A TDCS information intelligent recognition method based on multi-dimensional fusion analysis of video signals

Abstract

The present invention discloses a method for intelligently identifying TDCS interface information based on multi-dimensional fusion analysis of video signals, comprising a video signal acquisition and decoding unit, a video signal multi-dimensional fusion analysis unit, and a business data encapsulation and distribution unit. First, the network video stream protocol of the TDCS driving log interface is parsed and encapsulated in real time through the video decoding and image enhancement algorithms of the video signal acquisition and decoding unit, and the processed video frame data is extracted. Then, based on the multi-dimensional fusion processing technology of the video signal, key information is identified from the video frame data according to text information, color information, change state, and prior structure. Finally, the extracted key information is encapsulated into standard business data in combination with business logic, and encrypted for message sending. This method integrates video decoding technology, OCR recognition technology, and image feature recognition technology, and has the characteristics of high data acquisition security and high information extraction accuracy, providing technical support for the safe acquisition of dispatching command and management system information.

Description

Technical Field

[0001] The present invention relates to the field of railway dispatching command and management system (TDCS) data acquisition. Specifically, a method for acquiring TDCS driving log interface data based on multi-dimensional analysis of video signals is proposed, thereby achieving the acquisition of train dispatching and receiving scheduling information through physical isolation. Background Art

[0002] TDCS (Train Operation Dispatching Command System) is a railway-wide dispatching and command management system. It provides timely and accurate modern dispatching and command management tools to all levels of railway management personnel, playing a vital role in the railway business domain. Furthermore, the development of relevant business domain application systems based on the train dispatching and receiving information provided by TDCS can further improve the efficiency of railway operations and management and the implementation of standardized operations.

[0003] However, due to the early development and implementation of the TDCS system, the system suffers from incomplete data sharing interfaces and imperfect data encryption measures. Furthermore, with each version iteration, the TDCS system versions deployed by various railway bureaus are inconsistent. These issues make obtaining data from the TDCS system through a unified interface very inconvenient and pose security risks.

[0004] In recent years, with the development of deep learning-based artificial intelligence technology and the continuous maturity of the combination of cutting-edge technologies and industry applications, the application of deep learning computer vision technology to video signals containing the TDCS driving log interface, through the combination of image processing technology and OCR recognition technology, is expected to provide another new approach to solving the above problems. Summary of the Invention

[0005] To address the above issues, the present invention proposes a TDCS information intelligent recognition method based on multi-dimensional fusion analysis of video signals. This method analyzes the dispatching and receiving information in real time from the video stream encoding the TDCS driving log interface. The method has the advantages of high security, strong real-time performance, and high accuracy. The method comprises the following steps:

[0006] (1) The signal acquisition and decoding unit uses video hard decoding technology to pull the network video stream protocol in real time to obtain the decoded video data frame I. The data structure is a tensor of h×w×c, and the pixel value range is [0,255].

[0007] (2) The signal acquisition and decoding unit performs digital filtering and contrast enhancement on the video data frame I through image data enhancement technology, removes noise data generated during the signal conversion process, improves the signal-to-noise ratio and clarity of the video data frame, and obtains the processed data frame I′;

[0008] (3) The video signal multi-dimensional fusion analysis unit extracts the information contained in the travel log interface, such as the arrival train number, arrival receiving track, agreed departure time at the rear station, actual departure time at the rear station, scheduled train arrival time at the rear station, actual arrival time of the train at the rear station, departure train number at this station, departure track at this station, departure time agreed by the front station, actual departure time of the train at this station, dispatching command, system time, etc., through the designed text recognition TDCSOCR algorithm. Preferably, the TDCSOCR algorithm architecture can be described in detail as including a travel log interface text area detection module, a travel log interface text recognition module, and a natural language post-processing module. Specifically: First, the text area detection module analyzes the real-time video stream and outputs the location area to be identified. Then, the text recognition module inputs the location area image to be identified and outputs the identified text information. Finally, the natural language post-processing module post-processes the identified text information to improve the accuracy of text recognition. The contents of each sub-module are as follows: The text area detection module of the driving log interface table includes two modes: global text detection and change text detection. The global text detection mode is started in the initialization stage and when the recognition information needs to be fully updated. The change text detection mode is used at other times. Global text detection is extracted by the intersection method of vertical and horizontal lines, and change text detection is extracted by the inter-frame difference method. Both detection modes output the rectangular position [x, y, w, h] of the text block to be identified. The text recognition module of the driving log interface table uses a text recognition algorithm based on deep learning. The algorithm input is a 32×160 grayscale image. The backbone network consists of 4 layers of residual convolution structure with BatchNorm, 1 layer of pooling layer, and 2 layers of multi-head Transformer structure. The loss function uses Attention Loss and CTC Loss, using the driving log text character training model, the natural language post-processing module builds a priori knowledge base based on the known train information subset, the preceding station name, and the following station name processed by the station history, and uses regular matching to correct the accuracy of text recognition;

[0009] (4) The multi-dimensional fusion analysis unit of the video signal identifies the color features of the background fill area of ​​the table cells, the table border lines, and the table text in the driving log interface. The trains above the green line are the trains that have been processed, the trains between the red and green lines are the trains that will be processed soon, the blue fill area of ​​the table cells indicates the trains that are being processed, the trains marked with pink text in the table are passenger trains, and the trains marked with light green text are freight trains;

[0010] (5) The video signal multi-dimensional fusion analysis unit monitors the areas where information changes in the driving log interface through the video frame difference algorithm, identifies the areas where changes occur, and does not process the areas that have not changed, thereby improving the efficiency of information extraction;

[0011] (6) The video signal multi-dimensional fusion analysis unit uses the designed driving log interface table structure reconstruction algorithm to arrange the recognized text information according to the table structure and match the layout of the actual key information in the table, further correcting the accuracy of the algorithm recognition; specifically:

[0012] (6-1) Find all vertical and horizontal lines according to Hough transform line detection, and project them horizontally and vertically.

[0013] Merge lines;

[0014] (6-2) Solve the intersection points of horizontal and vertical lines in sequence to reconstruct the row and column distribution of the table;

[0015] (6-3) Match the row and column distribution of the reconstructed table with the number of rows and columns of the prior table to determine whether the reconstructed table has

[0016] omission;

[0017] (6-4) If there are any omissions, ignore the reconstruction results and repeat steps (6-1) to (6-3).

[0018] (7) The business data encapsulation and distribution unit combines the text information and color feature information identified by the video signal multi-dimensional fusion analysis unit according to the business logic to form a train receiving and sending signal that meets the needs of the business application, encrypts the data through an encryption algorithm, and then completes the signal distribution through a service request.

[0019] Beneficial effects

[0020] The present invention discloses a method for intelligently recognizing TDCS information based on multi-dimensional fusion analysis of video signals. Deep learning text recognition technology is used to construct a TDCSOCR algorithm for recognizing text information on the TDCS driving log interface, and an image recognition algorithm is used to extract color information, change information, and structural information. This method has high operating efficiency and high recognition accuracy, can be deployed on both the server and edge ends, and provides technical support for securely obtaining information from the dispatching, command, and management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a module structure diagram of a TDCS information intelligent recognition method for multi-dimensional fusion analysis of video signals of the present invention;

[0022] Figure 2This is the interface diagram of the TDCS driving log decoded by the present invention;

[0023] Figure 3 This is a structural diagram of the TDCSOCR deep learning algorithm of the present invention;

[0024] Figure 4 This is a diagram showing the text detection and text recognition effects on the driving log interface of a specific example of the present invention; DETAILED DESCRIPTION

[0025] The following, in conjunction with the accompanying drawings and specific examples, illustrates the implementation effects of the TDCS information intelligent recognition method for multi-dimensional fusion analysis of video signals proposed by the present invention through a specific operational process. It should be understood that these embodiments are merely illustrative of the present invention and are not intended to limit the scope of the present invention. After reading this document, modifications of various equivalent forms of the present invention made by those skilled in the art fall within the scope defined by the appended claims.

[0026] Real scene data of the TDCS3.0 driving log interface of a certain station on May 20, 2023 was selected to verify the effectiveness of the TDCS information intelligent recognition method based on multi-dimensional fusion analysis of video signals proposed in this paper.

[0027] The overall system flow chart of the present invention is as follows: Figure 1 The specific implementation steps are as follows:

[0028] (1) The signal acquisition and decoding unit uses the video hard decoding technology to pull the network video stream protocol in real time to obtain the decoded video data frame I. The data structure is a tensor of 1920×1080×3, and the pixel value range is [0,255]. Figure 2 ;

[0029] (2) The signal acquisition and decoding unit performs digital filtering and contrast enhancement on the video data frame I based on the Gaussian filtering and contrast enhancement algorithms of the OPENCV image algorithm library, removes the noise data generated during the signal conversion process, improves the signal-to-noise ratio and clarity of the video data frame, and obtains the processed data frame I′;

[0030] (3) The initial data frame I′ is input into the text region detection module, and the global text detection mode is used to extract the vertical and horizontal lines by detecting the intersection of the lines. The rectangular positions of all text blocks are identified, such as [130, 310, 105, 40]. Then, the extracted text region image is scaled and grayed out and input into the trained deep learning TDCSOCR algorithm. The algorithm structure includes a grayscale image with a size of 32×160 as input. The backbone network consists of a 4-layer residual convolution structure with BatchNorm, a 1-layer pooling layer, and a 2-layer multi-head Transformer structure. The network structure is shown in the figure. Figure 3As shown, the text effect diagram recognized by the TDCSOCR algorithm is as follows Figure 4 As shown, finally, according to the known train number and business domain range of a certain station, the natural language post-processing module performs post-processing on the recognized text information to improve the text recognition accuracy;

[0031] (4) Use the color feature matching algorithm to identify the color features of the background fill area of ​​the I′ table cell, the table border lines, and the table text. The train numbers that have been processed include "21024", "82797", "82979", and "X238". The train number being processed is "K8251". At the same time, the time information of each train can be extracted;

[0032] (5) For the changed video text information, the video frame difference algorithm is used to monitor the area where the information changes in the driving log interface, and the changed text information is identified by the TDCSOCR algorithm and color feature matching algorithm;

[0033] (6) All vertical and horizontal lines are found through Hough transform line detection. By solving the intersection of vertical and horizontal lines, and based on the known header positions, such as "train number", "receiving track", "agreement to depart from the adjacent station", etc., the travel log interface table is reconstructed. The reconstructed table structure position information is used to correct the meaning of the text information recognized in step 5.

[0034] (7) Finally, the extracted train number, time and other information are combined to form a train receiving and dispatching signal that meets the needs of business applications, such as "This station agrees to the departure of K8251 at 16:36 on May 20, 2023" and "K8251 departs from this station at 17:02 on May 20, 2023". The data is encrypted using an encryption algorithm and then the signal is distributed through a service request.

[0035] The above analysis shows that the TDCS information intelligent recognition method based on multi-dimensional fusion analysis of video signals proposed in this invention can identify key information about receiving and dispatching trains from collected TDCS video stream signals. This method has high operating efficiency and recognition accuracy, can be deployed on both the server and edge sides, and provides technical support for securely obtaining information from the dispatching command and management system.

Claims

1. A method for intelligent recognition of TDCS interface information based on multi-dimensional fusion analysis of video signals, characterized in that The identification method obtains the following information: (1) The signal acquisition and decoding unit uses the video hard decoding technology to pull the network video stream protocol in real time to obtain the decoded TDCS driving log interface video data frame I. The data structure is a tensor of h×w×c, where h represents the height of the video frame, w represents the width of the video frame, and c represents the channel of the video frame. The pixel value range is [0, 255]. (2) The signal acquisition and decoding unit performs digital filtering and contrast enhancement on the video data frame I through image data enhancement technology, removes the noise data generated during the signal conversion process, improves the signal-to-noise ratio and clarity of the video data frame, and obtains the data frame I′ with enhanced TDCS driving log interface signal. I′ is used as the input of the algorithm to extract TDCS driving log interface information; (3) The video signal multi-dimensional fusion analysis unit extracts the arrival train number, arrival track, agreed departure time of the next station, actual departure time of the next station, scheduled train arrival time of the next station, actual arrival time of the train at the next station, departure train number of this station, departure track of this station, departure time agreed by the front station, actual departure time of the train at this station, dispatching command, and system time information contained in the TDCS driving log interface through the text recognition TDCSOCR algorithm; (4) The video signal multi-dimensional fusion analysis unit identifies the color features of the background fill area, table border lines, and table text in the TDCS driving log interface. Trains above the green line are processed trains, trains between the red and green lines are trains to be processed, and blue fill areas in the table cells indicate trains that are being processed. Trains marked with pink text in the table are passenger trains, and trains marked with light green text are freight trains. (5) The video signal multi-dimensional fusion analysis unit identifies the area where information changes in the driving log interface through the following formula: g=Thr(abs(I c ′-I′ p ),θ) Where I c ′ represents the current frame, I′ p Represents the previous frame, abs represents the absolute value, θ represents the threshold of the threshold function Thr, and the pixel point whose pixel value changes by more than θ has a value of 1 in g. The area where the change occurs is identified by counting the blocks with a value of 1 in g. No processing is performed on areas that have not changed, thus improving the efficiency of information extraction; (6) The video signal multi-dimensional fusion analysis unit uses the driving log interface table structure reconstruction algorithm to arrange the recognized text information according to the table structure and match the layout of the actual key information in the table, further correcting the accuracy of the algorithm recognition; (7) The business data encapsulation and distribution unit combines the text information and color feature information identified by the video signal multi-dimensional fusion analysis unit according to the business logic to form a train receiving and sending signal that meets the needs of the business application, encrypts the data through an encryption algorithm, and then completes the signal distribution through a service request.

2. The method according to claim 1, characterized in that The text recognition TDCSOCR algorithm includes a driving log interface text area detection module, a driving log interface text recognition module, and a natural language post-processing module. The text area detection module analyzes the real-time video stream and outputs the location area to be identified. The text recognition module inputs the location area image to be identified and outputs the recognized text information. The natural language post-processing module post-processes the recognized text information to improve the text recognition accuracy.

3. The method according to claim 2, characterized in that The text area detection module in the driving log interface table includes two modes: global text detection and changing text detection. The global text detection mode is started during the initialization phase and when the recognition information needs to be fully updated. The changing text detection mode is used at other times. Global text detection is extracted by detecting the intersection of vertical and horizontal lines, and changing text detection is extracted by the inter-frame difference method. Both detection modes output the rectangular position [x, y, w, h] of the text block to be identified, where x represents the horizontal coordinate of the upper left corner of the rectangle, y represents the vertical coordinate of the upper left corner of the rectangle, w represents the width of the rectangular area, and h represents the height of the rectangular area.

4. The method according to claim 2, characterized in that The text recognition module for the table in the driving log interface uses a deep learning-based text recognition algorithm. The algorithm input is the output of the text region detection module. The backbone network consists of a 4-layer residual convolution structure with BatchNorm, a 1-layer pooling layer, and a 2-layer multi-head Transformer structure. The loss function uses Attention Loss and CTC Loss, and the model is trained using driving log text characters. The algorithm outputs the recognized text information.

5. The method according to claim 2, characterized in that The natural language post-processing module establishes a priori knowledge base based on the subset of known train information, the name of the preceding station, and the name of the following station processed historically, and uses regular matching to correct the accuracy of text recognition.

6. The method according to claim 1, wherein Driving log interface table structure reconstruction algorithm: (6-1) Find all vertical and horizontal lines using Hough transform line detection, and merge the lines by horizontal and vertical projection; (6-2) Solve the intersection points of horizontal and vertical lines in sequence to reconstruct the row and column distribution of the table; (6-3) Match the row and column distribution of the reconstructed table with the number of rows and columns of the prior table to determine whether there are omissions in the reconstructed table; (6-4) If there are omissions, ignore the reconstruction result and repeat steps (6-1) to (6-3); if there are no omissions, the reconstruction is completed.

Images