Deep learning-based multi-layer rate charging method for cross-garage parking in parking lot

Abstract

The invention discloses a multi-layer rate charging method for cross-garage parking in a parking lot based on deep learning, and the method comprises the steps: collecting multi-source original data generated by a vehicle at entrances and exits of a plurality of garages, and generating a feature vector set; inputting the feature vector set into an improved product unit residual network model, and generating a cross-library roaming parking session initial record based on a session aggregation result; obtaining a cross-garage roaming parking session cleaning record; constructing a charging rule map, and generating an amount change record on each charging rule layer; executing global path search in the charging rule atlas according to the amount change record to obtain a final charging result; and executing charging rule atlas complete calculation, checking and complementing a temporary charging result, and completing cross-library account checking and settlement entry operations to form a fare settlement result. The cross-garage parking behavior identification method can significantly reduce the probability of error aggregation and error splitting in cross-garage parking behavior identification, and reduces the phenomena of repeated charging and missed counting.

Description

Technical Field

[0001] This invention relates to the field of parking technology, and in particular to a multi-level rate billing method for cross-garage parking within a parking lot based on deep learning. Background Technology

[0002] With the continuous development of intelligent urban transportation and multi-format joint-operation parking lots, the movement of vehicles across multiple independent garages or parking areas has become an increasingly common parking behavior. In existing parking management systems, session recognition and multi-level fee calculation for cross-garage parking mainly rely on fixed rules or static parameters for segmented splicing and fee calculation. Common practices include using time windows, spatial distance thresholds, or simple entry and exit record comparisons to determine the start and end of a parking behavior, and directly implementing the fee rules in the business logic code in the form of nested or sequential judgments.

[0003] With the increasing complexity of parking lot layouts and the growing uncertainty of the recognition environment, existing technologies face numerous challenges in cross-database session aggregation. On the one hand, aggregation methods based on static thresholds are prone to mis-aggregation or mis-splitting under conditions of fluctuating recognition confidence and diverse vehicle behavior, resulting in the same parking process being split into multiple billing segments, or multiple short pauses being merged, leading to missed billing.

[0004] Existing systems often mix basic duration billing, time-period price differences, and cross-database settlement into a single process or code segment when handling multi-level rates, forming a highly coupled rule chain. This makes it difficult to meet the requirements of rule maintenance, hot updates, canary releases, and dispute tracing. Once the operational strategy or regulatory requirements change, the relevant code needs to be completely reconstructed, which can easily lead to errors and compliance risks. It also cannot effectively support the automatic selection of the optimal cost path and full-process auditing in complex scenarios. Summary of the Invention

[0005] One objective of this invention is to propose a multi-level rate billing method for cross-garage parking within a parking lot based on deep learning. This invention can significantly reduce the probability of mis-aggregation and mis-splitting in cross-garage parking behavior recognition, and reduce duplicate billing and omissions.

[0006] A method for calculating multi-level parking rates across parking garages based on deep learning according to an embodiment of the present invention includes: Collect multi-source raw data generated by vehicles at multiple garage entrances and exits, convert it into a structured spatiotemporal recognition dataset, preprocess the spatiotemporal recognition dataset, and generate a set of feature vectors. The feature vector set is input into the improved product unit residual network model to obtain the session aggregation result and the corresponding confidence threshold, and the initial record of cross-database roaming parking session is generated based on the session aggregation result. Based on the initial records of cross-garage roaming parking sessions, conflict resolution processing is performed to remove multi-garage conflict records within the same time window, and anomaly detection processing is performed to screen out suspected license plate replacement anomaly records and obstruction anomaly records, resulting in cleaned cross-garage roaming parking session records. A billing rule graph is constructed, which divides the billing rules into a basic time-based billing rule layer, a time-area adjustment rule layer, and a cross-depot settlement rule layer. The cross-depot roaming parking session cleaning record is used as input. The billing rule layers of the billing rule graph are triggered in sequence to perform rule subgraph matching calculations, thereby obtaining the basic time-based billing amount, the time-area adjustment billing amount, and the cross-depot merged billing amount. A record of amount change is generated in each billing rule layer. Based on the amount change record, a global path search is performed in the billing rule graph to determine the optimal cost path that meets regulatory requirements and operational strategies, and the final billing result is obtained. At the edge computing node, the improved product unit residual network model inference is executed to generate temporary billing results and some preprocessing results of billing rules. The complete calculation of the billing rule graph is then performed, the temporary billing results are checked and supplemented, and cross-database reconciliation and settlement are completed to form the fare settlement result.

[0007] Optionally, the multi-source raw data includes license plate recognition image sequences, video trajectory data, and parking lot topology information, and the feature vector set includes entry and exit time interval features, inter-parking space connectivity distance features, and recognition confidence features.

[0008] Optionally, the improved product unit residual network model includes: The feature vector set is normalized to obtain the normalized input vector set; For each normalized input vector, a cross-garage connectivity kernel is constructed based on the topological and continuity characteristics of the parking lot cross-garage recognition scenario; In the improved product unit residual network model structure of the first layer, the main response of the k-th product unit neuron is calculated, and the arithmetic mean of the exponential parameters of the main responses of all product unit neurons in the first layer is taken to obtain the layer sensitivity coefficient of the first layer. The sensitivity coefficient of the first layer is input into the gated residual structure to calculate the dynamic gate factor of the first layer, and the output feature vector of the residual block of the first layer product unit is calculated based on the dynamic gate factor of the first layer. For the product unit neural network structure from the 2nd to the Lth layer, the same calculation method as the 1st layer is used to obtain the output feature vector and layer sensitivity coefficient of the product unit residual block of each layer. The arithmetic mean of the layer sensitivity coefficients of all layers is taken to obtain the full layer sensitivity coefficient. Calculate the pairwise normalized entry and exit time interval characteristics, the pairwise normalized inter-library spatial connectivity distance characteristics, the pairwise normalized identification confidence characteristics, and the pairwise full-layer sensitivity coefficient; The session pairwise uncalibrated score is calculated by outputting the feature vector of the residual block of the product unit. The temperature calibration factor is generated based on the confidence feature of the pairwise normalization and the pairwise full-layer sensitivity coefficient. The session pairwise uncalibrated score is then calibrated based on the temperature calibration factor to obtain the session pairwise aggregation probability value. Based on the pairwise normalized entry and exit time interval features, the pairwise normalized inter-library spatial connectivity distance features, the pairwise normalized identification confidence features, and the pairwise full-layer sensitivity coefficient, an adaptive session aggregation threshold is generated by calculating the parameterized Sigmoid function. The probability value of each pair of sessions is compared with the corresponding adaptive session aggregation threshold. When the probability value of the pair of sessions is greater than or equal to the adaptive session aggregation threshold, the two identified records are identified as belonging to the same cross-garage parking session. All identified records are established with an undirected adjacency relationship according to this rule. The connected components of the adjacency relationship are extracted. Each connected component represents a set of initial records of cross-garage roaming parking sessions aggregated by sessions.

[0009] Optionally, the anomaly detection process to filter out suspected license plate replacement and obstruction anomaly records includes: Time alignment is performed on the initial record of each cross-garage roaming parking session, and the time span of the initial record of each cross-garage roaming parking session is defined as the interval from the earliest entry timestamp to the latest exit timestamp of all identified segments. For any two different initial records of cross-garage roaming parking sessions, if the time span of the two records overlaps and the license plate number strings of the two records are the same, then it is determined that the two initial records of cross-garage roaming parking sessions have a cross-garage session time overlap conflict. For each pair of cross-database roaming parking session initial records that have cross-database session time overlap conflicts, calculate the spatiotemporal consistency score; For each pair of cross-database session time overlap conflict records, a consistency threshold judgment is performed to obtain the intermediate cross-database parking session set; For each session record in the intermediate cross-garage parking session set after conflict resolution, occlusion anomaly detection and suspected license plate replacement detection are performed, and the corresponding recognition record sequence is extracted. Each recognition record is treated as an independent license plate recognition segment. Extract the recognition confidence score of the bar recognition record image and calculate the recognition confidence variance of the session record; When the identification confidence variance is greater than or equal to the occlusion anomaly identification variance threshold, and there are at least three consecutive frames of identification segments in the session record with identification confidence lower than the confidence threshold, the cross-garage parking session is determined to have an occlusion anomaly and is removed or split into multiple session records. Compare the license plate character sequences of any two recognition records, calculate the edit distance between them, and divide the edit distance by the larger value of the lengths of the two license plate strings to obtain the shortest edit distance ratio. When the shortest edit distance ratio is greater than or equal to the license plate replacement anomaly ratio threshold, the corresponding session record is determined to be a suspected license plate replacement anomaly record. Delete all session records that meet the conditions for occlusion anomaly or license plate change anomaly, and output the remaining session records as the final cross-garage roaming parking session cleaning record.

[0010] Optionally, the consistency threshold judgment rule includes: When the spatiotemporal consistency score is less than the consistency threshold, it is judged as a pseudo-conflict with temporal overlap but spatial unreachability, and the initial record of the cross-database roaming parking session with a low spatiotemporal consistency score is deleted. When the spatiotemporal consistency score is greater than or equal to the consistency threshold, retain the two initial records of the cross-database roaming parking session and mark them as tolerable short-term overlap; All retained initial records of cross-garage roaming parking sessions form an intermediate set of cross-garage parking sessions after conflict resolution.

[0011] Optionally, constructing the billing rule graph includes: Construct a billing rule graph, taking cross-database roaming parking session cleaning records as input, and triggering the basic time-based billing rule layer, time-area adjustment rule layer, and cross-database settlement rule layer in hierarchical order. Execute the corresponding rule subgraph matching under each level, and output the amount change record and rule hit sequence. Perform rule subgraph matching of the basic time-based billing rule layer on the cross-garage roaming parking session cleaning records to calculate the basic time-based billing amount; Based on the basic time-based billing amount, perform rule subgraph matching of the time-based area adjustment rule layer to obtain the billing amount after time-based area adjustment; Based on the billing amount after time period and regional adjustment, the rule subgraph matching of the cross-database settlement rule layer is performed to generate the cross-database merged billing amount; The basic time-based billing amount, the billing amount after time-area adjustment, and the cross-database merged billing amount are summarized in hierarchical order to form a complete billing hierarchy. At the same time, the rule hit sequence and amount change records of all levels are combined in sequence to form a complete execution trajectory of the billing rule graph.

[0012] Optionally, the rule subgraph matching of the basic time-of-use billing rule layer includes: The difference between the earliest entry time stamp and the latest exit time stamp is taken as the effective dwell time of the parking session. The effective dwell time is divided into basic time segments according to the basic time unit. The corresponding basic billing unit price is found for each basic time segment. The basic billing unit price of each basic time segment is multiplied by the length of the basic time unit and then summed to obtain the basic time-segment billing amount.

[0013] Optionally, the rule subgraph matching of the time-segment adjustment rule layer includes: The earliest entry time stamp and the latest exit time stamp of the parking session record are mapped to all basic time segments corresponding to the parking lot’s preset time area adjustment rule set. For each basic time segment, the time area to which it belongs is found and the corresponding adjustment coefficient is obtained. The basic billing unit price of the basic time segment is multiplied by the basic time unit length, and then multiplied by the adjustment coefficient and accumulated to obtain the billing amount after time area adjustment.

[0014] Optionally, the rule subgraph matching of the cross-database settlement rule layer includes: For cases where the same license plate number string flows continuously between multiple parking garages, the effective dwell time of all corresponding cross-garage parking session cleaning records is summed to obtain the total effective parking time after cross-garage merging. The merging price rate defined by the cross-garage merging rule is found, and the merging price rate is multiplied by the total effective parking time after cross-garage merging to obtain the cross-garage merging billing amount. If the parking lot has a capping rule, the minimum value between the combined cross-parking billing amount and the capping amount will be used as the final cross-parking settlement amount.

[0015] Optionally, the step of performing a global path search within the billing rule graph based on amount change records includes: Based on the billing rule graph, the amount change records and rule hit sequences are encoded as path nodes and directed edges respectively, and a cost path search graph structure is constructed. Perform global path enumeration on the cost path search graph structure to form a set of directed paths for all feasible cost paths, and obtain the total cost corresponding to each path. Determine whether each feasible cost path meets regulatory compliance constraints. If it does, the result of the regulatory compliance constraint is one; otherwise, it is zero. Determine whether the feasible cost path meets operational strategy constraints. If it does, the result of the operational strategy constraint is one; otherwise, it is zero. Only when a feasible cost path meets both regulatory compliance constraints and operational strategy constraints is the corresponding feasible cost path determined as a valid path. Among all valid paths that pass the constraint filtering, the path with the optimal execution cost is selected, and the path with the lowest final cost is determined. The billing amount corresponding to the path with the lowest cost is the final billing result.

[0016] The beneficial effects of this invention are: (1) This invention introduces an improved product unit residual network in the field of multi-rate billing for cross-garage parking in parking lots. By designing a cross-garage connectivity kernel and a hierarchical gating residual mechanism, it realizes the multiplicative interaction modeling of the core features of entry and exit time interval, inter-garage spatial connectivity distance and recognition confidence. The improved layer sensitivity coefficient and confidence temperature calibration mechanism of the product unit residual network output further enhance the robustness of the model in scenarios of abnormal occlusion and license plate recognition fluctuations. The session initial record obtained by PURe aggregation is directly decoupled and connected with the subsequent hierarchical rule engine, realizing the bidirectional linkage between model judgment and rule arrangement. It can significantly reduce the probability of mis-aggregation and mis-splitting in cross-garage parking behavior recognition and reduce the phenomenon of duplicate billing and omission.

[0017] (2) The present invention uses a three-layer structure of basic time-based billing, time-based regional adjustment and cross-database settlement to form a graph-based layer. Each layer is independently maintained and matched by pluggable rule subgraphs. All amount change records and rule hit sequences are encoded as directed path nodes and edges in the graph structure, supporting global path enumeration and multi-constraint filtering. Through automatic comparison of regulatory and operational strategies, only the billing result with the best cost and compliance is output. It not only supports rule hot update, rollback and canary release in complex scenarios, but also provides users and operators with full path audit and review basis, improving the system's maintenance efficiency and billing transparency. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a multi-level rate calculation method for cross-garage parking within a parking lot based on deep learning, as proposed in this invention. Figure 2 This is a structural block diagram of the improved product unit residual network model in a deep learning-based multi-level rate billing method for cross-garage parking within a parking lot, as proposed in this invention. Detailed Implementation

[0019] Example 1: Reference Figure 1 A deep learning-based multi-level pricing method for cross-garage parking within a parking lot, comprising: Collect multi-source raw data generated by vehicles at multiple garage entrances and exits, convert it into a structured spatiotemporal recognition dataset, preprocess the spatiotemporal recognition dataset, and generate a set of feature vectors. In this embodiment, the multi-source raw data includes license plate recognition image sequences, video trajectory data, and parking lot topology information. The feature vector set includes entry and exit time interval features, inter-parking space connectivity distance features, and recognition confidence features.

[0020] refer to Figure 2 The feature vector set is input into the improved product unit residual network model to obtain the session aggregation result representing the probability of the same cross-garage parking session and the corresponding confidence threshold, and the initial record of the cross-garage roaming parking session is generated based on the session aggregation result. In this embodiment, the improved residual network model of the product unit includes: The feature vector set is normalized to obtain the normalized input vector set; In Example 1, the entry / exit time interval feature and the inter-library spatial connectivity distance feature are linearly scaled and normalized to obtain normalized entry / exit time interval features and normalized inter-library spatial connectivity distance features. The normalized identification confidence feature is then set to equal the original identification confidence feature to form a normalized input vector set.

[0021] For each normalized input vector, a cross-garage connectivity kernel is constructed based on the topological and continuity characteristics of the parking lot cross-garage recognition scenario; The cross-repository connectivity kernel is used to quantify the impact of normalized entry-exit time interval features and normalized inter-repository spatial connectivity distance features on the spatiotemporal consistency of cross-repository systems. The closer the cross-repository connectivity kernel value is to 1, the higher the spatiotemporal consistency. The closer the cross-repository connectivity kernel value is to 0, the weaker the cross-repository correlation. The cross-repository connectivity kernel is calculated by weighted exponential decay of normalized entry-exit time interval features and normalized inter-repository spatial connectivity distance features.

[0022] In the improved product unit residual network model structure of the first layer, the main response of the k-th product unit neuron is calculated, and the arithmetic mean of the exponential parameters of the main responses of all product unit neurons in the first layer is taken to obtain the layer sensitivity coefficient of the first layer. In Example 1, in the improved product unit residual network model structure of the first layer, the normalized entry-exit time interval feature, the normalized inter-library spatial connectivity distance feature, and the normalized identification confidence feature are each added with a small constant and used as the base. For each base, the exponential parameter of the main response of the associated first layer product unit neuron is raised to obtain the main response of the input vector on all first layer product unit neurons.

[0023] ; ; in, This indicates that in the first product unit layer of the improved product unit residual network model, the... The neuron targets the first The main response of the product unit neuron in the recognition record. Indicates the first The first identification record after normalization Each feature component Corresponding to the normalized entry and exit time interval characteristics, Corresponding to the normalized inter-repository spatial connectivity distance feature, Corresponding to normalized identification confidence features This indicates that in the first product unit layer of the improved product unit residual network model, the... The first neuron pairs with the first The exponential parameter of each feature, Indicates that for the first The cross-depot connectivity kernel calculated from each identification record is used to measure the actual connectivity probability of vehicles flowing between multiple depots. Indicates that for the first Normalized identification confidence features of each identification record Number the neurons in the current product unit layer. Represents the regularization constant. This represents the first layer of the improved product unit residual network model. The first neuron The basic learnable parameters for each feature channel Represents a smooth activation function. This indicates that the first layer reflects the first level. The first neuron The sensitivity adjustment coefficient for each feature channel to low confidence in license plate recognition.

[0024] For all exponential parameters involved in the main response of all product unit neurons in the first layer, the arithmetic mean of the exponential parameter values ​​corresponding to all product unit neurons and all input features in the first layer is calculated. The average value is defined as the first layer sensitivity coefficient. The first layer sensitivity coefficient is used to quantitatively measure the overall sensitivity of the current identification record in the first layer to the multiplicative interaction contribution of the normalized entry and exit time interval feature, the normalized inter-library spatial connectivity distance feature, and the normalized identification confidence feature.

[0025] ; in, This represents the number of neurons in the residual block of the first-layer product unit. Indicates the first The input is the sensitivity coefficient of the residual layer of the first product unit.

[0026] The sensitivity coefficient of the first layer is input into the gated residual structure to calculate the dynamic gate factor of the first layer, and the output feature vector of the residual block of the first layer product unit is calculated based on the dynamic gate factor of the first layer. In Example 1, the sensitivity coefficient of the first layer is input into the gated residual structure. The gated residual structure adds a small constant to each component of the normalized input vector, takes the logarithm, multiplies it by the gate weight vector, adds the gate bias, and then adds it element-wise with the sensitivity coefficient of the first layer and the bias of the first layer gate coefficient before inputting it into the Sigmoid function to obtain the dynamic gate factor of the first layer. The dynamic gate factor of the first layer is used to perform weighted fusion of the main path output features and the identity mapping path features. The main path output features are the result of the output vector of the first layer product unit after linear transformation of the fully connected weight matrix and the bias vector and element-wise nonlinear function processing. The identity mapping path features are the output vector of the first layer product unit itself. The main path output features and the identity mapping path features are multiplied by the dynamic gate factor of the first layer and 1 minus the dynamic gate factor of the first layer, respectively, and then added element-wise to obtain the output feature vector of the residual block of the first layer product unit.

[0027] For the product unit neural network structure from the 2nd to the Lth layer, the same calculation method as the 1st layer is used to obtain the output feature vector and layer sensitivity coefficient of the product unit residual block of each layer. The arithmetic mean of the layer sensitivity coefficients of all layers is taken to obtain the full layer sensitivity coefficient. Calculate the pairwise normalized entry and exit time interval characteristics, the pairwise normalized inter-library spatial connectivity distance characteristics, the pairwise normalized identification confidence characteristics, and the pairwise full-layer sensitivity coefficient; In Example 1, each pair of normalized input vectors in the feature vector set is denoted as the first normalized input vector and the second normalized input vector, respectively.

[0028] The smaller value between the first normalized in-out time interval feature and the second normalized in-out time interval feature is taken as the pairwise normalized in-out time interval feature.

[0029] The smaller value between the first and second normalized inter-repository spatial connectivity distance features is taken as the pairwise normalized inter-repository spatial connectivity distance feature.

[0030] The smaller value between the first normalized identification confidence feature and the second normalized identification confidence feature is taken as the pairwise normalized identification confidence feature.

[0031] The arithmetic mean of the first and second full-layer sensitivity coefficients is taken as the paired full-layer sensitivity coefficient.

[0032] The session pairwise uncalibrated score is calculated by outputting the feature vector of the residual block of the product unit. The temperature calibration factor is generated based on the confidence feature of the pairwise normalization and the pairwise full-layer sensitivity coefficient. The session pairwise uncalibrated score is then calibrated based on the temperature calibration factor to obtain the session pairwise aggregation probability value. In Example 1, the output feature vectors of each pair of product units residual blocks are multiplied element-wise, and the result is multiplied with the output weight vector and the output bias is added to obtain the session pair uncalibrated score. Based on the pair normalization identification confidence feature and the pair full layer sensitivity coefficient, a temperature calibration factor is generated through a fixed term and weighted sum method. The session pair uncalibrated score is calibrated according to the temperature calibration factor and then processed by the Sigmoid function to obtain the session pair aggregation probability value. Based on the pairwise normalized entry and exit time interval features, the pairwise normalized inter-library spatial connectivity distance features, the pairwise normalized identification confidence features, and the pairwise full-layer sensitivity coefficient, an adaptive session aggregation threshold is generated by calculating the parameterized Sigmoid function. In Example 1, the paired normalized entry-exit time interval feature, paired normalized inter-library spatial connectivity distance feature, paired normalized identification confidence feature, and paired full-layer sensitivity coefficient of each pair of identification records are used as inputs. Each set of input features is multiplied by the corresponding threshold weight parameter and then summed. The threshold weight parameter includes the entry-exit time interval weight, inter-library spatial connectivity distance weight, identification confidence weight, and full-layer sensitivity coefficient weight. The resulting weighted sum is then added to the threshold bias parameter. The weighted sum result is input into the Sigmoid function to obtain an adaptive session aggregation threshold with a value range between 0 and 1.

[0033] The entry / exit time interval weight is used to adjust the influence of the pairwise normalized entry / exit time interval feature on the threshold; the inter-depot spatial connectivity distance weight is used to adjust the influence of the pairwise normalized inter-depot spatial connectivity distance feature on the threshold; the identification confidence weight is used to adjust the influence of the pairwise normalized identification confidence feature on the threshold; the full-layer sensitivity coefficient weight is used to adjust the influence of the pairwise full-layer sensitivity coefficient on the threshold; the threshold bias parameter is used for global offset adjustment; and the adaptive session aggregation threshold is used as the decision criterion for determining whether the two identification records belong to the same cross-depot parking session.

[0034] The probability value of each pair of sessions is compared with the corresponding adaptive session aggregation threshold. When the probability value of the pair of sessions is greater than or equal to the adaptive session aggregation threshold, the two identified records are identified as belonging to the same cross-garage parking session. All identified records are established with an undirected adjacency relationship according to this rule. The connected components of the adjacency relationship are extracted. Each connected component represents a set of initial records of cross-garage roaming parking sessions aggregated by sessions.

[0035] Based on the initial records of cross-garage roaming parking sessions, conflict resolution processing is performed to remove multi-garage conflict records within the same time window, and anomaly detection processing is performed to screen out suspected license plate replacement anomaly records and obstruction anomaly records, resulting in cleaned cross-garage roaming parking session records. In this embodiment, anomaly detection processing is implemented to filter out suspected license plate replacement anomaly records and obscuring anomaly records, including: Time alignment is performed on the initial record of each cross-garage roaming parking session, and the time span of the initial record of each cross-garage roaming parking session is defined as the interval from the earliest entry timestamp to the latest exit timestamp of all identified segments. The earliest entry timestamp is used to mark the time when a vehicle in the record was first identified, and the latest exit timestamp is used to mark the time when a vehicle in the record was last identified.

[0036] For any two different initial records of cross-garage roaming parking sessions, if the time span of the two records overlaps and the license plate number strings of the two records are the same, then it is determined that the two initial records of cross-garage roaming parking sessions have a cross-garage session time overlap conflict. For each pair of cross-database roaming parking session initial records that have cross-database session time overlap conflicts, calculate the spatiotemporal consistency score; The spatiotemporal consistency score is used to evaluate the physical spatial rationality between the initial records of the two cross-garage roaming parking sessions. The spatiotemporal consistency score is calculated as follows: calculate the shortest non-overlapping time difference between the initial records of the two cross-garage roaming parking sessions. The time difference is the earliest entry timestamp of the latter record minus the latest exit timestamp of the former record. If the result is negative, it is counted as zero. Calculate the shortest topological connectivity path distance between the garages where the two records are located. Sum the shortest non-overlapping time difference and the shortest topological connectivity path distance in a weighted manner. Calculate the exponent using the base of the natural logarithm to obtain the spatiotemporal consistency score.

[0037] For each pair of cross-database session time overlap conflict records, a consistency threshold judgment is performed to obtain the intermediate cross-database parking session set; The consistency threshold judgment rules include: When the spatiotemporal consistency score is less than the consistency threshold, it is judged as a pseudo-conflict with temporal overlap but spatial unreachability, and the initial record of the cross-database roaming parking session with a low spatiotemporal consistency score is deleted. When the spatiotemporal consistency score is greater than or equal to the consistency threshold, retain the two initial records of the cross-database roaming parking session and mark them as tolerable short-term overlap; All retained initial records of cross-garage roaming parking sessions form an intermediate set of cross-garage parking sessions after conflict resolution.

[0038] For each session record in the intermediate cross-garage parking session set after conflict resolution, occlusion anomaly detection and suspected license plate replacement detection are performed, and the corresponding recognition record sequence is extracted. Each recognition record is treated as an independent license plate recognition segment. In Example 1, for each session record in the intermediate cross-garage parking session set after conflict resolution, all license plate recognition records contained in the session record are obtained. Each license plate recognition record is an independent license plate recognition segment. The recognition record sequence is an ordered set formed by sorting all the license plate recognition segments aggregated in the session record by time, which is used for occlusion anomaly detection and suspected license plate replacement detection.

[0039] Extract the recognition confidence score of the bar recognition record image and calculate the recognition confidence variance of the session record; In Example 1, the difference between the image recognition confidence of each recognition record in the session record and the average image recognition confidence of all recognition records in the session record is squared. All squared differences are summed and then divided by the total number of recognition records in the session record to obtain the recognition confidence variance. The recognition confidence variance reflects the degree of fluctuation of recognition confidence within the session record.

[0040] When the identification confidence variance is greater than or equal to the occlusion anomaly identification variance threshold, and there are at least three consecutive frames of identification segments in the session record with identification confidence lower than the confidence threshold, the cross-garage parking session is determined to have an occlusion anomaly and is removed or split into multiple session records. Compare the license plate character sequences of any two recognition records, calculate the edit distance between them, and divide the edit distance by the larger value of the lengths of the two license plate strings to obtain the shortest edit distance ratio. When the shortest edit distance ratio is greater than or equal to the license plate replacement anomaly ratio threshold, the corresponding session record is determined to be a suspected license plate replacement anomaly record. Delete all session records that meet the conditions for occlusion anomaly or license plate change anomaly, and output the remaining session records as the final cross-garage roaming parking session cleaning record.

[0041] A billing rule graph is constructed, which divides the billing rules into a basic time-based billing rule layer, a time-area adjustment rule layer, and a cross-depot settlement rule layer. The cross-depot roaming parking session cleaning record is used as input. The billing rule layers of the billing rule graph are triggered in sequence to perform rule subgraph matching calculations, thereby obtaining the basic time-based billing amount, the time-area adjustment billing amount, and the cross-depot merged billing amount. A record of amount change is generated in each billing rule layer. In this embodiment, the construction of a billing rule graph includes: Construct a billing rule graph, taking cross-database roaming parking session cleaning records as input, and triggering the basic time-based billing rule layer, time-area adjustment rule layer, and cross-database settlement rule layer in hierarchical order. Execute the corresponding rule subgraph matching under each level, and output the amount change record and rule hit sequence. In Example 1, the billing rule graph consists of multiple rule subgraph nodes and their connections. Each rule subgraph node corresponds to a set of rule entries of a certain type. The process of constructing the billing rule graph includes dividing all rule entries into hierarchical categories according to their responsibilities, encapsulating rule entries belonging to the same level into independently executable rule subgraphs, and establishing chain connections between rule subgraphs of different levels according to a fixed execution order, so that during execution, cross-database roaming parking session cleaning records flow from the first-level rule subgraph into the rule subgraphs of subsequent levels in sequence.

[0042] When a cross-garage roaming parking session cleaning record is input into the billing rule graph, the rule subgraph entry matching the cleaning record is searched in the basic time-segmented billing rule layer. The matching conditions are used to determine whether the corresponding cleaning record meets the triggering conditions of a certain rule subgraph. If the match is successful, the billing calculation logic corresponding to the rule subgraph is executed, and the amount change generated during the execution of the rule subgraph is written into the amount change record. The triggered rule subgraph node is written into the rule hit sequence.

[0043] After completing the subgraph matching of the basic time-of-use billing rule layer, the basic time-of-use billing result is passed as input to the time-of-use area adjustment rule layer. In the time-of-use area adjustment rule layer, the above matching and amount recording process is repeated for each rule subgraph. Based on the successfully matched rule subgraph, the adjustment logic is executed and new amount change records and rule hit sequences are generated.

[0044] After the time-area adjustment rule layer is completed, the adjustment result is passed as input to the cross-database settlement rule layer. The same rule subgraph matching process is repeated in the cross-database settlement rule layer. By identifying rule subgraphs such as cross-database session markers, cross-database merging relationships, and capped entries, the cross-database settlement logic is completed and the amount change records and rule hit sequences in the final level are generated.

[0045] Perform rule subgraph matching of the basic time-based billing rule layer on the cross-garage roaming parking session cleaning records to calculate the basic time-based billing amount; The rule subgraph matching of the basic time-of-use billing rule layer includes: The difference between the earliest entry time stamp and the latest exit time stamp is taken as the effective dwell time of the parking session. The effective dwell time is divided into basic time segments according to the basic time unit. The corresponding basic billing unit price is found for each basic time segment. The basic billing unit price of each basic time segment is multiplied by the length of the basic time unit and then summed to obtain the basic time-segment billing amount. The basic time-based billing amount serves as the basic billing result, and the rule hit sequence and amount change record corresponding to the basic billing result are recorded.

[0046] Based on the basic time-based billing amount, perform rule subgraph matching of the time-based area adjustment rule layer to obtain the billing amount after time-based area adjustment; The rule subgraph matching of the time-region adjustment rule layer includes: The earliest entry time stamp and the latest exit time stamp of the parking session record are mapped to all basic time segments corresponding to the parking lot’s preset time area adjustment rule set. For each basic time segment, the time area to which it belongs is found and the corresponding adjustment coefficient is obtained. The basic billing unit price of the basic time segment is multiplied by the basic time unit length, and then multiplied by the adjustment coefficient and accumulated to obtain the billing amount after time area adjustment.

[0047] Based on the billing amount after time period and regional adjustment, the rule subgraph matching of the cross-database settlement rule layer is performed to generate the cross-database merged billing amount; The rule subgraph matching of the cross-database settlement rule layer includes: For cases where the same license plate number string flows continuously between multiple parking garages, the effective dwell time of all corresponding cross-garage parking session cleaning records is summed to obtain the total effective parking time after cross-garage merging. The merging price rate defined by the cross-garage merging rule is found, and the merging price rate is multiplied by the total effective parking time after cross-garage merging to obtain the cross-garage merging billing amount. If the parking lot has a capping rule, the minimum value between the combined cross-parking billing amount and the capping amount will be used as the final cross-parking settlement amount.

[0048] The basic time-based billing amount, the billing amount after time-area adjustment, and the cross-database merged billing amount are summarized in hierarchical order to form a complete billing hierarchy. At the same time, the rule hit sequence and amount change records of all levels are combined in sequence to form a complete execution trajectory of the billing rule graph.

[0049] Based on the amount change record, a global path search is performed in the billing rule graph to determine the optimal cost path that meets regulatory requirements and operational strategies, and the final billing result is obtained. In this embodiment, a global path search is performed within the billing rule graph based on the amount change record, including: Based on the billing rule graph, the amount change records and rule hit sequences are encoded as path nodes and directed edges respectively, and a cost path search graph structure is constructed. In Example 1, the cost path search graph structure is as follows: For each cross-depot roaming parking session cleaning record, amount change records and rule hit sequences are generated at the basic time-segmented billing rule layer, the time-segmented area adjustment rule layer, and the cross-depot settlement rule layer, respectively. These are numbered in hierarchical order to form a set of path nodes. The set of path nodes corresponds to the basic time-segmented billing node, the time-segmented area adjustment billing node, and the cross-depot settlement billing node, respectively. The billing amount corresponding to each level is in yuan. A set of directed edges is constructed between the path nodes according to the billing level order. Each directed edge carries the amount change value from the billing result of the previous level to the billing result of the next level. The amount change value is equal to the billing amount of the next level minus the billing amount of the previous level.

[0050] Perform global path enumeration on the cost path search graph structure to form a set of directed paths for all feasible cost paths, and obtain the total cost corresponding to each path. In Example 1, a depth-first search is used to generate all feasible cost paths from the basic time-segment billing node to the cross-database settlement node. Each feasible cost path consists of several path nodes. For each feasible cost path, the total cost is obtained by summing the billing amounts on the path nodes in the order of the path.

[0051] Determine whether each feasible cost path meets regulatory compliance constraints. If it does, the result of the regulatory compliance constraint is one; otherwise, it is zero. Determine whether the feasible cost path meets operational strategy constraints. If it does, the result of the operational strategy constraint is one; otherwise, it is zero. Only when a feasible cost path meets both regulatory compliance constraints and operational strategy constraints is the corresponding feasible cost path determined as a valid path. The regulatory compliance constraint is to compare the rule sequence of each feasible cost path with the current parking policies and the billing standards issued by the local price authorities. If the rules applicable to each billing node of the feasible cost path do not violate the current industry standards and the regulations of the competent authorities, the feasible cost path is deemed to meet the regulatory compliance constraint; otherwise, it does not.

[0052] The operational strategy constraint is to compare the rule hit sequence of each feasible cost path with the operational documents set by the parking lot operator, including the operating strategy, user rights policy, promotional discount rules, and internal financial control regulations. If every rule hit node in the feasible cost path conforms to the currently effective operating strategy and user rights rules, the feasible cost path is determined to meet the operational strategy constraint; otherwise, it is not.

[0053] Among all valid paths that pass the constraint filtering, the path with the optimal execution cost is selected, and the path with the lowest final cost is determined. The billing amount corresponding to the path with the lowest cost is the final billing result.

[0054] In this embodiment, the total cost of each valid path after constraint filtering is calculated, and the path with the lowest total cost is selected as the final path with the lowest cost. The index of the final path with the lowest cost is equal to the number of the valid path with the lowest cost among all valid paths. The final path with the lowest cost is the path with the lowest cost, and the final billing result is the billing amount corresponding to the path with the lowest cost.

[0055] The node sequence of the path with the lowest final cost is taken as the final rule hit sequence, and the amount change value between adjacent path nodes in the path with the lowest final cost is taken as the final amount change record. The final rule hit sequence and the final amount change record are taken together as the final billing link output of the cross-garage parking session.

[0056] At the edge computing node, the improved product unit residual network model inference is executed to generate temporary billing results and partial billing rule preprocessing results. Based on the partial billing rule preprocessing results, the complete billing rule graph calculation is performed. The temporary billing results are checked and supplemented, and cross-database reconciliation and settlement are completed to form the fare settlement result.

[0057] Example 2: During the deployment of an intelligent parking system in a large commercial complex, the operator collected the following actual parking data snippets: The parking lot has three garages, A, B, and C, where vehicles can move freely between them. Each garage is equipped with license plate recognition, geomagnetic sensing, and Bluetooth access detection devices. If a vehicle (license plate number example: XN·12345) appears in all three garages on the same day, the system will capture the following data: 14:03:45 Entered Library A, geomagnetic activation, image recognition confidence level 0.97; 14:09:31 Left Library A, image recognition confidence level 0.95; 14:12:17 Entered Library B, image recognition confidence level 0.96; 16:55:04 Left Library B, image recognition confidence level 0.91; 17:02:20 Entered Library C, Bluetooth recognition + geomagnetic confirmation, image recognition confidence level 0.90; 17:08:05 Left Library C, image recognition confidence level 0.94.

[0058] All original records and device-collected confidence levels are truthfully uploaded to the central data management platform, and the relevant sensor data has been timestamped and aligned.

[0059] The system organizes the entry and exit data into feature vectors, the components of which include: entry and exit time interval (AB: 6 minutes, BC: 7 minutes), spatial connectivity distance (AB: 150 meters, BC: 120 meters), and recognition confidence (greater than 0.90 for multiple frames, minimum 0.91 for a single frame). Each feature vector is normalized to form a standard input.

[0060] The system calls the product unit residual network on the normalized feature input to calculate the cross-library connectivity kernel: For the flow from database A to database B, the normalized flow time interval is 0.11, the normalized spatial connectivity distance is 0.14, and the normalized confidence level is 0.93. The cross-database connectivity kernel index decay is 0.89. For the flow from database B to database C, the normalized flow time interval is 0.13, the normalized spatial connectivity distance is 0.12, and the normalized confidence level is 0.92. The cross-database connectivity kernel index decay is 0.88.

[0061] The main responses of the residual network of the first-layer product unit are as follows: ; The layer sensitivity coefficient is the average of the two master response parameters, which is 0.98. After dynamic gating and residual superposition, the aggregation probabilities of each pair of sessions output by the PURe network are 0.95 and 0.94, respectively, both of which are greater than the adaptive threshold (0.85).

[0062] The system automatically performs time alignment on all aggregated initial records. The time spans of records in databases A, B, and C are as follows: Database A: 14:03:45~14:09:31, Database B: 14:12:17~16:55:04, Database C: 17:02:20~17:08:05.

[0063] Time conflict detection: There is no overlap of cross-database sessions with the same license plate, all records meet the physical reachability conditions, and the spatiotemporal consistency scores are all above 0.90, so there is no need to remove conflicts.

[0064] Occlusion anomaly detection: The confidence variance of all identified segments is less than 0.01 (mean 0.925), and the number of consecutive frames below 0.85 is 0, which does not trigger an anomaly.

[0065] License plate swapping detection: All records have an edit distance ratio of 0, indicating no license plate swapping behavior. The system outputs the only cleaned cross-garage roaming parking session record.

[0066] Upon entering the tiered billing process, the parking lot's time-based billing rules are as follows: The fee is 4 yuan per 30 minutes, with any time less than 30 minutes counted as 30 minutes; the peak period is from 17:00 to 19:00, with an adjustment coefficient of 1.2; the daily cap for cross-warehouse merging is 30 yuan.

[0067] The basic time-segmented billing layer covers a total of 6 time segments for the effective duration of databases A, B, and C (6 minutes + 162 minutes + 6 minutes = 174 minutes, rounded up to 630 = 180 minutes). The basic billing amount is 64 = 24 yuan.

[0068] The time-slot adjustment rule layer identifies that the departure of warehouse B and the entry and exit of warehouse C both fall during peak hours, and the corresponding adjustment segment amount = (2 segments * 4 yuan) * 1.2 = 9.6 yuan.

[0069] The cross-database settlement rule layer merged the billing amount (24 + 9.6 = 33.6 yuan). Because it exceeded the daily cap, the final billing amount was 30 yuan. The system generated the rule hit sequence as {basic time-sharing billing, peak adjustment, cap}, and the amount change record was {24, +9.6, -3.6}.

[0070] In the cost path search and constraint verification, the system generates all valid billing paths hierarchically based on the path nodes {Library A, Library B, Library C}, enumerates all possible paths (in Example 2, Library A is billed separately + Library B + Library C are billed separately, or AB is merged + C is billed separately), and finally only the cross-library merged + peak adjustment + capping path simultaneously meets regulatory and operational constraints, and the total billing amount is at least 30 yuan. The system outputs the final minimum cost path with an amount of 30 yuan, and the hit sequence and amount changes are fully recorded in the audit link.

[0071] To demonstrate the beneficial effects of this invention, the experimental statistics of comparing traditional methods and this invention on the same batch of real cross-database data (sample size 8000 groups) are shown in Table 1 below: Table 1. Traditional methods and the present invention in the same batch of real cross-database data

[0072] The training samples (partial) are as follows: Feature vector 1: [Time interval 0.14, Spatial distance 0.13, Confidence 0.93], PURe main response 0.111, Aggregation probability 0.93; Feature vector 2: [Time interval 0.09, Spatial distance 0.11, Confidence 0.95], PURe main response 0.108, Aggregation probability 0.96; Feature vector 3: [Time interval 0.17, Spatial distance 0.12, Confidence 0.91], PURe main response 0.102, Aggregation probability 0.91.

[0073] Example 2 enables automatic and accurate session aggregation and billing tiering for a large amount of real parking lot entry and exit traffic, improving cross-database billing accuracy, system maintenance efficiency, and user experience.

[0074] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-level fee calculation method for cross-garage parking within a parking lot based on deep learning, characterized in that, include: Collect multi-source raw data generated by vehicles at multiple garage entrances and exits, convert it into a structured spatiotemporal recognition dataset, preprocess the spatiotemporal recognition dataset, and generate a set of feature vectors. The feature vector set is input into the improved product unit residual network model to obtain the session aggregation result and the corresponding confidence threshold, and the initial record of cross-database roaming parking session is generated based on the session aggregation result. Based on the initial records of cross-garage roaming parking sessions, conflict resolution processing is performed to remove multi-garage conflict records within the same time window, and anomaly detection processing is performed to screen out suspected license plate replacement anomaly records and obstruction anomaly records, resulting in cleaned cross-garage roaming parking session records. A billing rule graph is constructed, which divides the billing rules into a basic time-based billing rule layer, a time-area adjustment rule layer, and a cross-depot settlement rule layer. The cross-depot roaming parking session cleaning record is used as input. The billing rule layers of the billing rule graph are triggered in sequence to perform rule subgraph matching calculations, thereby obtaining the basic time-based billing amount, the time-area adjustment billing amount, and the cross-depot merged billing amount. A record of amount change is generated in each billing rule layer. Based on the amount change record, a global path search is performed in the billing rule graph to determine the optimal cost path that meets regulatory requirements and operational strategies, and the final billing result is obtained. At the edge computing node, the improved product unit residual network model inference is executed to generate temporary billing results and some preprocessing results of billing rules. The complete calculation of the billing rule graph is then performed, the temporary billing results are checked and supplemented, and cross-database reconciliation and settlement are completed to form the fare settlement result.

2. The method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning as described in claim 1, characterized in that, The multi-source raw data includes license plate recognition image sequences, video trajectory data, and parking lot topology information. The feature vector set includes entry and exit time interval features, inter-parking space connectivity distance features, and recognition confidence features.

3. The method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning as described in claim 1, characterized in that, The improved product unit residual network model includes: The feature vector set is normalized to obtain the normalized input vector set; For each normalized input vector, a cross-garage connectivity kernel is constructed based on the topological and continuity characteristics of the parking lot cross-garage recognition scenario; In the improved product unit residual network model structure of the first layer, the main response of the k-th product unit neuron is calculated, and the arithmetic mean of the exponential parameters of the main responses of all product unit neurons in the first layer is taken to obtain the layer sensitivity coefficient of the first layer. The sensitivity coefficient of the first layer is input into the gated residual structure to calculate the dynamic gate factor of the first layer, and the output feature vector of the residual block of the first layer product unit is calculated based on the dynamic gate factor of the first layer. For the product unit neural network structure from the 2nd to the Lth layer, the same calculation method as the 1st layer is used to obtain the output feature vector and layer sensitivity coefficient of the product unit residual block of each layer. The arithmetic mean of the layer sensitivity coefficients of all layers is taken to obtain the full layer sensitivity coefficient. Calculate the pairwise normalized entry and exit time interval characteristics, the pairwise normalized inter-library spatial connectivity distance characteristics, the pairwise normalized identification confidence characteristics, and the pairwise full-layer sensitivity coefficient; The session pairwise uncalibrated score is calculated by outputting the feature vector of the residual block of the product unit. The temperature calibration factor is generated based on the confidence feature of the pairwise normalization and the pairwise full-layer sensitivity coefficient. The session pairwise uncalibrated score is then calibrated based on the temperature calibration factor to obtain the session pairwise aggregation probability value. Based on the pairwise normalized entry and exit time interval features, the pairwise normalized inter-library spatial connectivity distance features, the pairwise normalized identification confidence features, and the pairwise full-layer sensitivity coefficient, an adaptive session aggregation threshold is generated by calculating the parameterized Sigmoid function. The probability value of each pair of sessions is compared with the corresponding adaptive session aggregation threshold. When the probability value of the pair of sessions is greater than or equal to the adaptive session aggregation threshold, the two identified records are identified as belonging to the same cross-garage parking session. All identified records are established with an undirected adjacency relationship according to this rule. The connected components of the adjacency relationship are extracted. Each connected component represents a set of initial records of cross-garage roaming parking sessions aggregated by sessions.

4. The method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning according to claim 1, characterized in that, The implementation of anomaly detection processing to screen out suspected license plate replacement and obscuring anomaly records includes: Time alignment is performed on the initial record of each cross-garage roaming parking session, and the time span of the initial record of each cross-garage roaming parking session is defined as the interval from the earliest entry timestamp to the latest exit timestamp of all identified segments. For any two different initial records of cross-garage roaming parking sessions, if the time span of the two records overlaps and the license plate number strings of the two records are the same, then it is determined that the two initial records of cross-garage roaming parking sessions have a cross-garage session time overlap conflict. For each pair of cross-database roaming parking session initial records that have cross-database session time overlap conflicts, calculate the spatiotemporal consistency score; For each pair of cross-database session time overlap conflict records, a consistency threshold judgment is performed to obtain the intermediate cross-database parking session set; For each session record in the intermediate cross-garage parking session set after conflict resolution, occlusion anomaly detection and suspected license plate replacement detection are performed, and the corresponding recognition record sequence is extracted. Each recognition record is treated as an independent license plate recognition segment. Extract the recognition confidence score of the bar recognition record image and calculate the recognition confidence variance of the session record; When the identification confidence variance is greater than or equal to the occlusion anomaly identification variance threshold, and there are at least three consecutive frames of identification segments in the session record with identification confidence lower than the confidence threshold, the cross-garage parking session is determined to have an occlusion anomaly and is removed or split into multiple session records. Compare the license plate character sequences of any two recognition records, calculate the edit distance between them, and divide the edit distance by the larger value of the lengths of the two license plate strings to obtain the shortest edit distance ratio. When the shortest edit distance ratio is greater than or equal to the license plate replacement anomaly ratio threshold, the corresponding session record is determined to be a suspected license plate replacement anomaly record. Delete all session records that meet the conditions for occlusion anomaly or license plate change anomaly, and output the remaining session records as the final cross-garage roaming parking session cleaning record.

5. A method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning, as described in claim 4, is characterized in that... The consistency threshold judgment rule includes: When the spatiotemporal consistency score is less than the consistency threshold, it is judged as a pseudo-conflict with temporal overlap but spatial unreachability, and the initial record of the cross-database roaming parking session with a low spatiotemporal consistency score is deleted. When the spatiotemporal consistency score is greater than or equal to the consistency threshold, retain the two initial records of the cross-database roaming parking session and mark them as tolerable short-term overlap; All retained initial records of cross-garage roaming parking sessions form an intermediate set of cross-garage parking sessions after conflict resolution.

6. The method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning according to claim 1, characterized in that, The construction of the billing rule graph includes: Construct a billing rule graph, taking cross-database roaming parking session cleaning records as input, and triggering the basic time-based billing rule layer, time-area adjustment rule layer, and cross-database settlement rule layer in hierarchical order. Execute the corresponding rule subgraph matching under each level, and output the amount change record and rule hit sequence. Perform rule subgraph matching of the basic time-based billing rule layer on the cross-garage roaming parking session cleaning records to calculate the basic time-based billing amount; Based on the basic time-based billing amount, perform rule subgraph matching of the time-based area adjustment rule layer to obtain the billing amount after time-based area adjustment; Based on the billing amount after time period and regional adjustment, the rule subgraph matching of the cross-database settlement rule layer is performed to generate the cross-database merged billing amount; The basic time-based billing amount, the billing amount after time-area adjustment, and the cross-database merged billing amount are summarized in hierarchical order to form a complete billing hierarchy. At the same time, the rule hit sequence and amount change records of all levels are combined in sequence to form a complete execution trajectory of the billing rule graph.

7. A method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning, as described in claim 6, is characterized in that... The rule subgraph matching of the basic time-based billing rule layer includes: The difference between the earliest entry time stamp and the latest exit time stamp is taken as the effective dwell time of the parking session. The effective dwell time is divided into basic time segments according to the basic time unit. The corresponding basic billing unit price is found for each basic time segment. The basic billing unit price of each basic time segment is multiplied by the length of the basic time unit and then summed to obtain the basic time-segment billing amount.

8. A method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning, as described in claim 6, is characterized in that... The rule subgraph matching of the time-region adjustment rule layer includes: The earliest entry time stamp and the latest exit time stamp of the parking session record are mapped to all basic time segments corresponding to the parking lot’s preset time area adjustment rule set. For each basic time segment, the time area to which it belongs is found and the corresponding adjustment coefficient is obtained. The basic billing unit price of the basic time segment is multiplied by the basic time unit length, and then multiplied by the adjustment coefficient and accumulated to obtain the billing amount after time area adjustment.

9. A method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning, as described in claim 6, is characterized in that... The rule subgraph matching of the cross-database settlement rule layer includes: For cases where the same license plate number string flows continuously between multiple parking garages, the effective dwell time of all corresponding cross-garage parking session cleaning records is summed to obtain the total effective parking time after cross-garage merging. The merging price rate defined by the cross-garage merging rule is found, and the merging price rate is multiplied by the total effective parking time after cross-garage merging to obtain the cross-garage merging billing amount. If the parking lot has a capping rule, the minimum value between the combined cross-parking billing amount and the capping amount will be used as the final cross-parking settlement amount.

10. A method for multi-level rate calculation of cross-garage parking within a parking lot based on deep learning, as described in claim 6, characterized in that, The step of performing a global path search within the billing rule graph based on amount change records includes: Based on the billing rule graph, the amount change records and rule hit sequences are encoded as path nodes and directed edges respectively, and a cost path search graph structure is constructed. Perform global path enumeration on the cost path search graph structure to form a set of directed paths for all feasible cost paths, and obtain the total cost corresponding to each path. Determine whether each feasible cost path meets regulatory compliance constraints. If it does, the result of the regulatory compliance constraint is one; otherwise, it is zero. Determine whether the feasible cost path meets operational strategy constraints. If it does, the result of the operational strategy constraint is one; otherwise, it is zero. Only when a feasible cost path meets both regulatory compliance constraints and operational strategy constraints is the corresponding feasible cost path determined as a valid path. Among all valid paths that pass the constraint filtering, the path with the optimal execution cost is selected, and the path with the lowest final cost is determined. The billing amount corresponding to the path with the lowest cost is the final billing result.

Images