Urban rail water accumulation identification method and identification system

Abstract

The application discloses a kind of urban rail waterlogging identification method and identification system, the main steps of urban rail waterlogging identification method include collecting the ground water image data of each urban rail scene, establishing semantic segmentation model for waterlogging identification, neural network training and verification are carried out, the model is handled using the training completion to image, and the result information is output to interface service program, identification system includes image acquisition unit, model training unit, visual identification unit, linear line segment detection unit, ROI quantization processing unit, interface service program and intelligent alarm unit.The application adopts original visual detection technology and the method based on the slope search and length accumulation of multiple feature ROI regions, can tolerate camera shaking, light transformation and other bad conditions, can accurately identify and timely alarm to urban rail area shallow waterlogging, deep waterlogging, improve the flood situation investigation ability of relevant department and subsequent emergency disposal efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to a water accumulation recognition method and system, in particular to a rail transit water accumulation recognition method and system based on neural network visual detection technology. BACKGROUND

[0002] Urban rail transit is the backbone of public transportation system in large cities and an important infrastructure for modern city construction. It plays an increasingly important role in leading and supporting urban development, meeting people's travel needs, relieving traffic congestion, and reducing environmental pollution, and has become an important means of transportation for people's daily travel in large cities and an important guarantee for the normal operation of cities.

[0003] The operation safety of urban rail transit is of great significance to the safety of people's lives and property, the maintenance of social stability, and the improvement of people's sense of gain. In recent years, with the continuous increase of rail lines, rapid growth of operation mileage and scale, and continuous increase of passenger volume, the safety guarantee of urban rail transit is becoming more and more difficult, and higher requirements are put forward for operation safety management. However, in terms of flood control, it still relies on regular manual inspection by staff, and the inspection points are complex, including rail, tunnel, bridge and other areas. The inspection process not only consumes a lot of manpower, but also is prone to risks such as missed detection and delayed inspection, which poses a threat to the safe operation of urban rail transit. SUMMARY

[0004] The purpose of the present application is to provide a rail transit water accumulation recognition method based on neural network visual detection technology. Another purpose of the present application is to provide a system using the method.

[0005] Technical solution: The urban rail transit water accumulation recognition method comprises the following steps:

[0006] (1) Collecting ground water image data under each urban rail transit scene based on the target use environment, combining random ground water data on the network, using Labelme tool to label the water accumulation area, thereby forming a water accumulation recognition deep neural network data set, and dividing the water accumulation recognition neural network training set and neural network test set based on the data set;

[0007] (2) Establishing a semantic segmentation model for water accumulation recognition based on the standard FCN-8s full convolutional neural network;

[0008] (3) For the irregular characteristics of the shape and boundary of the water accumulation, and the variable characteristics of the water surface image information caused by the water surface reflection, a reflection-based attention mechanism unit is used to compare and identify the relevant image properties of the water surface reflection. Specifically, the reflection-based attention mechanism unit is added in step (2) to generate a new FCN-8s full convolutional neural network model, i.e. FCN-8s-FL model, and a loss function is further added in the new FCN-8s full convolutional neural network model to generate FCN-8s-FL-5RAU model;

[0009] (4) Based on the FCN-8s-FL model and FCN-8s-FL-5RAU model proposed in step (3) and the neural network training set completed in step (1), the neural network is trained;

[0010] (5) The model trained in step (4) is verified;

[0011] (6) The camera code stream data is obtained, and the pictures are extracted and framed in time to form time sequence frames G0(x, y), G1(x, y)……G n (x, y). The sequence frames G0(x, y), G1(x, y)……G n (x, y) are identified by the model trained in step (4), and the binary recognition result images G0(x, y)', G1(x, y)'……G n (x, y)' are output;

[0012] (7) The binary result images G0(x, y)', G1(x, y)'……G n (x, y)' are scanned and counted in pixels, the number of water accumulation pixel points is calculated, and the water accumulation conclusion is given according to the threshold value. The analysis and identification of the water pit and shallow water accumulation in each frame of picture are completed, and the identification result, position information and time information of the identification point are output to the interface service program;

[0013] (8) The images of the time sequence frames G0(x, y), G1(x, y)……G n (x, y) are filtered, light processed, enhanced and binary identified to obtain images H0(x, y), H1(x, y)……H n (x, y);

[0014] (9) The connected domain analysis is performed on the images H0(x, y), H1(x, y)……H n (x, y), and all linear rectangular support regions R0(x, y), R1(x, y)……R m(x, y), all connected domain contour R0(x, y), R1(x, y)……R m (x, y) of the pixel point, find the linear rectangular support area in which each pixel point is located, and preferentially extract it to the linear rectangular support area with a larger Length value, and update the area to R0(x, y)', R1(x, y)'……R m (x, y)';

[0015] (10) Update the area R0(x, y)', R1(x, y)'……R m (x, y), all black pixels in the rectangle are detected by neighborhood detection, and the isolated points without the same pixels in the neighborhood are removed, and all rectangular support areas with Width>α or Length<β are removed by re-traversing all rectangular support areas, to obtain R0(x, y)", R1(x, y)"……R m (x, y)", and then connect the rectangular wide edge midpoint line segments of all R0(x, y)", R1(x, y)"……R m (x, y)" to mark back to the original image G0(x, y), G1(x, y)……G n (x, y) in;

[0016] (11) Rotate the rails, sleepers, ballast, and turnout parts in all time sequence frames G0(x, y), G1(x, y)……G n (x, y) by linear line segment angle in reverse, realize horizontal positioning of key linear line segments, and after horizontal positioning, the rectangular frame can be used to extract the ROI of the local linear line segment area of the rails, sleepers, ballast, and turnout, so as to realize ROI extraction of different features, complete recognition and detection of deep accumulated water based on the quantization method of multi-feature ROI area slope search and length accumulation, and output the recognition result, position information, and time information of the recognition point to the interface service program;

[0017] (12) The interface service program matches the pit and shallow accumulated water recognition results output by steps (1)-(7) and the deep accumulated water results output by steps (8)-(11) according to the space-time relationship, outputs the recognition result, position information, and time information of the matched current recognition point, and stores them in the database, and sends them to the intelligent alarm unit, and pushes them through different alarm levels.

[0018] Further, the step (4) is based on the model construction method of the reflection-based attention mechanism unit: given an input feature map I with size [h, w, c], average pooling is performed in the horizontal direction to reduce it to [h, w / 2, c]; then vertical average pooling is performed to reduce it to X with size [n, w / 2, c]; then, each row X of X has size [1, w / 2], which is tiled or self-replicated to size [n, w / 2], and the feature map obtained from all rows is concatenated into a new feature map along the feature axis, with size [n, w / 2, c*n]; finally, this feature map is upsampled to size [h, w, c*n], and is denoted as X', I is concatenated along the feature axis n times to obtain I' with size [h, w, c*n], I' is subtracted from X' to obtain D with size [h, w, c*n], and the subtracted feature map is concatenated with I' again, fed into a convolutional layer and activated by a ReLU function to generate the final output feature map with the same size as I. i

[0019] Further, the step (3) based on the reflection-based attention mechanism unit identifies the relevant image properties of the water surface reflection, specifically including searching for the image reflected by the water surface, searching for the reflected pixel points by matching the image regions along the pixel columns of the image, and using multiple resolutions to search vertically.

[0020] Further, the step (5) verifies the trained model, specifically by applying the test set completed in step (1) to the FCN-8s-FL model and the FCN-8s-FL-5RAU model in step (3), and comparing them with the FCN-8s neural network model without using the reflection-based attention mechanism unit.

[0021] Further, the step (11) is based on the quantification method of the slope search and length accumulation of the multi-feature ROI region: the number and length accumulation of all line segments in the ROI region whose slope value difference with the line segments of the rail, sleeper, ballast, and turnout is less than k are calculated, and the number and length accumulation of the linear line segments of the rail, sleeper, ballast, and turnout in the ROI region are Counts-g, Counts-d, Counts-c1, Counts-c2, Lenth-g, Lenth-d, Lenth-c1, and Lenth-c2, respectively. When there is deep standing water, the rail, sleeper, ballast, and turnout are covered by the standing water, and the linear line segments in the corresponding ROI region disappear, resulting in a sharp decrease in the values of Counts-g, Counts-d, Counts-c1, Counts-c2, Lenth-g, Lenth-d, Lenth-c1, and Lenth-c2. When the above values are less than the threshold value for n consecutive frames, it is determined to be in a deep standing water state.​

[0022] The urban rail water accumulation identification system comprises an image acquisition unit, a model training unit, a visual identification unit, a linear line segment detection unit, an ROI quantization processing unit, an interface service program and an intelligent alarm unit, wherein the image acquisition unit is a camera conforming to a standard code stream output protocol; the model training unit is used for constructing an improved full convolution integral neural network, completing water accumulation identification model training and providing a basic model for visual identification; the visual identification unit performs image identification on water pits and shallow water accumulation under a monitoring area through the model trained by the model training unit; the linear line segment detection unit is internally provided with an original linear line segment identification algorithm, and realizes identification on linear line segment features of fixed scenes in urban rail scene monitoring images; the ROI quantization processing unit is used for processing images and extracting ROIs of local linear line segment areas of rails, sleepers, ballast beds and turnouts, and completes identification and detection on deep water accumulation through a quantization method based on multi-feature ROI area slope search and length accumulation; the interface service program is used for providing data flow service for each subsystem, and matches according to the space-time relationship of result data, and outputs matched information to the intelligent alarm unit; and the intelligent alarm unit is used for sending identification results to an urban rail emergency management center platform and pushing through different alarm levels.

[0023] Advantages: Compared with the prior art, the present application has the following remarkable advantages:

[0024] (1) The original visual detection technology is adopted to accurately identify and timely alarm shallow water accumulation and deep water accumulation in urban rail areas, and is suitable for all-weather scenes such as bridges and tunnels, and outdoor scenes, thereby improving the flood investigation ability of relevant departments and the subsequent emergency disposal work efficiency;

[0025] (2) The water surface mirror feature search is fused, and the full convolution neural network is improved, thereby strengthening the identification performance of the model, solving the identification difficulty caused by irregular shape and boundary of water accumulation, and improving the accuracy of identification on water pits and shallow water accumulation;

[0026] (3) The original quantization method based on multi-feature ROI area slope search and length accumulation is adopted to quickly distinguish the feature difference between different components such as rails, sleepers, ballast beds and turnouts and deep water accumulation, effectively identify the deep water accumulation state of water accumulation covering a large area of rails, and tolerate bad conditions such as camera shaking and light conversion, thereby having strong scene versatility;

[0027] (4) The regional shallow water accumulation and deep water accumulation are identified and result output through different algorithms, thereby providing strong data support for risk identification and rapid decision of a flood control emergency system, and helping relevant departments to quickly, efficiently and orderly complete emergency disposal work;

[0028] (5) The system adopts a standard protocol to obtain camera code stream data, can utilize the existing camera of the urban rail system to be connected, greatly reduces the construction cost and deployment difficulty, and is convenient for the relevant departments to quickly popularize and use. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The flowchart of the application;

[0030] Figure 2 The semantic segmentation model network structure diagram;

[0031] Figure 3 The structure diagram of the reflection attention mechanism unit;

[0032] Figure 4 The structure diagram of the new FCN-8s full convolutional neural network;

[0033] Figure 5 The semantic segmentation water accumulation identification binary effect diagram;

[0034] Figure 6 The picture preprocessing effect diagram;

[0035] Figure 7 The linear rectangular support area;

[0036] Figure 8 The linear line segment rotation and bit position process diagram;

[0037] Figure 9 The ROI extraction and quantization diagram;

[0038] Figure 10 The linear line segment detection result diagram of different water accumulations. DETAILED DESCRIPTION

[0039] The technical solutions of the application will be further described below in combination with the drawings.

[0040] The urban rail water accumulation identification system disclosed by the application comprises an image acquisition unit, a model training unit, a visual identification unit, a linear line segment detection unit, an ROI quantization processing unit, an interface service program and an intelligent alarm unit, as shown in the figure. Figure 1 The system adopts a standard protocol to obtain camera code stream data, can utilize the existing camera of the urban rail system to be connected, greatly reduces the construction cost and deployment difficulty, and is convenient for the relevant departments to quickly popularize and use.

[0041] The image acquisition unit is a camera conforming to a standard code stream output protocol, and is configured to acquire real-time image data. The camera can be a device that meets the requirements of monitoring a region in an original urban rail system. The model training unit is configured to construct an improved fully convolutional integral neural network, and to complete training of a waterlogging identification model, thereby providing a basic model for visual identification. The visual identification unit is configured to perform image identification on water pits and shallow waterlogging under a monitoring region by using the model trained by the model training unit, and to output the results to the interface service program. The linear line segment detection unit is configured to implement identification of linear line segment features of fixed scenery in a monitoring image of an urban rail scene, and to output identification results of all linear line segment features, by using a unique linear line segment identification algorithm. The ROI quantization processing unit is configured to process an image, and to extract ROIs of local linear line segment regions of a rail, a rail tie, a track bed, and a turnout, and to complete identification and detection of deep waterlogging by using a unique quantization method based on slope searching and length accumulation of multi-feature ROIs, and to output the results to the interface service program. The interface service program is configured to provide data flow conversion services for each subsystem, and to match results data according to a space-time relationship, and to output identification results, position information, and time information of a current identification point that has been matched to the intelligent alarm unit. The intelligent alarm unit is configured to send the identification results to an urban rail emergency management center platform, to push different alarm levels, and to provide a decision basis for a superior management system, and to help relevant departments to quickly, efficiently, and orderly complete emergency disposal work.

[0042] An urban rail waterlogging identification method includes the following steps:

[0043] (1) Collecting ground water image data under each urban rail scene based on a target use environment, combining random ground water data on a network, using a Labelme tool to label waterlogging regions, thereby constituting a waterlogging identification deep neural network data set, and dividing a waterlogging identification neural network training set and a neural network test set based on the data set.

[0044] (2) Establishing a semantic segmentation model for waterlogging identification based on a standard FCN-8s fully convolutional neural network. The network structure is as shown in Figure 2 The standard fully convolutional neural network is used, and the difference from a general convolutional neural network is that all the fully connected layers are replaced by convolutional layers to obtain a two-dimensional feature map, and then a softmax function is used to obtain the classification results of the corresponding pixels.

[0045] (3) In view of the irregular characteristics of the shape and boundary of waterlogging, and the variable characteristics of water surface image information caused by water surface reflection, a reflection-based attention mechanism unit is used to compare and identify the relevant image properties of water surface reflection. The reflection-based attention mechanism unit is added in step (2) to generate a new FCN-8s fully convolutional neural network (FCN-8s-FL model), and the network structure is as shown inFigure 4 As shown in the new FCN-8s full convolutional neural network model, the loss function (FCN-8s-FL-5RAU model) is added, and the focal loss is used to solve the problem of sample imbalance.

[0046] The reflection-based attention mechanism unit is used to search the upper image reflected by the water surface, and the reflection pixel points are searched by matching the image area along the pixel column of the image. In the vertical matching, multiple resolution traversal searches are used to improve the tolerance to errors caused by perspective distortion, slight camera rotation (angle with the horizon) and blurred reflection.

[0047] The structure diagram of the reflection-based attention mechanism unit is shown in Figure 3 As shown in the figure, specifically, given an input feature map I with size [h, w, c], average pooling is performed in the horizontal direction to reduce it to [h, w / 2, c], and vertical average pooling is performed to reduce it to X with size [n, w / 2, c]; then, each row X i of X has a size of [1, w / 2], which is tiled or self-replicated to a size of [n, w / 2], and the feature maps obtained from all rows are concatenated into a new feature map along the feature axis, with a size of [n, w / 2, c*n]; then, this feature map is upsampled to a size of [h, w, c*n], and represented as X', concatenated with I along the feature axis n times to obtain I' with a size of [h, w, c*n], and subtracted from X' to produce D with a size of [h, w, c*n], which encodes the reflection relationship. The feature map subtracted is concatenated with I' again, fed into a convolutional layer and activated by a ReLU function to generate a final output map with the same size as I.

[0048] (4) Based on the model proposed in step (3) and the training set completed in step (1), the neural network is trained, for example, on a 12GB memory NVIDIA TITAN XP GPU, with BatchSize set to 1 and learning rate set to 10 -6 times the initial value, and reduced by 0.2 times every 5,000 iterations.

[0049] (5) Verify the model trained in step (4) by applying the test set completed in step (1) to the new FCN-8s full convolutional neural network model in step (3), and compare it with the FCN-8s neural network model without using the reflection-based attention mechanism unit. It can be concluded that the model obtained in step (4) has better road water recognition function, and the relevant data is as follows:

[0050] Table 1 Comparison of model test results

[0051] F value Precision Recall Accuracy FCN-8s 65.21％ 69.81％ 61.18％ 99.05％ FCN-8s-FL 70.62％ 74.38％ 67.22％ 99.19％ FCN-8s-FL-5RAU 76.91％ 78.03％ 75.81％ 99.34％

[0052] (6) Obtain the camera code stream data and convert it into a video file, extract the image frames regularly to form a time series of image frames G0(x, y), G1(x, y)...G n (x, y), through the model trained in step (4), the sequence graph frames G0(x, y), G1(x, y)...G n (x, y) is identified and the output is as follows Figure 5 The binary recognition result graphs G0(x, y)′, G1(x, y)′...G n (x, y)′.

[0053] (7) For the binarized result graphs G0(x, y)′, G1(x, y)′, G n (x, y)′ performs pixel scanning statistics, calculates the number of waterlogged pixels, draws a conclusion on waterlogging based on the threshold, completes the analysis and identification of puddles and shallow waterlogging in each frame of the image, and outputs the identification results, location information, and time information of the identification point to the interface service program.

[0054] (8) For the time series frames G0(x, y), G1(x, y)...G n (x, y) is filtered, light processed, enhanced and binarized to obtain images H0(x, y), H1(x, y)...H n (x, y), such as Figure 6 shown.

[0055] (9) For images H0(x, y), H1(x, y) ... H n (x, y) performs connected domain analysis, and traverses all linear rectangular support regions R0(x, y), R1(x, y)...R m (x, y), such as Figure 7 As shown, for all connected domain contours R0(x, y), R1(x, y)...R m Traverse the pixel points of (x, y), find the linear rectangular support area where each pixel point is located, and preferentially extract it to the linear rectangular support area with a larger Length value and update the area to R0(x, y)′, R1(x, y)′...R m (x, y)′.

[0056] (10) Update the regions R0(x, y)′, R1(x, y)′, ... R m(x, y) ', all the black pixels in all the rectangles in (x, y)'are detected, the isolated points without the same pixels in the neighborhood are removed, and all the rectangular support regions are traversed again, the rectangular support regions with Width > a or Length < b are removed, to obtain R0(x, y) ", R1(x, y) ",... R m (x, y) ", and all R0(x, y) ", R1(x, y) ",... R m (x, y) " are connected, and the original image G0(x, y), G1(x, y),... G n (x, y) is marked, as shown in Figure 8 .

[0057] (11) The rails, sleepers, ballast, and turnout parts in all the time sequence frames G0(x, y), G1(x, y),... G n (x, y) are inversely rotated according to the linear line segment angle, to realize horizontal positioning of the key linear line segments, and after the horizontal positioning, the ROI (region of interest) of the local linear line segment region of the rails, sleepers, ballast, and turnout can be extracted through the rectangular frame, to realize ROI extraction of different features,

[0058] The number and length cumulative sum of all the line segments in the ROI region with a difference in slope value from the rails, sleepers, ballast, and turnout line segments less than k are calculated, as shown in Figure 9 The number and length cumulative sum of the ROI linear line segments of the rails, sleepers, ballast, and turnout are Counts-g, Counts-d, Counts-c1, Counts-c2, Lenth-g, Lenth-d, Lenth-c1, Lenth-c2, respectively, when there is deep accumulated water, the rails, sleepers, ballast, and turnout are covered by the accumulated water, the linear line segments in the corresponding ROI region disappear, and the values of Counts-g, Counts-d, Counts-c1, Counts-c2, Lenth-g, Lenth-d, Lenth-c1, Lenth-c2 decrease sharply, as shown in Figure 10 When it is determined that the above values are less than the threshold value for n consecutive frames, it is determined that there is a deep accumulated water state, and the recognition result, position information, and time information of the recognition point are output to the interface service program.

[0059] The algorithms in steps (8) to (11) are used for feature detection of a random accumulated water test picture, as shown in Figure 10 When there are light and shadow and ripples in the accumulated water, there is a small probability that a small amount of linear line segment features can be detected, however, Figure 9 and Figure 10It can be concluded that when the rail is covered by the depth of accumulated water, although the linear segment feature of the accumulated water area is completely different from the fixed slope linear segment feature generated by the rail, tie, ballast and turnout in the rail, and the stability of the linear segment feature is poor with water wave movement, but it can also be accurately distinguished through continuous identification of multiple frames of pictures. In the actual application process, the algorithm through the above steps can effectively identify the accumulated water of the urban rail.

[0060] (12) The interface service program matches the pit and shallow accumulated water recognition results recognized and output in steps (1) to (7) and the deep accumulated water results recognized and output in steps (8) to (11) according to the space-time relationship, outputs the recognition results, position information and time information of the current recognition point matched to the alarm unit, stores them into the database, simultaneously sends them to the intelligent alarm unit, and provides a decision basis for the flood control emergency system through the push of different alarm levels.

Claims

1. A method for identifying waterlogging in urban rail transit, characterized by: The method comprises the following steps: (1) Based on the target use environment, ground waterlogging image data of various urban rail scenarios were collected. Combined with random ground waterlogging data on the Internet, the Labelme tool was used to annotate the waterlogging areas to form a deep neural network dataset for waterlogging recognition. Based on this dataset, the neural network training set and neural network test set were divided; (2) Establish a semantic segmentation model for waterlogging recognition based on the standard FCN-8s fully convolutional neural network; (3) Add the reflection-based attention mechanism unit in step (2) to generate the FCN-8s-FL model, and further add the loss function to the FCN-8s-FL model to generate the FCN-8s-FL-5RAU model; (4) Performing neural network training based on the FCN-8s-FL model and FCN-8s-FL-5RAU model proposed in step (3) and the neural network training set completed in step (1); (5) Verify the model trained in step (4); (6) Obtain the camera code stream data, extract the frames of the pictures at regular intervals, and form the time series frames G0(x, y), G1(x, y) ... Gn(x, y). The model trained in step (4) is used to extract the time series frames G0(x, y), G1(x, y) ... Gn(x, y). n (x, y) is recognized and the binary recognition result graph G0(x, y)′, G1(x, y)′…G n (x, y)′; (7) For the binarized result graphs G0(x, y)′, G1(x, y)′…G n (x, y)′ performs pixel scanning statistics, calculates the number of waterlogged pixels, draws conclusions based on the threshold, completes the analysis and identification of puddles and shallow water in each frame, and outputs the identification results, location information, and time information of the identified points to the interface service program; (8) For the time series frames G0(x, y), G1(x, y)...G n The image of (x, y) is filtered, light processed, enhanced and binarized to obtain images H0(x, y), H1(x, y)...H n (x, y); (9) For images H0(x, y), H1(x, y)...H n (x, y) is used to perform connected domain analysis, and all linear rectangular support regions R0(x, y), R1(x, y)...R1(x, y) in the image that satisfy Width≤α and Length≥β are traversed according to the connected domain contour pixels. m (x, y), for all connected domain contours R0(x, y), R1(x, y)...R m Traverse the pixel points of (x, y), find the linear rectangular support area where each pixel point is located, and extract it to the linear rectangular support area with a larger Length value first and update the area to R0(x, y)′, R1(x, y)′...R m (x, y)′; (10) Update the regions R0(x, y)′, R1(x, y)′, ... R m Perform neighborhood detection on all black pixel pixels in the rectangle of (x, y)′, remove isolated points with no identical pixels in the neighborhood, and re-traverse all rectangular support areas, remove rectangular support areas with Width>α or Length<β, and obtain R0(x, y)″, R1(x, y)″...R m (x, y)″, then all R0(x, y)″, R1(x, y)″...R m (x, y)″ is connected by a line segment at the midpoint of the rectangle width, and marked back to the original image G0(x, y), G1(x, y)...G n (x, y); (11) The rails, sleepers, roadbeds, and switches in all time series frames G0(x, y), G1(x, y) ... Gn(x, y) are locally rotated inversely according to the linear segment angle to achieve horizontal positioning of the key linear segments. After the horizontal positioning, the local linear segment areas of the rails, sleepers, roadbeds, and switches are extracted using a rectangular frame, thereby achieving ROI extraction of different features. According to a quantization method based on multi-feature ROI area slope search and length accumulation, the deep water accumulation recognition and detection is completed, and the recognition result, position information, and time information of the recognition point are output to the interface service program; the quantization method based on multi-feature ROI area slope search and length accumulation is as follows: the number and length accumulation sum of all segments in the ROI area whose difference with the rail, sleeper, roadbed, and switch line segment slope value is less than k are calculated. When the number and length accumulation sum of the ROI linear segments of the rails, sleepers, roadbeds, and switches are less than a threshold for n consecutive frames, it is determined to be a deep water accumulation state; (12) The interface service program matches the puddle and shallow water identification results output by steps (1)-(7) and the deep water identification results output by steps (8)-(11) according to the spatiotemporal relationship, outputs the identification result, location information, and time information of the current identification point after matching, stores them in the database, and sends them to the intelligent alarm unit at the same time, and pushes them through different alarm levels.

2. The method for identifying waterlogging in urban rail transit according to claim 1, wherein: The step (3) identifies the relevant image properties of the water surface reflection based on the reflection attention mechanism unit, specifically including searching for the upper image reflected by the accumulated water surface, searching for reflected pixel points by matching the image area along the pixel columns of the image, and using multiple resolutions for traversal search in vertical matching.

3. The method for identifying waterlogging in urban rail transit according to claim 1, wherein: The model construction method of the reflection-based attention mechanism unit in step (3) is as follows: given an input feature map I of size [h, w, c], average pooling is performed in the horizontal direction to reduce it to [h, w / 2, c]; then vertical average pooling is performed to reduce it to X of size [n, w / 2, c]; then, each row X of X is averaged. i The size of the feature map is [1, w / 2], which is flattened or copied to the size of [n, w / 2]. The feature maps obtained from all rows are concatenated along the feature axis into a new feature map of size [n, w / 2, c*n]; finally, this feature map is upsampled to [h, w, c*n] and represented as X'. It is concatenated n times along the feature axis to obtain I' of size [h, w, c*n]. I' is subtracted from X' to produce D of size [h, w, c*n]. The subtracted feature map is concatenated with I' again, fed into the convolutional layer and activated by the ReLU function to generate the final output feature map of the same size as I.

4. The method for identifying waterlogging in urban rail transit according to claim 1, wherein: The specific operation of verifying the trained model in step (5) is as follows: applying the test set completed in step (1) to the FCN-8s-FL model and the FCN-8s-FL-5RAU model in step (3), and comparing them with the FCN-8s neural network model that does not use the reflection-based attention mechanism unit.

5. The method for identifying waterlogging in urban rail transit according to claim 1, wherein: The quantification method based on multi-feature ROI area slope search and length accumulation in step (11) is as follows: calculate the number and length accumulation of all line segments in the ROI area whose slope value difference with the rail, sleeper, roadbed, and switch line segment is less than k, and the number and length accumulation of the ROI linear segments of the rail, sleeper, roadbed, and switch are Counts-g, Counts-d, Counts-c1, Counts-c2, Lenth-g, Lenth-d, Lenth-c1, and Lenth-c2 respectively. When deep water accumulation occurs, the rail, sleeper, roadbed, and switch are covered by water, and the linear segments in the corresponding ROI area disappear, and the values ​​of Counts-g, Counts-d, Counts-c1, Counts-c2, Lenth-g, Lenth-d, Lenth-c1, and Lenth-c2 drop sharply. When the above values ​​are less than the threshold for n consecutive frames, it is determined to be a deep water accumulation state.

6. A system for identifying water accumulation on urban rail, using the method for identifying water accumulation on urban rail according to any one of claims 1 to 5, characterized in that: It includes image acquisition unit, model training unit, visual recognition unit, linear segment detection unit, ROI quantization processing unit, interface service program and intelligent alarm unit. Among them, the image acquisition unit is a camera that complies with the standard code stream output protocol; the model training unit is used to build the FCN-8s-FL-5RAU model, complete the water accumulation recognition model training, and provide a basic model for visual recognition; the visual recognition unit uses the model trained by the model training unit to perform image recognition of puddles and shallow water in the monitoring area; the linear segment detection unit has an original linear segment recognition algorithm built in to realize urban rail scene monitoring The linear line segment features of fixed scenes in the measured image are identified; the ROI quantization processing unit is used to process the image and extract the ROI of the local linear line segment area of ​​the rail, sleeper, roadbed and switch, and complete the identification and detection of deep water accumulation through a quantification method based on multi-feature ROI area slope search and length accumulation; the interface service program is used to provide data flow services for each subsystem, and match the result data according to the spatiotemporal relationship, and output the matched information to the intelligent alarm unit; the intelligent alarm unit is used to send the recognition results to the urban rail emergency management center platform and push them through different alarm levels.

Images