Emergency rescue vehicle auxiliary positioning method fusing multi-source data

Abstract

The application discloses an emergency rescue vehicle auxiliary positioning method fusing multi-source data, and the method is: collecting ETC issuing data, transaction data and service area data and performing multi-source fusion; extracting a mobile phone number and an alarm time from an alarm event, binding a vehicle license plate number and generating a unique identifier; extracting a vehicle time sequence passing track based on an ETC gantry passing database, taking the nearest gantry node as an upstream reference point; fusing service area data and road network topology information to calculate an actual passing time; generating a path probability distribution and a prior context vector by using a global topology attention mechanism through a path prediction module, and predicting a speed reference line through a task condition recursion unit based on a speed estimation module; finally, combining the upstream gantry post number, the actual passing time and the predicted speed to calculate the current position of the vehicle and fusing mobile phone GPS data to perform abnormal correction positioning to obtain route number, downstream gantry and post number information. The application realizes lane-level accurate positioning and effectively improves rescue efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of expressways, and particularly relates to an emergency rescue vehicle auxiliary positioning method fusing multi-source data. BACKGROUND

[0002] In the process of expressway digitization upgrading, the rescue call platform is an important entrance of expressway emergency management and the first channel for drivers and passengers to seek help in emergency situations. The operation efficiency and information processing capacity of the platform directly affect the speed and quality of emergency response. In vehicle accident alarm processing, rescue efficiency and positioning accuracy are directly related to the safety of drivers and passengers, traffic smoothness and emergency management capacity. If the specific position and driving direction of the vehicle cannot be obtained in time and accurately, the rescue personnel will face the problems of insufficient information and chaotic dispatch, which may lead to rescue delay and even cause secondary accidents, resulting in more serious personnel casualties and property losses. At present, the emergency management department of the expressway urgently needs a new method capable of accurately predicting the real-time position and driving direction of the vehicle to realize rapid rescue and scientific control of the accident site. At the same time, by accurately predicting the position and driving direction of the vehicle, basic data support can also be provided for the fine management of the expressway, promoting the digitization and intelligent development of the expressway network, and ultimately improving the overall operation efficiency and safety guarantee level.

[0003] Under the existing technology and management system, the expressway accident emergency management is faced with the following problems: 1. The positioning accuracy is insufficient and the misjudgment rate is high. At present, the vehicle positioning of the rescue call platform mainly relies on the GPS information of mobile phone signaling data and manual inquiry. However, using mobile phone signaling data for GPS positioning can only be positioned in a rough range, and cannot accurately position the vehicle position in the expressway scene, nor can it distinguish the specific lane direction and lane position of the vehicle. When the alarm person provides the stake number information through oral description, due to the complex network structure of the expressway, the discontinuous stake number identification or the unclear description of the driver and passenger, the positioning information is often inaccurate. This positioning ambiguity directly leads to the fact that the rescue vehicle cannot determine the driving route, so that the rescue resource scheduling has great uncertainty, delays the rescue speed and reduces the rescue efficiency. 2. The positioning is delayed and cannot respond in real time. The traditional manual patrol or video monitoring means has a long information collection and transmission chain and slow reaction, which cannot meet the real-time requirement. Especially in the initial stage of the accident, the positioning lag will cause the rescue vehicle to fail to arrive at the scene in the first time, miss the golden rescue time and bring serious threat to the safety of life and traffic order. 3. The data is fragmented and lacks unified integration. At present, the various data sources involved in the expressway management, including ETC transaction records, service area parking information, license plate recognition data and telephone alarm information, are scattered in different systems. These data formats are different, the codes are not unified, there is lack of effective interface between systems, so that efficient data fusion cannot be carried out. The fragmented data makes the prediction algorithm lack reliable input basis, thereby reducing the accuracy and reliability of the prediction result. 4. The emergency decision support is insufficient. Due to the lag of information transmission and the inaccuracy of positioning, the emergency management department lacks real-time and reliable data support when formulating the rescue plan. The dispatching decision often depends on manual experience, and the rescue action may be blind and delayed, and even cause further deterioration of the on-site traffic, cause serious traffic congestion and increase the risk of secondary accidents. The lack of scientific decision support has become an important bottleneck restricting the improvement of the expressway accident emergency management capability. 5. The coupling gap of the existing algorithm model: the emergency calculation relies on the kinematic formula (position = speed × time), and the core challenge lies in the strong coupling of “driving direction” and “speed reference line”. The speed reference line of the vehicle is highly dependent on the path it selects (such as the main road vs. the ramp). The existing technology, whether it is a basic statistical model (such as multiple linear regression) or a single and homogeneous deep learning model, lacks an explicit mechanism to decouple this dependence.

[0004] These “path blind” algorithms cannot effectively distinguish the macro intention of the vehicle when facing a fork, so that the speed prediction is fundamentally inaccurate. SUMMARY

[0005] The application aims to solve the problems of insufficient vehicle positioning accuracy, lagging positioning information, non-uniform data, insufficient emergency decision support and the like in highway emergencies, and provides an emergency rescue vehicle auxiliary positioning method fusing multi-source data, which realizes accurate positioning and driving direction prediction of highway vehicles by fusing ETC transaction data, service area passing records, deep learning technology and road network topology information.

[0006] The technical scheme adopted by the application is:

[0007] The multi-source data fusion emergency rescue vehicle auxiliary positioning method comprises the following steps:

[0008] Collecting multi-source data and performing fusion processing to obtain a multi-source data set;

[0009] Receiving an alarm event from a rescue call platform, extracting an alarm mobile phone number and an alarm time;

[0010] Querying ETC card issuance data based on the mobile phone number, binding a vehicle license plate number and generating a unique identifier;

[0011] Extracting a time sequence passing trajectory of the corresponding vehicle from an ETC gantry passing database, and taking the last gantry node closest to the alarm time as an upstream reference gantry;

[0012] Fusing service area data and road network topology information to calculate the actual passing time of the vehicle in the accident section;

[0013] Cascade prediction is performed through a space-time network multi-task relationship model (ST-MTR): first, a path prediction module (PPM) generates a path probability distribution and a path prior context vector, and then a speed estimation module (VPM) inputs and predicts a vehicle speed reference line based on the context vector;

[0014] Based on the gantry pile number of the upstream reference point, the actual passing time and the predicted speed, the estimated pile number position of the vehicle is calculated, and the abnormal correction is performed in combination with the mobile phone GPS data;

[0015] Specifically, the estimated pile number position of the current vehicle is calculated according to the corresponding position of the upstream reference gantry , the actual passing time of the vehicle in the accident section and the estimated speed reference line value of the future driving of the vehicle . .

[0016] The vehicle route number, downstream gantry and pile number information are output to the rescue platform to provide position information support for traffic management and emergency rescue.

[0017] Further, the step of extracting the time sequence passing trajectory of the corresponding vehicle is as follows:​

[0018] Extract all transaction records of the corresponding vehicle from the ETC gantry passage database, and each transaction record is in the format of:

[0019] ;

[0020] Wherein, Passid is the unique track number, Node is the identification of the toll station or gantry, and Time is the corresponding transaction time;

[0021] Sort the transaction record data in chronological order, and group them according to the track number Passid to generate the time sequence passage track of the vehicle. The expression of the time sequence passage track is:

[0022] ;

[0023] Wherein, is the first node of the time sequence passage track; is the transaction time of the first node of the time sequence passage track; and in the same way, is the node of the time sequence passage track; is the node of the time sequence passage track;

[0024] Based on the alarm time (TimeAlarm), the track closest to the alarm time is extracted from the time sequence passage track data, and the last node node of the corresponding track is regarded as the latest node before the accident.

[0025] The lane information of the latest node before the accident is extracted from the gantry information table as the upstream reference gantry of the accident, and the expression is:

[0026] ;

[0027] Wherein, is the route name; is the lane; is the stake number.

[0028] Further, the actual passage time of the vehicle in the accident section is calculated as follows:

[0029] Calculate the time difference between the two time points when the vehicle enters and leaves the accident section , and the expression is:

[0030] ;

[0031] Wherein, is the exit time of the vehicle in the accident section; This refers to the time the vehicle entered the accident area.

[0032] Based on the vehicle's actual driving trajectory, combined with information from toll stations and service area entrances / exits, the actual driving time of the vehicle in the accident section is calculated; by integrating service area data, the actual passage time of the vehicle in the accident section is calculated.

[0033] upstream reference gantry Starting from a node, the system retrieves all adjacent road segments originating from that node based on the road network topology, and integrates toll station information to construct a target toll station set. Its expression is as follows:

[0034]

[0035] in, Indicates the relationship with the reference node The total number of downstream nodes or toll stations with direct topological connections; the purpose of constructing this set is to delineate the possible departure range of vehicles, thereby retrieving whether there is a corresponding departure record for the vehicle.

[0036] Traverse the segment to find its road topology and related service area entrance / exit information, then traverse the service area entrance / exit data to check if the corresponding vehicle license plate number exists. The corresponding passage record; if it exists, the dwell time of the corresponding vehicle in the service area is calculated. Calculate the actual travel time of the vehicle within the accident area. :

[0037] ;

[0038] in, A function to calculate the time difference between two points in time. For alarm time, This refers to the passage time of a vehicle through the upstream reference gantry.

[0039] Furthermore, the path prediction module (PPM) employs a global topology attention mechanism to perform the following steps:

[0040] Convert vehicle gantry passage records into a spatiotemporal state sequence;

[0041] The path probability distribution is generated by calculating the road network node dependency weight matrix through self-attention;

[0042] Specifically, depending on the weight matrix The formula is as follows:

[0043] ;

[0044] in, is the input trajectory feature tensor, is the scaling factor of the feature space.

[0045] ;

[0046] where, is the weight matrix of the path prediction layer, is the bias parameter of the path prediction layer, is the global intent distribution, is the normalized exponential function.

[0047] The predicted probability distribution vector generates the path prior context vector .

[0048] where, is the weight matrix of the feature compression operator, is the bias parameter of the feature compression operator, is the activation function.

[0049] The vector condenses the topological intent of the macroscopic path, which will be passed as a "conditional constraint" to the subsequent VPM module.

[0050] Further, the velocity estimation module (VPM) employs a task-conditioned recurrent unit (TC-RU) to perform the following steps:

[0051] concatenating the historical velocity features with the path prior context vector into a hybrid vector;

[0052] Specifically, at each time step of the VPM, the input space is reconstructed into an "observation-prior" hybrid vector ; ; where, is the historical velocity feature at the current time.

[0053] The memory manifold is updated through the dynamic gating mechanism of the information preservation gate , the information injection gate , and the state output gate , and outputs the velocity prediction value under the path prior constraint.

[0054] Specifically, the dynamic evolution equation is as follows:

[0055] (1) Gating activation calculation:

[0056] ;

[0057] ;

[0058] where, and denote the gating value of information retention gate and information injection gate, respectively; denotes the hidden state at the last time step; denotes the mixed vector after concatenating the historical speed feature and the path prior context at the current time step; denotes the Sigmoid activation function; and denote the weight matrix and bias parameter of information retention gate, respectively; and denote the weight matrix and bias parameter of information injection gate, respectively.

[0059] (2) Memory manifold update: compute the candidate memory state at the current time step and update the global memory manifold :

[0060] ;

[0061] ;

[0062] where, denotes the candidate memory state computed at the current time step; and denote the global memory manifold at the last time step and the updated global memory manifold at the current time step, respectively; and denote the weight matrix and bias parameter of the candidate memory state computation layer, respectively; denotes the hyperbolic tangent activation function; denotes the Hadamard product.

[0063] (3) Hidden state and speed mapping: update the hidden state at the current time step and map to the final speed reference value :

[0064]

[0065]

[0066]

[0067] where, denotes the gating value of state output gate; denotes the hidden state at the current time step; is the final speed reference value after mapping; and denote the weight matrix and bias parameter of state output gate, respectively; represents a regression function for speed prediction.

[0068] Specifically, the speed estimation module (VPM) divides the original historical speed sequence and the path prior context vector according to a preset fixed window length during training, forming multiple training samples. Each training sample is composed of an input sequence (containing multiple time step feature vectors) and a corresponding target speed value (i.e., the "personalized normal speed baseline" at the future time).

[0069] Further, the steps of combining mobile phone GPS data for anomaly correction are as follows:

[0070] Compare the estimated stake number position with the feasible stake number interval determined by the mobile phone GPS data;

[0071] When the estimated stake number position is not within the feasible stake number interval, recalculate the position based on the downstream gantry;

[0072] When the estimated stake number position is within the interval but the deviation exceeds the threshold, use weighted average to fuse the GPS position and the predicted position.

[0073] Specifically, compare the estimated stake number with the GPS position, associate the mobile phone GPS position information with the gantry GPS position information, use the real-time GPS coordinates provided by the vehicle mobile phone, and based on the topological structure of the road network and the geographic information system (GIS) data. By mapping the mobile phone GPS point to the nearest road segment and combining the gantry information, obtain the gantry section and the start and end stake numbers, form a feasible stake number interval , compare the estimated stake number position calculated by the basic position with the possible road segment range determined by the GPS information. If the estimated stake number position exceeds the feasible road segment range determined by the mobile phone GPS data, it is considered that the estimated position is abnormal.

[0074] Abnormal correction mechanism, when it is detected that the estimated stake number position is abnormal (i.e. or ), the system will activate the abnormal correction mechanism. The correction strategy is as follows:

[0075] (1) Predicted error downstream gantry; if the estimated position significantly exceeds the downstream gantry (v ) that the vehicle can pass through, and the mobile phone GPS data points to the vicinity of the downstream gantry, then preferentially adopt the mobile phone GPS data, and use the downstream gantry as the new reference point to recalculate the vehicle position. At this time, the new estimated position can be updated as:

[0076] .

[0077] (2) position information correction; if the estimated position Although in the feasible section, but there is a large deviation from the nearest road stake number of the mobile phone GPS position beyond the set tolerance threshold, the weighted average method is used to fuse the basic position deduction result and the mobile phone GPS position information to obtain more accurate position estimation. The position after fusion is expressed as:

[0078] ;

[0079] Among them, is a weight factor, which is dynamically adjusted according to the confidence of mobile phone GPS data, the reliability of gantry data and the confidence of prediction model output. is the estimated position before fusion.

[0080] Further, the multi-source data includes ETC transaction records, service area parking information, road network topology, time periodicity characteristics and group preference profile.

[0081] Further, it also includes a sliding update mechanism: dynamically updating the vehicle passing trajectory and gantry identification result during the alarm process, supporting license plate number change or supplementary recording.

[0082] Further, the data involving user sensitive information is encrypted during transmission.

[0083] The above technical scheme is adopted in the present application, and the speed estimation is optimized by using path prediction. The path prediction (PPM) module is constructed by using the self-attention mechanism to learn the road network topology and the global path preference of the vehicle, and the trajectory prediction problem is solved. Then, the path prior context vector output by the PPM module is input as a condition to realize the relationship modeling between multiple tasks, and the speed prediction (VPM) module is constructed and optimized. The expected driving speed of the predicted path is estimated.

[0084] The present application has the following advantages and beneficial effects over the prior art:

[0085] 1. Fast positioning, realizing second-level response, fusing rescue call platform incoming information and ETC gantry real-time data, and completing positioning calculation and driving direction judgment within a few seconds after the phone is connected, realizing true second-level response, greatly shortening the rescue response time, and reducing the risk of personnel casualties and traffic congestion.

[0086] 2. High-precision positioning and direction prediction (solving the coupling gap). Traditional mobile phone signaling or GPS positioning can only provide approximate latitude and longitude information, and cannot be accurately positioned to highway lane, AB lane and stake number. More importantly, the traditional method cannot solve the strong coupling of vehicle "path selection" and "speed dynamics".

[0087] 3. Data fusion and generalization capability. Through a unified data access interface, multi-source data such as alarm calls, ETC gantries, group preference portraits, and geographic topologies are fused, and the system realizes full-link collaborative analysis. The key is that the present application does not rely on the individualized historical portrait of the vehicle, but learns the road network topology and group preferences through a global topology attention mechanism, so that it can still achieve high prediction accuracy for "strange vehicles" that are completely new in the verification set. At the same time, the introduction of the sliding update mechanism enables the vehicle information to be dynamically updated and supplemented during the alarm process. Even if the license plate number is changed or supplemented, the calculation results can be corrected in real time to maintain data continuity and accuracy, significantly enhancing the fault tolerance and practicality of the system.

[0088] 4. Multiple positioning comparison, improving research and judgment efficiency, supporting the generation of multiple positioning results for real-time comparison by alarm personnel, effectively reducing the dependence on manual inquiries. Reducing the positioning error caused by unclear description or chaotic post number by the alarm person, making the rescue command more efficient and smooth.

[0089] 5. High compatibility, easy to integrate and promote. The system uses standardized interface specifications to seamlessly integrate with rescue call platforms and highway emergency management platforms, is easy to deploy, and can be quickly promoted on existing infrastructure, reducing upgrade costs and operational maintenance difficulties. BRIEF DESCRIPTION OF DRAWINGS

[0090] The present application will be further described in detail below in conjunction with the drawings and specific embodiments;

[0091] Fig. 1 The flowchart of the emergency rescue vehicle auxiliary positioning method of the present application fusing multi-source data;

[0092] Fig. 2 The vehicle probability portrait calculation flowchart of the present application;

[0093] Fig. 3 The algorithm architecture diagram of the spatiotemporal network multi-task relationship model (ST-MTR) of the present application;

[0094] Fig. 4 The rescue call platform access algorithm architecture diagram of the present application. DETAILED DESCRIPTION

[0095] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0096] As shown in one of the drawings, Figs. 1 to 4 The present application discloses an emergency rescue vehicle auxiliary positioning method fusing multi-source data, comprising the following steps:

[0097] Step 1, data collection and preprocessing, mainly through the use of big data technology for multi-source data collection and fusion processing, the key data sources include: ETC issuance data: data collected when ETC card is issued, including vehicle license plate, ETC card number, reserved phone and other information. ETC system transaction data: vehicle pass data collected by ETC toll system on the highway, including license plate number, passid (pass unique identifier), transaction time, transaction node, which includes ETC gantry and toll station. Service area data: vehicle records in and out of service area, including service area name, license plate number, entry time, whether to make out, make out time, used to judge whether the vehicle enters the service area, so as to exclude the misjudgment of the vehicle caused by service area stay.

[0098] Step 2, receiving alarm event and extracting vehicle identification; after receiving the alarm event from the rescue call platform, the system extracts the alarm mobile phone number (PhoneID) and the alarm time (TimeAlarm). By querying the ETC issuance data, the system binds PhoneID with license plate number, and then obtains the unique identification (PlateID) of the vehicle. This process provides necessary identification data for subsequent vehicle positioning and accident analysis.

[0099] Step 3, vehicle current trajectory information extraction; once the vehicle identification PlateID is determined, the system will extract all transaction records of the vehicle from the ETC gantry pass database. Each record consists of:

[0100] ;

[0101] Among them, Passid is the unique trajectory number, Node is the identification of toll station or gantry, Time is the corresponding transaction time.

[0102] By sorting the transaction data in chronological order and grouping according to Passid, the system generates the time sequence pass trajectory of the vehicle:

[0103] ;

[0104] Among them, is the first node of the time sequence pass trajectory; is the transaction time of the first node of the time sequence pass trajectory; in turn, is the first node of the time sequence pass trajectory; is the transaction time of the first node of the time sequence pass trajectory.

[0105] Based on the alarm time (TimeAlarm), the system extracts the trajectory closest to the alarm time from the above trajectory data, and the last node of the trajectory is regarded as the latest node before the accident occurs.

[0106] The latest node before the accident occurs is extracted from the gantry information table The lane information (such as lane , stake number , route name ) of the latest node before the accident occurs is taken as the upstream reference gantry information of the accident , and the expression is: .

[0107] Step 4, fusion of service area data and road network topology analysis; in order to further improve the accuracy of location prediction, the system fuses the service area data with the road network topology structure, and calculates the actual passing time of the vehicle in the accident section. First, the system calculates the time difference between the two time points by the following formula: ; is the vehicle exit time in the accident section; is the vehicle entry time in the accident section.

[0108] Then, according to the actual driving trajectory of the vehicle, combined with the toll station and service area entrance and exit information, the real driving time of the vehicle in the accident section is calculated. The actual passing time of the vehicle in the accident section is calculated by fusing the service area data, taking the upstream reference gantry node as the starting point, searching all adjacent road sections based on the road network topology structure, and fusing the toll station information to construct the target toll station set . The expression is as follows:

[0109] ;

[0110] wherein, represents the total number of downstream nodes or toll stations that have a direct topological connection relationship with the reference node ; the purpose of constructing this set is to delimit the possible driving-off range of the vehicle, so as to search whether there is a corresponding driving-off passing record of the vehicle.

[0111] Traverse the section to find the topological structure of the road section it is located in and the related service area entrance and exit information, traverse the service area entrance and exit data, and search whether there is a passing record corresponding to the license plate number. If there is, calculate the residence time of the vehicle in the service area . According to this, the actual driving time of the vehicle in the accident section is calculated :

[0112] ;​

[0113] wherein, is a function of the time difference between two time points, is the alarm time, is the transit time of the vehicle passing the upstream gantry.

[0114] Table 1 Comparison of prediction performance of different model positions

[0115]

[0116] Step 5, path prediction based on ST-MTR architecture (PPM module construction); the present application constructs a path prediction module (PPM) to deconstruct the potential topological intention of the vehicle from the historical discrete trajectory of the vehicle. Unlike traditional sequence models, the present module adopts a global topological attention mechanism (Global Topological Attention Mechanism) to capture the global path preference of the vehicle in a complex road network by calculating the correlation strength between road network nodes.

[0117] Vehicle spatiotemporal trajectory tensorization: extracting the gantry transit records of the vehicle to construct the vehicle spatiotemporal state sequence wherein contains node identification embedding and timestamp encoding.

[0118] Global topological feature capture: constructing a self-attention-based topological encoder to convert the input sequence into a query space (Query), a key space (Key), and a feature space (Value) through a mapping function, respectively denoted as . The dependence weight matrix between road network nodes is calculated using multi-subspace projection , and the formula is as follows:

[0119] ;

[0120] wherein, is the input trajectory feature tensor, is the scaling factor of the feature space. The process outputs the hidden layer state representing the global intention distribution of the vehicle under the current road network topology.

[0121] Path probability distribution generation: projecting to the road network node space through a fully connected mapping layer to output the predicted probability distribution vector of the next gantry node.

[0122] ;

[0123] wherein, is the weight matrix of the path prediction layer, bias parameters for path prediction layer, global intent distribution, normalization exponential function.

[0124] Build inter-task dependency bridge: in order to realize the asymmetric conditional dependence between multi-tasks, the PPM module not only outputs the prediction result, but also generates a "path prior context vector" through a learnable feature compression operator :

[0125] ;

[0126] where, is the weight matrix of the feature compression operator, is the bias parameter of the feature compression operator, is the activation function.

[0127] This vector condenses the topological intent of the macroscopic path, and will be passed to the subsequent VPM module as a "conditional constraint".

[0128] Step 6, speed estimation based on asymmetric conditional dependence (VPM module construction). As shown in Table 2, the present application constructs a speed prediction module (VPM), and the core innovation lies in breaking the limitation of traditional speed prediction being carried out independently, and constructing a task-condition recursive unit (TC-RU). When performing time series deduction, the unit introduces the generated in step 5 as a global constraint, realizing the logical closed loop of "path optimization speed".

[0129] State space reconstruction: at each time step of VPM , the input space is reconstructed into a "observation-prior" mixed vector. Define as the historical speed feature at the current time, then the reconstructed input tensor is defined as:

[0130] ;

[0131] Task-condition recursive dynamics equation: use the TC-RU unit to capture the time series evolution law of speed data. In order to accurately control the information flow, define three dynamic gates: information retention gate , information injection gate and state output gate . The dynamics evolution equation is as follows:

[0132] (1) Gate activation calculation:

[0133] ;

[0134] ;

[0135] in, These represent the gate control values ​​for the information retention gate and the information injection gate, respectively. This represents the hidden state at the previous moment; This represents a hybrid vector at the current moment, which concatenates historical velocity features with path prior context. This represents the Sigmoid activation function; and These represent the weight matrix and bias parameters of the information retention gate, respectively. and These represent the weight matrix and bias parameters of the information injection gate, respectively.

[0136] (2) Memory manifold update: Calculate the candidate memory states at the current time step. And update the global memoiform manifold. :

[0137] ;

[0138] ;

[0139] in, This represents the candidate memory state calculated at the current moment; and Let these represent the global memory manifold at the previous time step and the global memory manifold at the current time step, respectively. and represent the weight matrix and bias parameters of the candidate memory state computation layer, respectively; tanh represents the hyperbolic tangent activation function. It represents the Hadamardi (or Hadama) stack.

[0140] (3) Hidden state and velocity mapping: Update the hidden state at the current time step. And mapped to the final velocity baseline value. :

[0141] ;

[0142] ;

[0143] ;

[0144] in, Indicates the gate value of the state output gate; This represents the hidden state at the current moment; This is the final velocity baseline value output after mapping; and These represent the weight matrix and bias parameters of the state output gate, respectively; represents the regression function for speed prediction. Through the above mechanism, the VPM module is not just doing time series regression, but is deducing the speed baseline that best fits the path characteristics within the delineated "path intent range".

[0145] Time series division in model training: The original historical speed sequence and path prior context vector (i.e. the input feature vector in step 6.1) are divided according to the pre-set fixed window length to form multiple training samples. Each training sample is composed of an input sequence (containing multiple time step feature vectors) and a corresponding target speed value (i.e. the "personalized normal speed baseline" at the future time). Feature normalization: Z-score standardization is performed on all input features to unify different dimensions and value ranges into similar numerical ranges, improving the stability and efficiency of model training.

[0146]

[0147] Table 2 Performance comparison of different models on speed prediction (speed baseline) task

[0148]

[0149] Step 7, calculation of vehicle precise position based on ST-MTR: aims to comprehensively use vehicle historical trajectory information, gantry pile number data, speed prediction results and time deduction algorithm to achieve high-precision calculation of the current position of the vehicle. Through multi-source information fusion technology, the system can accurately estimate the real-time position of the vehicle in the case of vehicle disconnection or abnormality, providing key spatial positioning support for emergency rescue and traffic control.

[0150] Basic position calculation: the system takes the last successfully detected gantry of the vehicle as the reference point. By combining the upstream gantry pile position in step 3 , the actual passing time of the vehicle in step 5 and the predicted future driving speed of the vehicle in step 7 , the current pile position of the vehicle can be deduced. The mathematical expression is as follows:

[0151] ;

[0152] Multi-source data fusion anomaly correction: this part is the key step of anomaly detection and position correction, which identifies and corrects the deviation of vehicle position estimation caused by model prediction error or data loss by fusing data from multiple sources.

[0153] ​(1) Estimation of pile number and GPS position comparison, mobile phone GPS position information and gantry GPS position information association, using the real-time GPS coordinates provided by the vehicle mobile phone, based on the topological structure of the road network and geographic information system (GIS) data. By mapping the mobile phone GPS point to the nearest road segment, and combining the gantry information, the gantry section And the start and end pile number, form a feasible pile number interval , the estimated pile number position calculated by the basic position And the possible range of road section determined by GPS information. If the estimated pile number position exceeds the feasible road section range determined by the mobile phone GPS data, it is considered that the estimated position is abnormal.

[0154] (2) Abnormal correction mechanism, when the estimated pile number position Abnormal (i.e. Or ), the system will activate the abnormal correction mechanism. The correction strategy is as follows:

[0155] (a) Predicting the error downstream gantry. If the estimated position Significantly exceeds the downstream gantry that the vehicle can pass through , and the mobile phone GPS data points to the vicinity of the downstream gantry, then the mobile phone GPS data is preferred, and the downstream gantry is taken as the new reference point to recalculate the vehicle position. At this time, the new estimated position Can be updated as: ;

[0156] (b) Position information correction. If the estimated position Although in the feasible road section, but there is a large deviation from the nearest road pile number, which exceeds the set tolerance threshold, then the weighted average method is used to fuse the basic position deduction result and the mobile phone GPS position information to obtain more accurate position estimation. The fused position Can be expressed as:

[0157] ;

[0158] Wherein, Is a weight factor, which is dynamically adjusted according to the confidence of the mobile phone GPS data, the reliability of the gantry data and the confidence of the prediction model output.

[0159] Step 8, the system outputs include vehicle route number, vehicle downstream gantry, vehicle pile number and other information, providing comprehensive position information support for traffic management and emergency rescue.

[0160] The application solves the path problem by using a path prediction module (PPM) first, and then uses the path prior as a condition to optimize the speed estimation module (VPM) based on asymmetric conditional dependence to solve the speed baseline problem. This cascading mechanism makes the prediction accuracy of the speed baseline higher than that of the independent model without path guidance, thereby accurately reconstructing the vehicle driving path, accurately inferring the current position, driving direction and stake point of the vehicle.

[0161] The application has the following advantages and beneficial effects over the prior art:

[0162] 1. Fast positioning, realizing second-level response, integrating rescue call platform incoming information and ETC gantry real-time data, positioning calculation and driving direction judgment can be completed within a few seconds after the phone is connected, realizing true second-level response, greatly shortening the rescue response time and reducing the risk of personnel casualties and traffic congestion.

[0163] 2. High-precision positioning and direction prediction (solving the coupling gap). Traditional mobile phone signaling or GPS positioning can only provide approximate latitude and longitude information, and cannot accurately reach the highway lane, AB lane and stake number position. More importantly, traditional methods cannot solve the strong coupling of vehicle "path selection" and "speed dynamics".

[0164] 3. Data fusion and generalization ability. Through a unified data access interface, multi-source data such as alarm phone, ETC gantry, group preference portrait and geographic topology structure are fused, and the system realizes full-link collaborative analysis. The key is that the application does not rely on the individualized historical portrait of the vehicle, but learns the road network topology and group preferences through a global topology attention mechanism, so that it can still achieve high prediction accuracy for "strange vehicles" that are completely new in the verification set. At the same time, the introduction of the sliding update mechanism enables the vehicle information to be dynamically updated and supplemented during the alarm process, so that even if the license plate number is changed or supplemented, the calculation result can be corrected in real time, maintaining data continuity and accuracy, and significantly enhancing the fault tolerance and practicality of the system.

[0165] 4. Multiple positioning comparison to improve judgment efficiency and support the generation of multiple positioning results for real-time comparison by the alarm personnel, effectively reducing the dependence on manual inquiry. Reducing the positioning error caused by unclear description by the alarm person or stake number confusion makes the rescue command more efficient and smooth.

[0166] 5. High compatibility, easy to integrate and promote. The system uses standardized interface specifications, seamlessly integrates with the rescue call platform and the highway emergency management platform, is easy to deploy, can be quickly promoted on existing infrastructure, and reduces the upgrade cost and difficulty of operation and maintenance.

[0167] It is apparent that the described embodiments are only some — but not all — of the embodiments of the present application. The embodiments described in this application and features in the embodiments can be combined with each other in cases without conflict. The components of the embodiments of the present application, which are generally described and shown in the accompanying drawings, can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work, shall fall within the scope of protection of the present application.

Claims

1. An emergency rescue vehicle assisted positioning method integrating multi-source data, characterized in that: It includes the following steps: Multi-source datasets are obtained by collecting and fusing data from multiple sources. Receive alarm events from the rescue call platform and extract the alarm mobile phone number and alarm time; ETC card issuance data can be queried based on mobile phone numbers, and vehicle license plate numbers can be bound to generate unique identifiers. Extract the time-series passage trajectory of the corresponding vehicle from the ETC gantry passage database, and use the last gantry node with the closest alarm time as the upstream reference gantry; By integrating service area data and road network topology information, the actual travel time of vehicles in the accident section can be calculated. Cascaded prediction is performed using a multi-task relationship model of a spatiotemporal network: first, the path prediction module generates the path probability distribution and the path prior context vector, and then the speed estimation module uses the context vector as a condition to input the predicted vehicle speed baseline. The estimated vehicle position is calculated based on the gantry station number of the upstream benchmark, the actual travel time, and the predicted speed, and anomaly correction is performed by combining mobile phone GPS data. It outputs vehicle route number, downstream gantry and mileage information to the rescue platform to provide location information support for traffic management and emergency rescue. The path prediction module employs a global topology attention mechanism, which includes the following steps: Convert vehicle gantry passage records into a spatiotemporal state sequence; The global intent distribution of vehicles in the current road network topology is obtained by calculating the dependency weight matrix of road network nodes through self-attention. Global intent distribution Projecting the path prediction probability distribution vector onto the road network node space generates the path prediction probability distribution vector for the next gantry node. ; ;in, This is the weight matrix of the path prediction layer. These are the bias parameters for the path prediction layer. For global intent distribution, It is a normalized exponential function; The predicted probability distribution vector generates a path prior context vector through a learnable feature compression operator. ;in, Here is the weight matrix of the feature compression operator. These are the bias parameters for the feature compression operator. For activation functions; The velocity estimation module is implemented using task-conditional recursive units, including the following steps: Historical speed features are concatenated with path prior context vectors to form a hybrid vector; Through the information retention gate Information injection gate and state output gate The dynamic gating mechanism is used to dynamically evolve and update the memory manifold, and output the velocity prediction value under the path prior constraint.

2. The emergency rescue vehicle assisted positioning method based on multi-source data according to claim 1, characterized in that: The steps to extract the time-series travel trajectory of the corresponding vehicle are as follows: Extract all transaction records for the corresponding vehicle from the ETC gantry passage database; By sorting transaction record data in chronological order and grouping it according to the trajectory number Passid, a time-series passage trajectory of the vehicle is generated. Extract the trajectory closest to the alarm time from the time-series traffic trajectory data, and then extract the last node of the corresponding trajectory. This is considered the latest point in time before the accident occurred. Extract the latest node before the event from the gantry information table. The lane information is used as the upstream reference gantry for the accident.

3. The emergency rescue vehicle auxiliary positioning method based on multi-source data according to claim 1, characterized in that: The steps for calculating the actual travel time of a vehicle in the accident area are as follows: Traverse the segment to find its road topology and related service area entrance / exit information, then traverse the service area entrance / exit data to check if the corresponding vehicle license plate number exists. The corresponding passage record; if it exists, the dwell time of the corresponding vehicle in the service area is calculated. Calculate the actual travel time of the vehicle within the accident area. : ； in, A function to calculate the time difference between two points in time. For alarm time, This refers to the passage time of a vehicle through the upstream reference gantry.

4. The emergency rescue vehicle auxiliary positioning method based on multi-source data according to claim 1, characterized in that: The dynamic evolution equations are as follows: (1) Gated activation calculation: ； ； in, These represent the gate control values ​​for the information retention gate and the information injection gate, respectively. This represents the hidden state at the previous moment; This represents a hybrid vector at the current moment, which concatenates historical velocity features with the path prior context. This represents the Sigmoid activation function; and These represent the weight matrix and bias parameters of the information retention gate, respectively. and These represent the weight matrix and bias parameters of the information injection gate, respectively. (2) Memory manifold update: Calculate the candidate memory states at the current time step. And update the global memoiform manifold. : ； ； in, This represents the candidate memory state calculated at the current moment; and Let represent the global memory manifold at the previous time step and the global memory manifold at the current time step, respectively; and These represent the weight matrix and bias parameters of the candidate memory state computation layer, respectively. Represents the hyperbolic tangent activation function; It represents the Hadamardi (or Hadama) stack; (3) Hidden state and velocity mapping: Update the hidden state at the current time step. And mapped to the final velocity baseline value. : ； ； ； in, Indicates the gate value of the state output gate; This represents the hidden state at the current moment; This is the final velocity baseline value output after mapping; and These represent the weight matrix and bias parameters of the state output gate, respectively; This represents the regression function used for velocity prediction.

5. The emergency rescue vehicle auxiliary positioning method based on multi-source data according to claim 1, characterized in that: The method for calculating the estimated vehicle station position is as follows: Based on the corresponding position of the upstream reference gantry Actual travel time of vehicles in the accident section and the baseline value of the estimated future speed of the vehicle The estimated station position of the current vehicle is calculated. ; .

6. The emergency rescue vehicle auxiliary positioning method based on multi-source data according to claim 1 or 4, characterized in that: The steps for anomaly correction using mobile phone GPS data are as follows: Compare the estimated station location with the feasible station range determined by mobile phone GPS data; If the estimated station location is not within the feasible station range, the location shall be recalculated based on the downstream gantry. If the estimated station location is within the range but the deviation exceeds the threshold, a weighted average fusion of the GPS location and the predicted location is used.

7. The emergency rescue vehicle auxiliary positioning method based on multi-source data according to claim 1, characterized in that: Multi-source data includes ETC transaction records, service area parking information, road network topology, time periodic characteristics, and group preference profiles.

8. The emergency rescue vehicle auxiliary positioning method based on multi-source data according to claim 1, characterized in that: It also includes a sliding update mechanism: dynamically updating the vehicle's travel trajectory and gantry recognition results during the alarm process, supporting license plate number changes or re-entry.

Images