A method for automatic identification, tracking and positioning of vehicles in a railway yard

Abstract

The application discloses a kind of railway station yard vehicle automatic identification and tracking positioning method, it is related to railway transport management technical field, it is difficult to solve the technical problem that three-level judging mechanism is combined with multi-sensor data to identify vehicle, it is also difficult to realize the automatic conversion of coordinate to label by digital twin model, lack of weight adjustment and joint loss function based on prediction variance, balance multi-task training effect;Including: by building comprehensive monitoring system to collect multi-source data, realize vehicle automatic identification in combination with three-level judging mechanism;Vehicle position information is calculated using Kalman filtering algorithm and railway station yard special coordinate system, and the coordinate is mapped into semantic position label by digital twin model;Based on historical data and semantic label, a multi-task neural network framework is constructed, the loss function weight is dynamically adjusted to optimize the joint loss function, efficient semantic prediction and classification of vehicle position are realized.

Description

Technical Field

[0001] This invention belongs to the field of railway transportation management technology, specifically relating to a method for automatic vehicle identification and tracking positioning within railway stations. Background Technology

[0002] With the rapid development of the railway transportation industry, the flow of vehicles in railway stations is increasing day by day. Traditional manual identification and positioning methods can no longer meet the needs of efficient and accurate operation of modern railway transportation. In the complex station environment, accurate and rapid identification of vehicle information and real-time monitoring of vehicle location and trajectory are crucial for optimizing station scheduling, improving operational efficiency and ensuring transportation safety.

[0003] However, while some existing vehicle recognition and localization technologies have achieved vehicle recognition and localization to a certain extent, they have the following problems: reliance on a single sensor leads to a decrease in recognition accuracy; lack of a dedicated coordinate system and multi-sensor fusion algorithm results in large position calculation errors, and semantic location labeling requires manual intervention, which is inefficient; and failure to consider the correlation between tasks, with each task being trained independently, leads to redundant model parameters, wasted resources, and difficulty in dynamically adapting to changes in data distribution.

[0004] Therefore, there is an urgent need for a vehicle identification and tracking method that integrates multi-source data, has semantic mapping capabilities, and supports dynamic optimization. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a method for automatic vehicle identification, tracking and positioning within railway stations, to solve the following technical problem:

[0006] Reliance on a single sensor leads to a decrease in recognition accuracy; the lack of a dedicated coordinate system and multi-sensor fusion algorithm results in large position calculation errors, and semantic position labeling requires manual intervention, which is inefficient; the failure to consider the correlation between tasks and the independent training of each task lead to redundant model parameters, wasted resources, and difficulty in dynamically adapting to changes in data distribution.

[0007] To address the above problems, this invention provides a method for automatic vehicle identification, tracking, and positioning within railway stations, comprising the following steps:

[0008] S1: Construct a comprehensive monitoring system for railway station areas to collect various types of data and perform synchronous fusion processing;

[0009] S2: Set up a three-level judgment mechanism to calculate the first, second and third matching degree scores respectively, add them with weights to obtain the comprehensive matching degree score, and automatically identify the vehicle based on the comparison result of the comprehensive matching degree score and the comprehensive matching degree threshold.

[0010] S3: After successful automatic vehicle identification, a tracking database is constructed, and a railway station coordinate system is established. The Kalman filter algorithm is used to integrate multi-source data to obtain the two-dimensional plane coordinates and height information of the vehicle in the railway station coordinate system.

[0011] S4: Utilize RFID readers and high-definition industrial cameras distributed within railway stations to dynamically acquire vehicle coordinates and synchronize them to the tracking database. Through the digital twin model of the railway station, the vehicle coordinates are mapped to manually labeled functional areas and associated track numbers. At the same time, complete semantic information is generated by combining status tags.

[0012] S5: Construct a multi-task neural network framework using historical data and semantic location labels, and define a joint optimization loss function through dynamic weight adjustment and joint loss optimization.

[0013] Preferably, the construction of the integrated monitoring system inside the railway station includes high-definition industrial cameras, laser sensors, radar, RFID tags and readers, and Beidou positioning terminals.

[0014] Preferably, the three-level judgment mechanism in step S2 includes the following steps:

[0015] The data collected by the integrated monitoring system inside the railway station are preprocessed respectively;

[0016] A three-level judgment mechanism is set up, specifically as follows:

[0017] Extract vehicle visual features, including vehicle body color, logo patterns, number of windows and shape of the front of the vehicle, match them with a pre-established vehicle feature database, and make a first visual judgment based on the matching results;

[0018] The vehicle's outline, dimensions, and distance from the track measured by the laser sensor are matched with a standard vehicle parameter library, and a second physical judgment is made based on the matching results.

[0019] The integrity and occlusion of vehicle tag information read by RFID readers are used to make third-party information judgments;

[0020] The overall matching score is obtained by weighted fusion of visual judgment, physical judgment and RFID information.

[0021] Preferably, the step of weightedly fusing visual judgment, physical judgment, and RFID information to obtain a comprehensive matching score includes the following steps:

[0022] The extracted vehicle visual features are matched with a preset vehicle feature database to calculate the first matching score; the physical parameters measured by the laser sensor are matched with a standard vehicle parameter database to calculate the second matching score; and the tag information read by RFID is used as the basis to calculate the third matching score based on the tag integrity and occlusion.

[0023] Weights are assigned to each data source based on historical data, and the weighted sums are then used to obtain a comprehensive matching score, specifically:

[0024]

[0025] in, To calculate the overall matching score, , and These are the first matching score, the second matching score, and the third matching score, respectively. , and These are the corresponding weights;

[0026] Based on a pre-set reliability, the confidence interval for the overall matching score is determined using historical training data, and the overall matching score threshold is defined as follows:

[0027]

[0028] in The overall matching degree threshold, This represents the average of the historical overall matching scores. Standard deviation, Confidence coefficient;

[0029] like > If the match is successful, the vehicle recognition result will be output; otherwise, it will be determined as a matching conflict and the matching inconsistency handling process will be triggered.

[0030] Preferably, the calculation of the third matching score based on tag integrity and occlusion includes the following steps:

[0031] The system reads vehicle tag information using an RFID reader, including key fields such as vehicle number and vehicle model, and also obtains the completeness and occlusion of the tag.

[0032] The integrity level is determined by whether the label is intact and undamaged; the occlusion level is determined by calculating the proportion of the label that is obscured by an object.

[0033] Based on label integrity and occlusion, label status is categorized into the following levels, with different weights assigned according to each quality level:

[0034] Level 1 Quality: Integrity ≥ 95%, Occlusion ≤ 5%, corresponding weight: 1;

[0035] Level 2 Quality: Integrity ≥ 70%, Occlusion ≤ 30%, corresponding weight: 0.7;

[0036] Level 3 Quality: Integrity < 70%, Occlusion > 30%, corresponding weight set to 0.4;

[0037] The corresponding weight is determined based on the tag quality level, and multiplied by 100 to obtain the third matching score.

[0038] Preferably, the process for triggering inconsistency matching includes the following steps:

[0039] S201: Data review, including reviewing the data collected by the integrated monitoring system inside the railway station to check for any data collection errors or omissions, and at the same time, activating backup monitoring equipment to obtain vehicle information;

[0040] S202: Cross-validation, which includes the process of matching the reviewed data and the newly collected data again using a three-level judgment mechanism;

[0041] S203: Manual assistance in judgment, including, if there are still inconsistencies in matching after data review and cross-validation, a manual assistance in judgment is introduced to perform the final vehicle identification.

[0042] Preferably, step S3 includes the following steps:

[0043] After a vehicle is successfully identified, a tracking database is built for it;

[0044] The tracking database records vehicle identification information and semantic location tags;

[0045] A railway station-specific coordinate system is adopted, with the track start point as the origin and the track direction as the reference axis.

[0046] By fusing data from multiple sensors using the Kalman filter algorithm, the two-dimensional plane coordinates and height information of the vehicle in the railway station coordinate system are calculated.

[0047] Preferably, step S4 includes the following steps:

[0048] The vehicle location coordinates are updated in real time using RFID readers and high-definition industrial cameras distributed within the railway station.

[0049] The location coordinates are mapped in real time to functional area labels, status labels, and associated labels using a data twin model, and then synchronized to the tracking database.

[0050] The tracking and positioning system updates the vehicle's location coordinates in the tracking file based on the received location. At the same time, after updating the vehicle coordinates, it queries the railway station's digital twin model in real time, maps the coordinates to functional area labels and associated labels, and stores them synchronously with the status labels in the tracking database.

[0051] The railway station digital twin model is created by converting railway station CAD drawings into a three-dimensional mesh model, manually annotating the functional areas of each mesh, and matching preset labels according to the mesh index where the vehicle coordinates are located.

[0052] Preferably, step S5, which involves constructing a multi-task neural network framework using historical data and semantic location labels, includes the following steps:

[0053] Vehicle data with location coordinates is collected by combining historical operational data of railway stations with manually labeled semantic location tags, which include functional area tags, status tags, and association tags.

[0054] The neural network extracts location coordinates, timestamps, raw sensor data, and external data features as input features. The neural network automatically learns the correlation between tasks by sharing underlying parameters.

[0055] The input features are processed through a shared preprocessing layer, which includes standardizing coordinate values, extracting temporal features, and encoding classification features.

[0056] Use a multi-task branching structure, including: functional area label branches, status label branches, and related label branches;

[0057] The functional area label branch: input coordinates, time and vehicle type characteristics, output grid function classification;

[0058] The status label branch takes speed, historical trajectory, and location features as inputs and outputs status labels.

[0059] The associated label branch: input coordinates and track status, output current track number.

[0060] Preferably, the joint optimization loss function is defined in step S5, which includes the following steps:

[0061] The weights of the loss function are dynamically adjusted based on the prediction variance of each task, specifically as follows:

[0062]

[0063] in, The weight coefficients corresponding to the functional area labels. The variance of the functional area label prediction results. The variance of the state region label prediction results. The variance of the associated label prediction results;

[0064] Calculate the weight coefficients corresponding to the state region labels and associated labels using the same method;

[0065] The weight coefficients corresponding to the labels are weighted and added to the loss functions corresponding to the labels to construct a joint optimization loss function;

[0066] The joint optimization loss function is specifically as follows:

[0067]

[0068] in, To jointly optimize the loss function, , and These are the weight coefficients corresponding to the labels. , and These are the loss functions corresponding to the labels, where... For focus loss function, and Both are cross-entropy loss functions.

[0069] The beneficial effects of this invention are:

[0070] This invention improves vehicle identification accuracy by fusing multi-source data through a three-level judgment mechanism (visual, physical, and RFID) and achieves accurate calculation of vehicle two-dimensional plane coordinates and height information based on the Kalman filter algorithm and a dedicated coordinate system.

[0071] This invention utilizes a digital twin model to map vehicle coordinates into functional areas, states, and associated track labels, reducing manual labeling costs and improving label generation efficiency. Combined with a multi-task neural network, it automatically learns the relationships between tasks, achieving end-to-end semantic prediction of "location-function-state".

[0072] This invention uses RFID and high-definition cameras to update vehicle coordinates in real time, and tracks semantic tags synchronously in the database, supporting fast query and trajectory backtracking. It is suitable for complex railway station environments, improves the level of intelligent management of railway stations, and dynamically adjusts the weights of the loss function based on the predicted variance, thereby enhancing the robustness of the model to difficult-to-classify samples. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0074] Figure 2 This is a schematic diagram of the module flow of the integrated monitoring system for railway station interiors of the present invention;

[0075] Figure 3 This is a schematic diagram of the three-level judgment mechanism method of the present invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Please see Figure 1 As shown, the present invention provides a method for automatic vehicle identification, tracking, and positioning within a railway station, comprising the following steps:

[0078] S1: Construct a comprehensive monitoring system for railway station areas to collect various types of data and perform synchronous fusion processing;

[0079] S2: Set up a three-level judgment mechanism to calculate the first, second and third matching degree scores respectively, add them with weights to obtain the comprehensive matching degree score, and automatically identify the vehicle based on the comparison result of the comprehensive matching degree score and the comprehensive matching degree threshold.

[0080] S3: After successful automatic vehicle identification, a tracking database is constructed, and a railway station coordinate system is established. The Kalman filter algorithm is used to integrate multi-source data to obtain the two-dimensional plane coordinates and height information of the vehicle in the railway station coordinate system.

[0081] S4: Utilize RFID readers and high-definition industrial cameras distributed within railway stations to dynamically acquire vehicle coordinates and synchronize them to the tracking database. Through the digital twin model of the railway station, the vehicle coordinates are mapped to manually labeled functional areas and associated track numbers. At the same time, complete semantic information is generated by combining status tags.

[0082] S5: Construct a multi-task neural network framework using historical data and semantic location labels, and define a joint optimization loss function through dynamic weight adjustment and joint loss optimization.

[0083] In one embodiment of the present invention, the construction of the integrated monitoring system inside the railway station includes a high-definition industrial camera, a laser sensor, radar, RFID tags and readers, and a Beidou positioning terminal.

[0084] Specifically, high-definition industrial cameras and supplementary lighting equipment are installed at high points (from a bird's-eye view) on both sides of the railway tracks within the railway station, covering multiple tracks and avoiding lens obstruction, to collect images of the locomotives and carriages entering and leaving the station, as well as marking information; laser sensors are installed on both sides or top of the tracks, triggered synchronously with the high-definition industrial cameras, to detect vehicle signals, speed, and distance information; radar is installed at both ends of the tracks or at bends for long-distance vehicle detection; RFID tags are affixed to the bottom of the vehicles to store unique identifiers such as car number and model; corresponding RFID readers are set up at key nodes on the tracks (entrance, shunting area); and Beidou positioning terminals are equipped on some vehicles or fixed base stations; NTP or PTP is used to align the clocks of all equipment to ensure that the data timestamp error is ≤10ms.

[0085] In one embodiment of the present invention, the three-level judgment mechanism in step S2 includes the following steps:

[0086] The data collected by the integrated monitoring system inside the railway station are preprocessed respectively;

[0087] A three-level judgment mechanism is set up, specifically as follows:

[0088] Extract vehicle visual features, including vehicle body color, logo patterns, number of windows and shape of the front of the vehicle, match them with a pre-established vehicle feature database, and make a first visual judgment based on the matching results;

[0089] The vehicle's outline, dimensions, and distance from the track measured by the laser sensor are matched with a standard vehicle parameter library, and a second physical judgment is made based on the matching results.

[0090] The integrity and occlusion of vehicle tag information read by RFID readers are used to make third-party information judgments;

[0091] The overall matching score is obtained by weighted fusion of visual judgment, physical judgment and RFID information.

[0092] Specifically, the system performs image denoising, grayscale conversion, and edge detection on visual data acquired by high-definition industrial cameras; extracts vehicle body color (HSV color space analysis), marking patterns (template matching or feature point recognition), number of windows (morphological analysis), and front shape (contour comparison); performs point cloud filtering and extracts vehicle contours from laser sensor data; calculates vehicle dimensions (length, width, height), centroid position, and distance from the track; decodes vehicle tag information, including key fields such as vehicle number and model, and assesses tag integrity (whether it is missing or damaged) and occlusion (the proportion of occluded area); performs coordinate transformation and timestamp alignment on BeiDou positioning data; and sets up a three-level judgment mechanism, where the first level... For visual feature matching, preprocessed visual features (color, logo, windows, front shape) are input, and the extracted visual features are compared with a preset vehicle feature database to calculate a first matching score. The second level is physical parameter matching, where the vehicle outline, dimensions, and distance to the track measured by a laser sensor are input. The vehicle's shape parameters are matched with a standard vehicle parameter database to calculate whether the distance between the vehicle and the track meets the standard safety range, generating a second matching score. The third level is RFID information judgment, where RFID tag information (vehicle number, vehicle type), completeness, and occlusion are input. Quality levels are assigned based on tag completeness and occlusion, and a third matching score is calculated. Among these, the preset... The vehicle feature database collects vehicle model samples, covering all common vehicle types in the station (such as freight locomotives, passenger cars, shunting locomotives, etc.), including images from different angles, lighting conditions, and weather conditions. High-definition industrial cameras are used to capture key parts of the vehicles, such as the front, sides, and front end, ensuring feature integrity. Vehicle body color, markings, number of windows, and front shape are labeled. Image processing software is used to extract features and store them as structured data. A standardized process is then implemented, including converting colors to the HSV color space, calculating the proportion of dominant colors, identifying specific markings through template matching or SIFT feature points, counting the number of window areas based on morphological analysis (contour detection), and extracting the front contour through edge detection. Define a shape library (trapezoidal, streamlined, etc.) for comparison, regularly supplement newly acquired image features, and eliminate old vehicle model data; the standard vehicle parameter library is obtained by extracting standard dimensions, centroid position, wheelbase and other parameters from vehicle manufacturer design manuals, using laser sensors or 3D measuring instruments to scan actual vehicles, obtain outline point cloud data, calculate actual dimensions through point cloud segmentation and fitting algorithms, and compare and verify with design values. The standardization process includes using the maximum outline of the vehicle as the standard, eliminating retractable parts, calculating the centroid based on design documents or through laser point cloud, and defining the safe distance range between the vehicle and the track. Regularly (re-measure the dimensions of in-service vehicles and update the parameters of aging or modified models in the database).

[0093] In one embodiment of the present invention, the step of weightedly fusing visual judgment, physical judgment, and RFID information to obtain a comprehensive matching score includes the following steps:

[0094] The extracted vehicle visual features are matched with a preset vehicle feature database to calculate the first matching score; the physical parameters measured by the laser sensor are matched with a standard vehicle parameter database to calculate the second matching score; and the tag information read by RFID is used as the basis to calculate the third matching score based on the tag integrity and occlusion.

[0095] Weights are assigned to each data source based on historical data, and the weighted sums are then used to obtain a comprehensive matching score, specifically:

[0096]

[0097] in, To calculate the overall matching score, , and These are the first matching score, the second matching score, and the third matching score, respectively. , and These are the corresponding weights;

[0098] Based on a pre-set reliability, the confidence interval for the overall matching score is determined using historical training data, and the overall matching score threshold is defined as follows:

[0099]

[0100] in The overall matching degree threshold, This represents the average of the historical overall matching scores. Standard deviation, Confidence coefficient;

[0101] like > If the match is successful, the vehicle recognition result will be output; otherwise, it will be determined as a matching conflict and the matching inconsistency handling process will be triggered.

[0102] Specifically, the first matching score is calculated by weighting the degree of matching between visual features, including vehicle color, logo patterns, number of windows, and front shape, and the vehicle feature database.

[0103]

[0104] in, The score is the first match score. For the first The weights of each feature, For the first The similarity of the features, where , , and The values ​​were 0.2, 0.3, 0.2, and 0.3, respectively.

[0105] The second matching score is calculated by weighting the degree of matching between the vehicle's external parameters (length, width, height, and center of gravity position) measured by the laser sensor and the standard vehicle parameter library. It is calculated using the same method as the first matching score, with the weights for length, width, height, and center of gravity position being 0.4, 0.3, 0.2, and 0.1, respectively.

[0106] , and These are the weights corresponding to the first, second, and third matching scores, with values ​​of 0.5, 0.3, and 0.2, respectively.

[0107] In one embodiment of the present invention, the calculation of the third matching score based on tag integrity and occlusion includes the following steps:

[0108] The system reads vehicle tag information using an RFID reader, including key fields such as vehicle number and vehicle model, and also obtains the completeness and occlusion of the tag.

[0109] The integrity level is determined by whether the label is intact and undamaged; the occlusion level is determined by calculating the proportion of the label that is obscured by an object.

[0110] Based on label integrity and occlusion, label status is categorized into the following levels, with different weights assigned according to each quality level:

[0111] Level 1 Quality: Integrity ≥ 95%, Occlusion ≤ 5%, corresponding weight: 1;

[0112] Level 2 Quality: Integrity ≥ 70%, Occlusion ≤ 30%, corresponding weight: 0.7;

[0113] Level 3 Quality: Integrity < 70%, Occlusion > 30%, corresponding weight set to 0.4;

[0114] The corresponding weight is determined based on the tag quality level, and multiplied by 100 to obtain the third matching score.

[0115] In one embodiment of the present invention, the triggering of the mismatch processing flow includes the following steps:

[0116] S201: Data review, including reviewing the data collected by the integrated monitoring system inside the railway station to check for any data collection errors or omissions, and at the same time, activating backup monitoring equipment to obtain vehicle information;

[0117] S202: Cross-validation, which includes the process of matching the reviewed data and the newly collected data again using a three-level judgment mechanism;

[0118] S203: Manual assistance in judgment, including, if there are still inconsistencies in matching after data review and cross-validation, a manual assistance in judgment is introduced to perform the final vehicle identification.

[0119] In one embodiment of the present invention, step S3 includes the following steps:

[0120] After a vehicle is successfully identified, a tracking database is built for it;

[0121] The tracking database records vehicle identification information and semantic location tags;

[0122] A railway station-specific coordinate system is adopted, with the track start point as the origin and the track direction as the reference axis.

[0123] By fusing data from multiple sensors using the Kalman filter algorithm, the two-dimensional plane coordinates and height information of the vehicle in the railway station coordinate system are calculated.

[0124] Specifically, the tracking database includes vehicle ID, recognition time, semantic location label, two-dimensional plane coordinates, height, speed, orientation angle, and confidence level. A new record is inserted each time a vehicle is successfully recognized and its coordinates are updated. If the timestamp of an old record for the same vehicle ID exceeds a threshold (e.g., 10 minutes), it is overwritten and updated. An index is created for "vehicle ID" and "recognition time" to accelerate historical trajectory queries. The track starting point is the intersection of the station signal tower centerline and the track. The local coordinate systems of sensors such as LiDAR and cameras are converted to the railway station coordinate system, specifically: P1 = T * P2 + R, where P1 is the railway station coordinate system, P2 is the sensor's local coordinate system, T is the rotation matrix calculated based on the installation angle, and R is the translation vector. Multi-sensor data, including LiDAR data, high-definition industrial camera data, and track encoder data, is fused using a Kalman filter algorithm.

[0125] In one embodiment of the present invention, step S4 includes the following steps:

[0126] The vehicle location coordinates are updated in real time using RFID readers and high-definition industrial cameras distributed within the railway station.

[0127] The location coordinates are mapped in real time to functional area labels, status labels, and associated labels using a data twin model, and then synchronized to the tracking database.

[0128] The tracking and positioning system updates the vehicle's location coordinates in the tracking file based on the received location. At the same time, after updating the vehicle coordinates, it queries the railway station's digital twin model in real time, maps the coordinates to functional area labels and associated labels, and stores them synchronously with the status labels in the tracking database.

[0129] The railway station digital twin model is created by converting railway station CAD drawings into a three-dimensional mesh model, manually annotating the functional areas of each mesh, and matching preset labels according to the mesh index where the vehicle coordinates are located.

[0130] Specifically, RFID readers and industrial cameras send vehicle location data to edge computing nodes. The edge nodes fuse multi-source data, using a Kalman filter algorithm to integrate RFID, camera, and track encoder data, and output vehicle location coordinates. The data twin model mapping includes locating the grid index in the 3D mesh model based on the vehicle's location coordinates (x, y), querying the rule base to return functional labels, and generating status labels. If the speed v ≤ 0.5 m / s and lasts > 60 seconds, it is marked "timeout stop"; if the trajectory deviates from the planned path by > 10 meters, it is marked "path deviation". The associated coordinate generation includes querying the train dispatch system, matching the task corresponding to the current vehicle ID, and associating the track occupancy status. The station CAD drawings are imported into BIM software to generate a 3D mesh model. The mesh precision is divided according to the complexity of the functional areas, for example, the grid of the turnout area is ≤ 1m × 1m, and the grid of the track is ≤ 5m × 5m. Each grid is assigned a functional label and associated attributes. The grid where the vehicle is located is queried based on its location coordinates (x, y), and a preset label is returned.

[0131] In one embodiment of the present invention, step S5, which involves constructing a multi-task neural network framework using historical data and semantic location labels, includes the following steps:

[0132] Vehicle data with location coordinates is collected by combining historical operational data of railway stations with manually labeled semantic location tags, which include functional area tags, status tags, and association tags.

[0133] The neural network extracts location coordinates, timestamps, raw sensor data, and external data features as input features. The neural network automatically learns the correlation between tasks by sharing underlying parameters.

[0134] The input features are processed through a shared preprocessing layer, which includes standardizing coordinate values, extracting temporal features, and encoding classification features.

[0135] Use a multi-task branching structure, including: functional area label branches, status label branches, and related label branches;

[0136] The functional area label branch: input coordinates, time and vehicle type characteristics, output grid function classification;

[0137] The status label branch takes speed, historical trajectory, and location features as inputs and outputs status labels.

[0138] The associated label branch: input coordinates and track status, output current track number.

[0139] Specifically, functional area labels, status labels, and associated labels are collected from the tracking database. Raw sensor data includes RFID trigger records, high-definition industrial camera trajectory coordinates, speed sensor data, and track encoder status. External data includes train scheduling plans and signal status. Data cleaning is performed, including removing duplicate records and filling in missing values. Features from location coordinates, timestamps, raw sensor data, and external data are extracted, including location features, time features, sensor features, vehicle type features, historical trajectory features, and external features. These features are then standardized. Semantic location labels include functional area labels, status labels, and associated labels. Functional area labels directly use grid functional labels generated from the digital twin model. Status labels are defined according to business rules, for example, 0 for stationary, 1 for moving, and 2 for speeding. Associated labels are track numbers. A neural network is constructed. The network model, with a shared preprocessing layer, includes an input layer that receives all input features. The input dimension is 11-dimensional, comprising [coordinate x, coordinate y, hour, minute, second, speed, acceleration, orientation angle, vehicle type code, trajectory entropy, and track state code]. A shared fully connected layer outputs a 128-dimensional hidden layer representation for extracting general features across tasks. The multi-task branch structure includes functional region label branches, state label branches, and association label branches. The loss function for the functional region label branches and association label branches is the cross-entropy loss function, while the loss function for the state label branch is the focus loss function. The optimizer is Adam with lr=0.001. The trained model is quantized and deployed to the station edge device. Input: Real-time reception of features such as coordinates, speed, and time from the vehicle tracking database. Output: Prediction results and confidence scores for the three main labels. The predicted labels and confidence scores are written to the tracking database to update the vehicle profile.

[0140] In one embodiment of the present invention, step S5 defines a joint optimization loss function, including the following steps:

[0141] The weights of the loss function are dynamically adjusted based on the prediction variance of each task, specifically as follows:

[0142]

[0143] in, The weight coefficients corresponding to the functional area labels. The variance of the functional area label prediction results. The variance of the state region label prediction results. The variance of the associated label prediction results;

[0144] Calculate the weight coefficients corresponding to the state region labels and associated labels using the same method;

[0145] The weight coefficients corresponding to the labels are weighted and added to the loss functions corresponding to the labels to construct a joint optimization loss function;

[0146] The joint optimization loss function is specifically as follows:

[0147]

[0148] in, To jointly optimize the loss function, , and These are the weight coefficients corresponding to the labels. , and These are the loss functions corresponding to the labels, where... For focus loss function, and Both are cross-entropy loss functions.

[0149] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for automatic identification and tracking and positioning of vehicles in a railway yard, characterized in that, The method comprises the following steps: S1: constructing a comprehensive monitoring system inside the railway station yard for collecting various types of data and performing synchronous fusion processing thereon; S2: setting up a three-level judgment mechanism, respectively calculating first, second and third matching degree scores, performing weighted addition to obtain a comprehensive matching degree score, and performing vehicle automatic identification on the comparison result of the comprehensive matching degree score and a comprehensive matching degree threshold value; S3: after successful vehicle automatic identification, constructing a tracking database, establishing a railway station yard coordinate system, and using a Kalman filtering algorithm to integrate multi-source data to obtain two-dimensional plane coordinates and height information of the vehicle in the railway station yard coordinate system; S4: using the RFID reader and high-definition industrial camera distributed inside the railway station yard to dynamically obtain vehicle coordinates and synchronize them to the tracking database, mapping the vehicle coordinates to the artificially labeled functional areas and associated track numbers through the railway station yard digital twin model, and generating complete semantic information in combination with the state label; S5: constructing a multi-task neural network framework through historical data and semantic location labels, defining a joint optimization loss function through dynamic weight adjustment and joint loss optimization; The three-level judgment mechanism in step S2 comprises the following steps: The data collected by the comprehensive monitoring system inside the railway station yard are preprocessed respectively; The three-level judgment mechanism is set up, specifically as follows: Vehicle visual features including body color, identification pattern, window number and head shape are extracted, matched with a pre-established vehicle feature database, and a first visual judgment is made according to the matching result; The vehicle contour, size and distance from the track measured by the laser sensor are matched with the standard vehicle parameter library, and a second physical judgment is made according to the matching result; The completeness and occlusion of the vehicle label information read by the RFID reader are judged as the third information judgment; The visual judgment, physical judgment and RFID information are weighted and fused to obtain a comprehensive matching degree score; The step S4 comprises the following steps: The RFID reader and high-definition industrial camera distributed inside the railway station yard are used to update the vehicle position coordinates in real time; The position coordinates are mapped into functional area labels, state labels and associated labels in real time through the data twin model and synchronized to the tracking database; The tracking positioning system updates the position coordinates of the vehicle in the tracking archive according to the received position, and after updating the vehicle coordinates, queries the railway station yard digital twin model in real time, maps the coordinates into functional area labels and associated labels, and synchronously stores them into the tracking database together with the state labels; The railway station yard digital twin model converts the railway station yard CAD drawing into a three-dimensional grid model, artificially labels the functional areas of each grid, and matches the preset labels according to the grid index where the vehicle coordinates are located; The step S5 of constructing a multi-task neural network framework through historical data and semantic location labels comprises the following steps: Collect vehicle data with position coordinates from the railway station yard historical operation data and artificially labeled semantic location labels, and the semantic location labels include functional area labels, state labels and associated labels; The neural network is used to extract position coordinates, timestamps, sensor raw data and external data features as input features, and the neural network automatically learns the correlation between tasks by sharing underlying parameters; The input features are processed through a shared preprocessing layer, including normalizing coordinate values, extracting time features and encoding classification features; A multi-task branch structure is used, including a functional area label branch, a state label branch and a correlation label branch; The functional area label branch inputs coordinates, time and vehicle features, and outputs grid function classification; The state label branch inputs speed, historical trajectory and position features, and outputs state labels; The correlation label branch inputs coordinates and trajectory state, and outputs the current trajectory number.

2. The method for automatic identification, tracking and positioning of vehicles in a railway yard according to claim 1, characterized in that, The railway station internal comprehensive monitoring system is constructed, including a high-definition industrial camera, a laser sensor, a radar, an RFID tag and reader and a Beidou positioning terminal.

3. The method for automatic identification, tracking and positioning of vehicles in a railway yard according to claim 1, characterized in that, The visual judgment, physical judgment and RFID information are weighted and fused to obtain a comprehensive matching degree score, including the following steps: The extracted vehicle visual features are matched with a preset vehicle feature database to calculate a first matching degree score; the physical parameters measured by the laser sensor are matched with a standard vehicle parameter library to calculate a second matching degree score; and the label information read by the RFID is used as a reference to calculate a third matching degree score according to the label completeness and the shielding degree; According to historical data, weights are assigned to each data source, and the weighted sum is added to obtain a comprehensive matching degree score, specifically: wherein, is a comprehensive matching score, , and are a first matching score, a second matching score and a third matching score, respectively, , and are corresponding weights, respectively. Based on the preset confidence, the confidence interval of the comprehensive matching degree score is determined through historical training data, and a comprehensive matching degree threshold is defined, specifically: wherein is a threshold value for the overall match degree, is the mean value of the historical overall match degree scores, is the standard deviation, is a confidence coefficient; If > the vehicle identification result is output; otherwise, it is determined that there is a matching conflict, and a matching inconsistency processing flow is triggered.

4. The method of claim 3, wherein, The third matching degree score is calculated according to the label completeness and the shielding degree, including the following steps: The RFID reader reads the vehicle label information, including the vehicle number and the vehicle type key field, and at the same time, the completeness and the shielding degree of the label are obtained; The completeness is determined by judging whether the label is complete and intact; and the shielding degree is determined by calculating the proportion of the label being shielded by an object; According to the label completeness and the shielding degree, the label state is divided into the following levels, and different weights are assigned according to the label quality level, specifically: First-class quality: completeness ≥ 95%, shielding degree ≤ 5%, corresponding weight: 1; Second-class quality: completeness ≥ 70%, shielding degree ≤ 30%, corresponding weight: 0.7; Third-class quality: completeness < 70%, shielding degree > 30%, corresponding weight: 0.4; According to the corresponding weight of the label quality level, the third matching degree score is obtained by multiplying 100.

5. The method of claim 3, wherein, The matching inconsistency processing procedure is triggered, including the following steps: S201: data review, including reviewing the data collected by the railway station internal comprehensive monitoring system, checking whether there is data collection error or omission, and starting the backup monitoring equipment to obtain vehicle information; S202: cross-validation, including matching the data after review and newly collected data again through the three-level judgment mechanism; S203: manual auxiliary judgment, including introducing manual auxiliary judgment if there is still matching inconsistency after data review and cross-validation, and finally identifying the vehicle.

6. The method of claim 1, wherein, The step S3 comprises the following steps: After the vehicle is successfully identified, a tracking database is constructed for the vehicle; The tracking database records the identification information and semantic location label of the vehicle; A railway station coordinate system is established, taking the starting point of the track as the origin and the track direction as the reference axis; Multi-sensor data is fused through Kalman filtering algorithm to calculate the two-dimensional plane coordinates and height information of the vehicle in the railway station coordinate system.

7. The method of claim 1, wherein, The joint optimization loss function defined in the step S5 comprises the following steps: The loss function weight is dynamically adjusted according to the prediction variance of each task, and the specific adjustment method is as follows: wherein, is a weight coefficient corresponding to the functional region label, is a variance of the functional region label prediction result, is a variance of the state region label prediction result, is a variance of the associated label prediction result; The same method is used to calculate the weight coefficients corresponding to the state area label and the associated label respectively; The label corresponding weight coefficients and the label corresponding loss function are weighted and added to construct the joint optimization loss function; The joint optimization loss function is specifically as follows: wherein, is a joint optimization loss function, , and are label corresponding weight coefficients, respectively, , and are label corresponding loss functions, respectively, wherein, is a focal loss function, and are cross-entropy loss functions.

Images