process will begin, that is, extracting conflict nodes, which are those nodes that contradict other nodes or relationship edges during the semantic consistency verification process. These conflict nodes may be nodes whose behavioral descriptions do not match those of other associated nodes, or nodes whose semantic direction is inconsistent with the edge relationship.For each conflict node, the algorithm will be revised based on its question-and-answer confidence score. Question-and-answer confidence score indicates the platform's level of trust in the node or its behavior. During the revision process, if a conflict node has a high question-and-answer confidence score, it indicates that the node's identification result is more reliable, and the node will be given a larger weight, prioritizing its impact on the overall abnormal behavior identification result. Conflict nodes with low confidence scores will be treated as a weaker factor and adjusted appropriately. Then, the abnormal behavior identification result is re-evaluated using a weighted average based on the adjusted conflict node weights. If the revision of the conflict nodes makes the abnormal behavior identification result more consistent with the semantic relationships of other nodes and edges in the graph, the abnormal behavior identification result will be updated, forming the revised final analysis result. This revised result will more accurately reflect the actual vehicle behavior in the parking lot, thus providing a more reliable basis for subsequent anomaly detection and fee management, and improving the intelligence level of parking lot management in anomaly detection and fee management.

[0041] In summary, the embodiments of this application have at least the following technical effects: First, a set of vehicle video feeds collected by cameras at the parking lot entrance, exit, and inside the lot is acquired. Frames are extracted from these video feeds to obtain the visual input information sequence of the target vehicle. Then, the visual input information sequence is analyzed using a visual semantic analyzer to generate a semantic long-description text sequence for the target vehicle. Next, a set of questions for querying abnormal vehicle behavior is constructed based on preset seamless parking fee payment rules and a set of historical fee anomaly records. Then, a multimodal question-answering model is invoked to answer the questions based on the semantic long-description text sequence, obtaining a fine-grained semantic question-answering result set and a question-answer confidence set. Then, the fine-grained semantic question-answering result set and question-answer confidence set are analyzed to construct a multi-node heterogeneous graph. Finally, abnormal behavior analysis is performed on the multi-node heterogeneous graph. If abnormal behavior is found in the analysis results, an anomaly warning is obtained, and fee management is performed based on the anomaly warning information. This solution addresses the technical challenges of inaccurate abnormal behavior detection and difficulty in tracing in existing contactless payment parking scenarios, caused by superficial understanding of visual data and isolated multi-source information. It achieves the technical effect of improving the accuracy of vehicle behavior recognition and the security of parking fee management in parking lots through deep semantic understanding and multi-relationship graph analysis.

[0042] Example 2, based on the same inventive concept as the contactless payment parking fee management method in the previous examples, such as... Figure 2 As shown, this application provides a contactless payment parking fee management platform, which includes: Video Acquisition Module 11: Acquires a set of vehicle videos captured by cameras at the parking lot entrance, exit, and inside the parking lot; extracts frames from the vehicle video set to obtain a visual input information sequence for the target vehicle; Video Analysis Module 12: Analyzes the visual input information sequence based on a visual semantic analyzer to generate a semantic long description text sequence for the target vehicle; Query Question Construction Module 13: Constructs a set of query questions for abnormal vehicle behavior according to preset seamless payment parking fee rules and a set of historical fee anomaly records; Semantic Question Answering Module 14: Calls a multimodal question answering model to answer the query questions for abnormal vehicle behavior based on the semantic long description text sequence, obtaining a set of fine-grained semantic question answering results and a set of question answering confidence scores; Question Answering Result Analysis Module 15: Combines the set of fine-grained semantic question answering results and the set of question answering confidence scores for analysis to construct a multi-node heterogeneous graph; Fee Management Module 16: Performs abnormal behavior analysis on the multi-node heterogeneous graph; if abnormal behavior is found in the analysis results, obtains abnormal warning information and manages fees based on the abnormal warning information.

[0043] Furthermore, the video analysis module 12 is used to perform the following methods: Each semantic long description text includes semantic features of vehicle appearance attributes, license plate visibility, vehicle behavior and actions, and vehicle location.

[0044] Furthermore, the query question construction module 13 is used to execute the following method: The preset rules for contactless payment parking fees include rules for normal vehicle entry and exit, rules for license plate recognition and visibility, rules for parking space usage, and rules for fee deduction.

[0045] Furthermore, the query question construction module 13 is used to execute the following method: The historical fee anomaly record set is aggregated by category to obtain multiple clustered historical fee anomaly record sets; a basic vehicle anomaly behavior query question set is constructed based on the preset contactless payment parking fee rules; the basic vehicle anomaly behavior query question set is refined and expanded according to the multiple clustered historical fee anomaly record sets to obtain a vehicle anomaly behavior query question set.

[0046] Furthermore, the semantic question-answering module 14 is used to perform the following methods: A multimodal question-answering model is invoked to answer the question set of abnormal vehicle behavior queries based on the semantically long descriptive text sequence, obtaining a fine-grained semantic question-answering result set and a set of associated semantically long descriptive fragments; based on the set of associated semantically long descriptive fragments, context iterative retrieval is performed on the semantically long descriptive text sequence to determine the iterative neighborhood set of associated semantically long descriptive fragments; based on the iterative neighborhood set of associated semantically long descriptive fragments, confidence is identified on the fine-grained semantic question-answering result set to obtain the question-answer confidence set.

[0047] Furthermore, the semantic question-answering module 14 is used to perform the following methods: Extract the first related semantic long description fragment iterative neighborhood and the first fine-grained semantic question-answering result from the related semantic long description fragment iterative neighborhood set and the fine-grained semantic question-answering result set; traverse and compare the support of the first related semantic long description fragment iterative neighborhood to the first fine-grained semantic question-answering result to obtain the number of positively consistent statements and the set of semantic relevance scores, the number of reverse conflicting semantics and the set of semantic conflict degrees; perform confidence weight analysis according to the number of positively consistent statements and the number of reverse conflicting semantics, and combine the analysis results to perform weighted analysis on the set of semantic relevance scores and the set of semantic conflict degrees to obtain the first question-answer confidence score, and add the first question-answer confidence score to the question-answer confidence score set.

[0048] Furthermore, the question-and-answer result analysis module 15 is used to perform the following methods: Structured semantic elements are extracted from the fine-grained semantic question-answering result set, and each extracted semantic element is treated as a node to obtain a semantic element node set; the semantic element node set is identified using the question-answering confidence set to obtain an identified semantic element node set; based on the association between semantic elements, a relational topology is performed on the identified semantic element node set to obtain the multi-node relational heterogeneous graph.

[0049] Furthermore, the cost management module 16 is used to perform the following methods: Abnormal behavior is identified by traversing each semantic element node in the heterogeneous multi-node relationship graph, and an abnormal behavior heterogeneous multi-node relationship subgraph is extracted. Abnormal behavior is integrated based on the abnormal behavior heterogeneous multi-node relationship subgraph to obtain an initial abnormal behavior integration result. The initial abnormal behavior integration result is corrected according to the heterogeneous multi-node relationship graph to obtain the analysis result.

[0050] Furthermore, the cost management module 16 is used to perform the following methods: Based on a multi-node heterogeneous graph, a global semantic consistency check is performed on the initial abnormal behavior integration result to obtain a semantic consistency score. If the semantic consistency score is greater than or equal to a preset threshold, no correction is made, and the initial abnormal behavior integration result is used as the analysis result. If the semantic consistency score is less than the preset threshold, conflict nodes are extracted, and the initial abnormal behavior integration result is corrected based on the question-answer confidence of the conflict nodes to obtain the analysis result.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for managing parking fees through contactless payment, characterized in that: The method includes: Acquire a set of vehicle videos collected by cameras at the entrance, exit, and inside the parking lot; extract frames from the vehicle video set to obtain a sequence of visual input information for the target vehicle. The visual input information sequence is analyzed using a visual semantic analyzer to generate a long semantic description text sequence of the target vehicle. Based on the preset rules for contactless payment parking fees and the set of historical abnormal fee records, a set of questions for querying abnormal vehicle behavior is constructed. A multimodal question-answering model is invoked to answer the question set of abnormal vehicle behavior queries based on the semantic long description text sequence, thereby obtaining a fine-grained semantic question-answering result set and a question-answering confidence set; By combining the fine-grained semantic question-answering result set and the question-answering confidence set, a multi-node relationship heterogeneous graph is constructed. Anomaly analysis is performed on the heterogeneous graph of the multi-node relationship. If the analysis results show abnormal behavior, anomaly warning information is obtained, and cost management is carried out based on the anomaly warning information.

2. The contactless payment parking fee management method as described in claim 1, characterized in that, Each semantic long description text includes semantic features of vehicle appearance attributes, license plate visibility, vehicle behavior and actions, and vehicle location.

3. The contactless payment parking fee management method as described in claim 1, characterized in that, The preset rules for contactless payment parking fees include rules for normal vehicle entry and exit, rules for license plate recognition and visibility, rules for parking space usage, and rules for fee deduction.

4. The contactless payment parking fee management method as described in claim 1, characterized in that, Based on the preset rules for contactless payment parking fees and the set of historical abnormal fee records, a set of questions for querying abnormal vehicle behavior is constructed, including: The historical expense anomaly record set is aggregated by type to obtain multiple clustered historical expense anomaly record sets; A basic set of questions for querying abnormal vehicle behavior is constructed based on the preset rules for contactless payment parking fees. The basic set of questions for querying abnormal vehicle behavior is refined and expanded based on the multiple clustered sets of historical fee anomaly records to obtain a new set of questions for querying abnormal vehicle behavior.

5. The contactless payment parking fee management method as described in claim 1, characterized in that, A multimodal question-answering model is invoked to answer the set of questions related to abnormal vehicle behavior queries based on the semantically long descriptive text sequence, obtaining a fine-grained semantic question-answering result set and a question-answer confidence set, including: A multimodal question-answering model is invoked to answer the question set of abnormal vehicle behavior queries based on the semantic long description text sequence, thereby obtaining a fine-grained semantic question-answering result set and a set of associated semantic long description fragments; Based on the set of associated semantic long description fragments, context iterative retrieval is performed on the semantic long description text sequence to determine the iterative neighborhood set of associated semantic long description fragments; Based on the iterative neighborhood set of the associated semantic long description fragment, confidence scores are identified for the fine-grained semantic question-answering result set to obtain the question-answer confidence set.

6. The contactless payment parking fee management method as described in claim 5, characterized in that, Based on the iterative neighborhood set of the associated semantic long description fragment, confidence scores are identified for each of the fine-grained semantic question-answering result sets to obtain the question-answering confidence set, including: Extract the first iterative neighborhood of the associated semantic long description fragment and the first fine-grained semantic question answering result from the iterative neighborhood set of the associated semantic long description fragment and the fine-grained semantic question answering result set; The support of the first fine-grained semantic question answering result to the long description fragment of the first associated semantics is traversed and compared in the neighborhood. The number of positively consistent statements and the set of semantic relevance scores, the number of reverse conflicting semantics and the set of semantic conflict degree are obtained. Confidence weight analysis is performed based on the number of normal consistency statements and the number of reverse conflict semantics. The semantic relevance score set and semantic conflict score set are then weighted and analyzed in combination with the analysis results to obtain the first question-and-answer confidence score. The first question-and-answer confidence score is then added to the question-and-answer confidence score set.

7. The contactless payment parking fee management method as described in claim 1, characterized in that, By combining the aforementioned fine-grained semantic question-answering result set and question-answering confidence set for analysis, a multi-node relationship heterogeneous graph is constructed, including: Structured semantic elements are extracted from the fine-grained semantic question-answering result set, and each extracted semantic element is treated as a node to obtain a semantic element node set; The semantic element node set is identified using the question-answer confidence set to obtain the identified semantic element node set. Based on the relationships between semantic elements, a relational topology is performed on the set of identifier semantic element nodes to obtain the multi-node relational heterogeneous graph.

8. The contactless payment parking fee management method as described in claim 1, characterized in that, Anomaly analysis is performed on the heterogeneous graph of multi-node relationships. If abnormal behavior is found in the analysis results, anomaly warning information is obtained, and cost management is performed based on the anomaly warning information, including: Traverse each semantic element node in the heterogeneous multi-node relationship graph to identify abnormal behavior and extract the heterogeneous multi-node relationship subgraph of abnormal behavior; Anomalies are integrated based on heterogeneous subgraphs of multi-node relationships of anomalous behaviors to obtain initial anomalous behavior integration results. The initial abnormal behavior integration results are corrected based on the multi-node heterogeneous relationship graph to obtain the analysis results.

9. The contactless payment parking fee management method as described in claim 8, characterized in that, The initial abnormal behavior integration results are corrected based on the multi-node heterogeneous relationship graph to obtain the analysis results, including: Based on the heterogeneous graph of multi-node relationships, a global semantic consistency check is performed on the initial abnormal behavior integration result to obtain a semantic consistency score. If the semantic consistency score is greater than or equal to a preset threshold, no correction is made, and the initial abnormal behavior integration result is used as the analysis result. If the semantic consistency score is less than the preset threshold, conflict nodes are extracted, and the initial abnormal behavior integration results are corrected by combining the question-answer confidence of the conflict nodes to obtain the analysis results.

10. A contactless payment parking fee management platform, characterized in that: The platform is used to implement the contactless payment parking fee management method according to any one of claims 1-9, and the platform includes: Video acquisition module: acquires a set of vehicle videos collected by cameras at the parking lot entrance, exit, and inside the parking lot, and extracts frames from the vehicle video set to obtain the visual input information sequence of the target vehicle; Video analysis module: Analyzes the visual input information sequence based on a visual semantic analyzer to generate a long semantic description text sequence of the target vehicle; The query question construction module constructs a set of query questions for abnormal vehicle behavior based on the preset rules for contactless payment parking fees and the set of historical fee exception records. Semantic question answering module: Invokes a multimodal question answering model to answer the question set of abnormal vehicle behavior queries based on the semantic long description text sequence, and obtains a fine-grained semantic question answering result set and a question answering confidence set; Question answering result analysis module: Combines the fine-grained semantic question answering result set and question answering confidence set for analysis, and constructs a multi-node relationship heterogeneous graph; Cost Management Module: Performs abnormal behavior analysis on the heterogeneous graph of multi-node relationships. If abnormal behavior is found in the analysis results, abnormal warning information is obtained, and cost management is performed based on the abnormal warning information.

Images