Method for extracting and identifying rail-mounted area in complex turnout area based on image identification

Abstract

The invention belongs to the technical field of smart urban rails and computer vision, and particularly provides an image recognition-based method for extracting and recognizing a rail-mounted area in a complex turnout area, and the method comprises the steps: processing an input image through an instance segmentation model to obtain an n-dimensional matrix mask; processing the matrix mask through a rail-mounted area processing module to obtain a matrix mask MaskRe; analyzing the number of rail-mounted areas of MaskRe, and intercepting a turnout area image through a turnout area image acquisition module; obtaining the positioning and reversing states of the turnout, namely the travel path direction of the train, through the target positioning model, so as to obtain the unique rail travel area of the train; and carrying out visualization processing, and outputting an RGB image result. According to the method, the scene image of the complex turnout area is processed through the instance segmentation model, the switch tip state is positioned and identified through the target positioning model, and the current running rail running area of the train in the complex turnout area image obtained at different scenes and shooting angles can be accurately identified and extracted.

Description

Technical Field

[0001] The present invention belongs to the field of automatic driving and computer vision technology, and specifically relates to a method for extracting and identifying track areas in a complex turnout area based on image recognition. Background Art

[0002] With the development of autonomous driving technology, the application of assisted driving technology in the field of rail transit is becoming more and more important. The core of assisted driving technology is to accurately identify the driving area of ​​the train, that is, the "track area". On the basis of identifying the track area, defining the vehicle spacing limit is an important step in the assisted driving technology in the field of rail transit. Considering the rapid development of the rail transit industry, the operating environment of national railways, subways, and trams has become more and more complex. There will be multiple rails intersecting in the switch area. Therefore, how to ensure that the train can identify the correct track area in the complex switch area and improve the safety of driving is an important problem to be solved.

[0003] At present, based on the track switch recognition technology, it is possible to evaluate the possible safety risks in advance. However, when faced with switch scenes with large changes in lighting differences and complex track crossings, the use of traditional track switch recognition technology has the problem of low recognition accuracy.

[0004] In this regard, in order to solve the problems existing in the prior art, technical personnel in this field have made many efforts. For example, Chinese patent application CN 111860137A proposes a vision-based track turnout recognition method. By using pure vision to perform track turnout recognition, color information and image distribution can be obtained, thereby improving the accuracy of turnout recognition. However, for turnout images obtained in different scenes, the recognition accuracy is still lacking. Summary of the invention

[0005] To solve the above problems, the present invention provides a method for extracting and identifying track areas in complex turnout areas based on image recognition. The scene image of the complex turnout area is processed by an instance segmentation model, and is located and identified by a target positioning model. It can accurately identify and extract the track area where the train is traveling in the turnout images obtained from different scenes and shooting angles.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is A method for extracting and identifying a track area in a complex turnout area based on image recognition, the method comprising the following steps: Step S1: constructing a track area extraction and recognition system in a complex turnout area based on image recognition; including an instance segmentation model, a target positioning model, a track area processing module and a visualization processing module; Step S2: Obtain the track image in front of the train, scale the image size to R*C, and input the image into the instance segmentation model for processing to obtain an n-dimensional matrix mask, where n is the number of track areas; Step S3: Analyze the number of track areas; if the number of track areas is equal to 1, input the mask matrix into the visualization module, and output the image result of unique track area identification; if the number of track areas is greater than 1, input the mask matrix into the track area processing module; Step S4: the track area processing module includes an invalid track area filtering module and a switch area image acquisition module; the mask matrix is ​​processed by the invalid track area filtering module to obtain a processed track area image matrix Mask_Re; Step S5: Analyze the number of track areas of Mask_Re. If the number of track areas is 1, input the Mask_Re matrix into the visualization module to output the image result of unique track area recognition. If the number of track areas is greater than 1, input Mask_Re into the switch area image acquisition module to capture the switch area image. Step S6: marking the turnout tips in the turnout branch area in the turnout area image as two types, open and close, as a turnout tip dataset; Step S7: using the turnout point data set as the training set and test set of the target positioning model, repeatedly training and testing to obtain the final model; Step S8: using the turnout area image as the input of the final model, positioning and identifying the turnout tip through the final model, obtaining the turnout direction of the current turnout branch area, and then obtaining the track area of ​​the train running in the turnout area; Step S9: Visualize the acquired track area through a visualization module and output an RGB image result.

[0007] Furthermore, in step S2, the size of the n-dimensional matrix mask is the same as the size of the input image, the area with a value of 0 in the matrix is ​​a non-track area, the area with a value other than 0 is a track area, and the value corresponding to each track area is different.

[0008] Furthermore, in step S4, the mask matrix is ​​processed by filtering the invalid track area module to obtain the processed track area image matrix Mask_Re in the following specific method: Step S4-1: Analyze the distribution of the mask matrix values ​​and calculate the center point coordinates (Rail_x, 10) of each track area when the ordinate is 10; Step S4-2: Calculate the difference between the horizontal coordinate of the center point of each track area when the vertical coordinate is 10 and the horizontal coordinate of the center point of the input image |Rail_x-R / 2|; obtain the track area Mask_N with the smallest distance from the center point of the image; wherein R / 2 is the horizontal coordinate of the center point of the input image; Step S4-3: extract the track area overlapping with Mask_N to obtain the processed track area image matrix Mask_Re.

[0009] Furthermore, in step S5, Mask_Re is input into the module for obtaining the image of the turnout area, and the specific method for intercepting the image of the turnout area is: Calculate all overlapping areas, select the overlapping area A with the largest number of track areas, calculate the minimum circumscribed rectangle coordinates, and intercept the turnout area image based on the obtained rectangle coordinates.

[0010] Furthermore, in step S8, the turnout tip is located and identified by the final model to obtain the turnout direction of the current turnout branch area, and then the specific method for obtaining the only track area in the turnout area is: Step S8-1: by locating and identifying the turnout tip, the center coordinates (Sx, Sy) of the turnout tip and the state of the turnout tip: open or closed are obtained; Step S8-2: Determine the direction of the turnout by combining the coordinates of the upper left corner point (Lx, Ly), the coordinates of the upper right corner point (Rx, Ry), the open and closed state of the turnout tip and the center coordinates of the turnout tip (Sx, Sy); Step S8-3: Select and extract the track area where the train runs from the overlapping area according to the direction of the turnout, and obtain the track area where the train runs in the turnout area.

[0011] Furthermore, in step S8-2, the specific method for determining the direction of the turnout is: when Sx is less than (Lx+Rx) / 2 and the state of the turnout tip is open, the turnout direction is left; when Sx is greater than (Lx+Rx) / 2 and the state of the turnout tip is open, the turnout direction is right.

[0012] Furthermore, in step S8-3, the specific method for selecting and extracting the track area for train operation from the overlapping area according to the opening direction of the turnout is as follows: when the turnout opens to the left, the track area on the left side of area A is searched within the minimum circumscribed rectangle of the overlapping area A and extracted; when the turnout opens to the right, the track area on the right side of A is searched within the minimum circumscribed rectangle of the overlapping area A and extracted.

[0013] The technical solution of the present invention has the following beneficial effects: The present invention processes an input image through an instance segmentation model to obtain an n-dimensional matrix representing a track area, and filters the matrix for invalid track area processing and obtains a turnout area image through a track area processing module to obtain an intercepted turnout area image, and then uses a target positioning model to determine the turnout direction in combination with the coordinate parameters of the upper left corner point, the coordinate parameters of the upper right corner point, the open and closed state of the turnout tip and the coordinate parameters of the center point of the turnout tip of the minimum circumscribed rectangle, and identifies and extracts the track area where the train is traveling in combination with the minimum circumscribed rectangle range of the overlapping area, thereby achieving accurate identification and extraction of the track area where the train is traveling in the turnout area in different scenarios, thereby improving the accuracy of the identification process; at the same time, it realizes pixel-level segmentation and instance distinction of each track area in the scenario where there are multiple track areas in the turnout area, thereby having better generalization and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of a method for extracting and identifying a track area in a complex turnout area based on image recognition according to the present invention; Figure 2 is a processing flow chart of the track area processing module of the present invention; Figure 3 is a diagram of a turnout region captured in an embodiment of the present invention; Figure 4 is a flow chart of segmentation processing according to an embodiment of the present invention; Figure 5 This is a diagram showing the effect of track area extraction and recognition according to an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0016] like Figure 1 and Figure 2 As shown, this embodiment proposes a method for extracting and identifying a track area in a complex turnout area based on image recognition, comprising the following steps: Step S1: construct a track area extraction and recognition system in a complex turnout area based on image recognition; including an instance segmentation model, a target positioning model, a track area processing module and a visualization processing module.

[0017] The DeepLab v3+ model used in the instance segmentation model is used to normalize the input image to match the training data distribution and obtain an n-dimensional matrix consisting of multiple track areas; Figure 4As shown in the figure, the structure design of the instance segmentation model includes an encoder and a decoder. The encoder part is based on the ResNet backbone network, which gradually extracts high-level semantic features through multiple residual blocks and dilated convolution layers. ResNet contains a 7x7 convolution layer, a maximum pooling layer and four residual blocks. The last two residual blocks use dilated convolution to expand the receptive field. At the end of the encoder, the spatial pyramid pooling (ASPP) module captures multi-scale information through parallel 1x1 convolution layers, three 3x3 dilated convolution layers with different dilation rates and a global average pooling layer. The output feature maps of all branches are reduced in dimension and concatenated together through a 1x1 convolution layer, and then further reduced in dimension through a 1x1 convolution layer. The decoder part upsamples the low-resolution feature map output by the ASPP module to the same size as the input image through bilinear interpolation, extracts low-level feature maps from the middle layer of the backbone network, and fuses them with the upsampled high-level feature maps by element-by-element addition to restore fine-grained spatial information, and finally outputs the category prediction of each pixel through a 1x1 convolution layer. Depthwise separable convolutions are widely used throughout the network to reduce computational costs, and batch normalization layers are added after each convolution layer to accelerate the training process and improve the generalization ability of the model.

[0018] The processing process is: Use convolutional neural networks (CNNs) to extract deep features of the image; these features will be used in subsequent region proposal networks (RPNs) and mask predictions; Slide a small convolution kernel over the feature map to generate a series of potential bounding box proposals and assign an objectness score to each proposal; Regress the proposed bounding boxes to obtain more accurate locations; Map the bounding box output by RPN to the feature map, and extract fixed-size feature patches through the region of interest (RoI) pooling operation; Each RoI pooled feature block is processed by two parallel fully connected layers: one for classification, that is, to determine whether it is an object; the other is used to generate a mask; Based on the output of the mask branch, a binary mask is generated for each proposed region. The generated mask is upsampled using a deconvolution layer to obtain finer mask edges. Masks with higher classification scores are subjected to non-maximum suppression (NMS) to remove overlapping masks and ensure that each object is segmented only once. The mask after NMS processing is output as the final segmentation result, and each mask corresponds to an instance in the image, that is, each track area.

[0019] The track area processing module is used to filter out invalid track areas from the obtained matrix and obtain turnout area images.

[0020] Among them, the target positioning model is used to locate and identify the switch tip in the acquired switch area image to obtain the center coordinates of the switch tip and the opening and closing status of the switch tip; and the opening direction of the switch is determined in combination with the coordinates of the upper left corner and the upper right corner of the minimum circumscribed rectangle in the acquired switch area image, so as to accurately extract the track area where the train is traveling.

[0021] Among them, the visualization module is used to visualize the track area and convert it into an RGB image.

[0022] Step S2: Get the track image in front of the train, scale the image size to R*C, and input the image into the instance segmentation model for processing to obtain an n-dimensional matrix mask, where n is the number of track areas. Specifically, the size of the n-dimensional matrix mask is the same as the size of the input image. The area with a value of 0 in the matrix is ​​a non-track area, and the area with a value other than 0 is a track area. The value corresponding to each track area is different.

[0023] For example, when there are three track areas, the mask matrix is ​​3-dimensional, the first dimension is Mask_1, the area with a value of 0 in Mask_1 is a non-track area, and the area with a value of 1 corresponds to track area instance 1. Take the 3-dimensional vector of the xth row and yth column in the mask matrix. If the number of non-zero values ​​in the vector is greater than or equal to 2, it means that there is an overlap in the track area. The pixel point with the horizontal coordinate x and the vertical coordinate y of the original input image is the overlap area of ​​the track area.

[0024] Step S3: Analyze the number of track areas; if the number of track areas is equal to 1, input the mask matrix into the visualization module and output the image result of unique track area identification; if the number of track areas is greater than 1, input the mask matrix into the track area processing module.

[0025] Step S4: The track area processing module includes an invalid track area filtering module and a switch area image acquisition module; the mask matrix is ​​processed by the invalid track area filtering module to obtain the processed track area image matrix Mask_Re. The specific method is: Step S4-1: Analyze the distribution of the mask matrix values ​​and calculate the center point coordinates (Rail_x, 10) of each track area when the ordinate is 10; Step S4-2: Calculate the difference between the horizontal coordinate of the center point of each track area when the vertical coordinate is 10 and the horizontal coordinate of the center point of the input image |Rail_x-R / 2|; obtain the track area Mask_N with the smallest distance from the center point of the image; wherein R / 2 is the horizontal coordinate of the center point of the input image; Step S4-3: extract the track area overlapping with Mask_N to obtain the processed track area image matrix Mask_Re.

[0026] Step S5: Analyze the number of track areas of Mask_Re. If the number of track areas is 1, input the Mask_Re matrix into the visualization module, perform visualization on the matrix, and output the image result of unique track area recognition; if the number of track areas is greater than 1, input Mask_Re into the switch area image acquisition module to capture the switch area image. The specific method is: like Figure 3 As shown, all overlapping areas are calculated, and the overlapping area A with the largest number of track areas is selected. For example, the n-dimensional vector of the x-th row and y-th column in the mask matrix is ​​taken, and different values ​​other than 0 are extracted to form a list. The set of pixels with the longest list length is the required overlapping area A, and then the minimum circumscribed rectangle coordinates are calculated, and the turnout area image is intercepted according to the obtained rectangular coordinates.

[0027] Specifically, the calculation formula for the minimum bounding rectangle coordinates is as follows. Let the point set be , the coordinates of the lower left and upper right corners of the minimum enclosing rectangle are: Lower left corner Top right corner Step S6: marking the turnout tips in the turnout branch area in the turnout area image as two types, open and close, as a turnout tip dataset.

[0028] Step S7: Use the turnout point dataset as the training set and test set of the target positioning model, and perform repeated training and testing to obtain the final model. Specifically, the target positioning model is also named the target detection model. The target detection model uses the yolov5 model, which divides the input image into S×S grids. Each grid unit is responsible for detecting the target whose center point falls within the unit. Each grid allows the prediction of 2 bounding boxes, which are understood as S×S×2 prediction areas, covering the entire area of ​​the image. Target detection is then performed in these prediction areas. The model uses Darknet53 as the backbone network, which consists of 53 convolutional layers. The network starts with a normal convolutional layer, followed by a batch normalization layer and a Leaky ReLU activation function, and then consists of multiple residual blocks. Each residual block contains multiple convolutional layers, which are optimized by applying the Cross Stage Partial (CSP) network strategy.

[0029] Step S8: Using the turnout area image as the input of the final model, the turnout tip is located and identified through the final model to obtain the turnout direction of the current turnout branch area, and then obtain the track area where the train runs in the turnout area. The specific method is: Step S8-1: by locating and identifying the turnout tip, the center coordinates (Sx, Sy) of the turnout tip and the state of the turnout tip: open or closed are obtained; Step S8-2: Determine the direction of the turnout based on the coordinates of the upper left corner point (Lx, Ly), the coordinates of the upper right corner point (Rx, Ry), the open and closed state of the turnout tip, and the center coordinates of the turnout tip (Sx, Sy); when Sx is less than (Lx+Rx) / 2 and the state of the turnout tip is open, the turnout direction is left; when Sx is greater than (Lx+Rx) / 2 and the state of the turnout tip is open, the turnout direction is right.

[0030] Step S8-3: Select and extract the track area for train operation from the overlapping area according to the direction of the turnout, and obtain the track area for train operation in the turnout area. When the turnout is left-opening, search for and extract the track area on the left side of area A within the minimum circumscribed rectangle of the overlapping area A; when the turnout is right-opening, search for and extract the track area on the right side of A within the minimum circumscribed rectangle of the overlapping area A.

[0031] Step S9: Visualize the acquired track area through a visualization module and output an RGB image result.

[0032] In order to verify the reliability and effectiveness of the method proposed in the present invention, the method proposed in the present invention is applied to a practical scenario, experimental verification is carried out, and the experimental results are analyzed.

[0033] In the verification experiment, a camera is used as the image input. A camera with a suitable focal length is installed in front of the train driver's cab, and the camera focal length is fine-tuned so that the picture can clearly display the track area. The software can obtain the current screen image in real time through the camera structure function, and the obtained image is used as the input image for experimental verification.

[0034] Through the method for extracting and identifying track areas in complex turnout areas based on image recognition described in the present invention, turnouts in different scenarios can be extracted and identified, so that the track area that the train needs to travel can be accurately identified, and pixel-level segmentation and instance distinction of each track area in the scenario where there are multiple track areas in the turnout area are achieved, which has better generalization and robustness. Figure 5 This is a diagram showing the effect of track area extraction and recognition according to an embodiment of the present invention, wherein (a) is the original image, (b) is the processing result of the instance segmentation model, (c) is the extraction result of the track area processing module, and (d) is the RGB image result output by the visualization processing.

[0035] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for extracting and identifying track areas in complex turnout areas based on image recognition, characterized in that: The following steps are involved: Step S1: constructing a track area extraction and recognition system in a complex turnout area based on image recognition; It includes an instance segmentation model, a target positioning model, a track area processing module and a visualization module; wherein the track area processing module includes an invalid track area filtering module and a switch area image acquisition module; Step S2: obtaining a track image in front of the train, scaling the image size to R*C, and inputting the image into the instance segmentation model for processing to obtain an n-dimensional matrix mask, wherein n is the number of track areas; Step S3: Analyze the number of the track areas; if the number of track areas is equal to 1, input the mask matrix into the visualization module, and output the image result of unique track area identification; if the number of track areas is greater than 1, input the mask matrix into the track area processing module; Step S4: Processing the mask matrix through the invalid track area filtering module in the track area processing module to obtain a processed track area image matrix Mask_Re; Step S5: Analyze the number of track areas of Mask_Re. If the number of track areas is 1, input the Mask_Re matrix into the visualization module to output the image result of unique track area recognition. If the number of track areas is greater than 1, input Mask_Re into the module for obtaining the image of the turnout area in the track area processing module to intercept the turnout area image. Step S6: marking the turnout tips in the turnout branch area in the turnout area image as two types, open and close, as a turnout tip dataset; Step S7: using the turnout point data set as a training set and a test set of a target positioning model, and repeatedly training and testing to obtain a final model; Step S8: using the turnout area image as the input of the final model, positioning and identifying the turnout tip through the final model, obtaining the turnout direction of the current turnout branch area, and then obtaining the track area of ​​the train running in the turnout area; Step S9: Visualize the acquired track area through the visualization module and output an RGB image result.

2. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 1 is characterized in that: In step S2, the size of the n-dimensional matrix mask is the same as the size of the input image, the area with a value of 0 in the matrix is ​​a non-track area, the area with a value other than 0 is a track area, and the value corresponding to each track area is different.

3. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 1, characterized in that: In step S2, the DeepLab v3+ model used by the instance segmentation model first uses a convolutional neural network to extract deep features of the image; slides a small convolution kernel on the feature map to generate a series of potential bounding box proposals, and assigns an objectness score to each proposal; regresses the proposed bounding box to obtain a more accurate position; maps the bounding box output by the RPN to the feature map, and extracts a fixed-size feature map block through the region of interest RoI pooling operation; processes each RoI-pooled feature map block through two parallel fully connected layers: one for classification, that is, judging whether it is an object; and the other for generating a mask; based on the output of the mask branch, a binary mask for each proposed region is generated; the generated mask is upsampled using a deconvolution layer to obtain a finer mask edge; non-maximum suppression (NMS) is performed on masks with higher classification scores to remove overlapping masks to ensure that each object is segmented only once; the mask processed by NMS is output as the final segmentation result, and each mask corresponds to an instance in the image; The model includes an encoder and a decoder; the encoder part is based on the ResNet backbone network, and gradually extracts high-level semantic features through multiple residual blocks and hole convolution layers. ResNet contains a 7x7 convolution layer, a maximum pooling layer and four residual blocks. The last two residual blocks use hole convolution to expand the receptive field; at the end of the encoder, the spatial pyramid pooling (ASPP) module captures multi-scale information through parallel 1x1 convolution layers, three 3x3 hole convolution layers with different hole rates and a global average pooling layer. The output feature maps of all branches are reduced in dimension and concatenated together through a 1x1 convolution layer, and then further reduced in dimension through a 1x1 convolution layer; the decoder part upsamples the low-resolution feature map output by the ASPP module to the same size as the input image through bilinear interpolation, and extracts low-level feature maps from the middle layer of the backbone network, which are fused with the upsampled high-level feature maps by element-by-element addition, and finally outputs the category prediction of each pixel through a 1x1 convolution layer.

4. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 1, characterized in that: In step S4, the mask matrix is ​​processed by filtering the invalid track area module to obtain the processed track area image matrix Mask_Re in the following specific method: Step S4-1: Analyze the distribution of the mask matrix values ​​and calculate the center point coordinates (Rail_x, 10) of each track area when the ordinate is 10; Step S4-2: Calculate the difference between the horizontal coordinate of the center point of each track area when the vertical coordinate is 10 and the horizontal coordinate of the center point of the input image |Rail_x-R / 2|; obtain the track area Mask_N with the smallest distance from the center point of the image; wherein R / 2 is the horizontal coordinate of the center point of the input image; Step S4-3: extract the track area overlapping with Mask_N to obtain the processed track area image matrix Mask_Re.

5. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 1, characterized in that: In step S5, Mask_Re is input into the module for obtaining the image of the turnout area. The specific method for capturing the image of the turnout area is as follows: Calculate all overlapping areas, select the overlapping area A with the largest number of track areas, calculate the minimum circumscribed rectangle coordinates, and intercept the turnout area image based on the obtained rectangle coordinates.

6. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 1, characterized in that: In step S7: The target localization model is a target detection model. The target detection model uses the yolov5 model, which divides the input image into S×S grids. Each grid unit is responsible for detecting the target whose center point falls within the unit. Each grid allows two bounding boxes to be predicted, which can be understood as S×S×2 prediction areas, covering the entire area of ​​the image. Then, target detection is performed in these prediction areas. The model uses Darknet53 as the backbone network, which consists of 53 convolutional layers. The network starts with a normal convolutional layer, followed by a batch normalization layer and a Leaky ReLU activation function, and then consists of multiple residual blocks. Each residual block contains multiple convolutional layers and is optimized by applying the Cross Stage Partial (CSP) network strategy.

7. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 1, characterized in that: In step S8, the specific method of locating and identifying the turnout point through the final model, obtaining the turnout direction of the current turnout branch area, and then obtaining the only track area in the turnout area is: Step S8-1: by locating and identifying the turnout tip, the center coordinates (Sx, Sy) of the turnout tip and the state of the turnout tip: open or closed are obtained; Step S8-2: Determine the direction of the turnout by combining the coordinates of the upper left corner point (Lx, Ly), the coordinates of the upper right corner point (Rx, Ry), the open and closed state of the turnout tip and the center coordinates of the turnout tip (Sx, Sy); Step S8-3: Select and extract the track area where the train runs from the overlapping area according to the direction of the turnout, and obtain the track area where the train runs in the turnout area.

8. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 7, characterized in that: In step S8-2, the specific method for determining the direction of the turnout is: when Sx is less than (Lx+Rx) / 2 and the state of the turnout tip is open, the turnout direction is left; when Sx is greater than (Lx+Rx) / 2 and the state of the turnout tip is open, the turnout direction is right.

9. The method for extracting and identifying the track area in a complex turnout area based on image recognition according to claim 7, characterized in that: In step S8-3, the specific method for selecting and extracting the track area for train operation from the overlapping area according to the direction of the turnout is as follows: when the turnout is opened to the left, the track area on the left side of area A is searched within the minimum circumscribed rectangle of the overlapping area A and extracted; when the turnout is opened to the right, the track area on the right side of A is searched within the minimum circumscribed rectangle of the overlapping area A and extracted.

Images