Railway external environment hidden danger intelligent identification method based on unmanned aerial vehicle inspection video

Abstract

The invention discloses a railway external environment hidden danger intelligent identification method based on an unmanned aerial vehicle inspection video, and the method comprises the steps: carrying out the preprocessing of data shot by an unmanned aerial vehicle, and forming a sample data set; training a target detection network by using the sample data set to obtain an optimal weight parameter; judging a target detection model applicable to a current scene according to an environment image acquired by the unmanned aerial vehicle in real time, and acquiring an identification frame of a risk target by using the trained target detection network; the identification frame comprises pixel coordinates of the risk target and category information of the risk target; constructing a rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system; and according to the identified pixel coordinates of the risk target, obtaining the Gaussian plane coordinates of the position of the unmanned aerial vehicle, and calculating the geographic coordinates of the position of the risk target by using the constructed rotation matrix. According to the invention, a set of efficient and intelligent troubleshooting and treatment system scheme for railway external environment hidden dangers is constructed, and the accuracy of hidden danger identification and the precision of spatial positioning are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle photography, and in particular to a method for intelligently identifying hidden dangers in the external environment of railways based on unmanned aerial vehicle inspection videos. Background Art

[0002] With the continuous growth of my country's railway operating mileage and the continuous complexity of the railway transportation network, the hidden dangers of the external environment along the railway are becoming increasingly prominent. The traditional railway external environment hidden danger inspection method mainly adopts human foot inspection, visual identification, etc., which has problems such as low operating efficiency and low identification rate, and cannot meet the needs of comprehensive, accurate and systematic analysis and judgment of potential risks in the external environment. With the rise of drone technology, the use of drone video for hidden danger inspection has gradually been applied to the railway external environment inspection work. However, most of the existing drone inspection technologies remain in the data collection and transmission stage, and the analysis and identification efficiency of image data is low, and the generation and systematic management of reports still mostly rely on manual execution. Therefore, in order to further improve the efficiency and accuracy of the railway external environment hidden danger inspection, the present invention uses artificial intelligence algorithms and photogrammetry-related principles based on drone video image data to achieve efficient extraction and storage of railway external environment hidden dangers, and at the same time can realize intelligent positioning of hidden danger targets and generate reports with one click, assisting relevant personnel in the management of hidden dangers. Based on the results of this algorithm, a demonstration application of railway external environment inspection and management has been carried out. Summary of the invention

[0003] Therefore, the purpose of the present invention is to provide a method for intelligently identifying hidden dangers in the external environment of railways based on drone inspection videos, to construct a set of efficient and intelligent inspection and management system solutions for hidden dangers in the external environment of railways, and to improve the accuracy of hidden danger identification and the precision of spatial positioning.

[0004] In order to achieve the above-mentioned object, the present invention provides an intelligent identification method of railway external environment hidden dangers based on drone inspection video, which is characterized by comprising: S1, pre-process the data captured by the drone to form a sample data set; S2. Use the sample data set to train the target detection network to obtain the optimal weight parameters; S3, using the environmental image collected by the drone in real time, determining the target detection model applicable to the current scene, and using the trained target detection network to obtain the identification frame of the risk target; the identification frame includes the pixel coordinates of the risk target and its category information; S4, constructing a rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system; S5. According to the pixel coordinates of the identified risk target, the Gaussian plane coordinates of the location of the UAV are obtained, and the geographic coordinates of the location of the risk target are calculated using the constructed rotation matrix.

[0005] Further preferably, in S1, preprocessing the data captured by the drone further includes: S101, classifying and extracting key frames from the drone's shooting data to form long-focus samples and short-focus samples; S102, annotating part of the data in the long-focus sample and the short-focus sample to obtain an annotated data set; S103, enhancing all sample data; S104: Divide the enhanced images and the corresponding label files into a training set and a validation set according to a specific ratio.

[0006] Further preferably, in S103, the enhancing of the sample data includes: enhancing the size of the sample data, enhancing the pixels of the sample data, and synchronously enhancing the sample image and the corresponding annotation file.

[0007] Further preferably, the sample image and the corresponding annotation file are enhanced synchronously, and the XML annotation file of the original image is read to obtain the original bounding box information; Enhance the original image and calculate the new coordinates of the bounding box in the enhanced image; Update the bounding box information in the XML markup file, including the file name, the upper left corner and lower right corner coordinates of the bounding box; The updated file names contain the enhanced sequence number to distinguish different enhanced images.

[0008] Use python code to save the updated XML markup file.

[0009] Further preferably, in S3, the environmental image collected in real time by the drone is used to obtain the identification frame of the risk target using the trained target detection network; the following steps are also included: S301, obtaining the drone video address and POS file transmitted in the previous section, and extracting the POS information of the 10th frame; S302, judging the target detection model applicable to the current scene according to the camera focal length and flight altitude at this time; S303, using Tornado's asynchronous mechanism, transmit the information of the selected target detection model to the front end in real time; S304: Infer each frame of the image according to the selected target detection model, and return JSON format data including the second of the frame image, the recognized type, the annotated image, the original image, and the latitude and longitude information of the image target.

[0010] Further preferably, in S302, determining the target detection model applicable to the current scene comprises the following steps: The flight height of the drone to obtain the current frame image is H, the focal length of the camera is f, and the pixel size is recorded as p. The spatial resolution (GSD) of the drone is calculated as follows: Image GSD attribute of the minimum object size marked in the preset attributes of each target detection model; The calculated GSD is compared with the GSD value attribute in each target detection model, and the model with the closest value is selected as the optimal model for identifying the incoming data.

[0011] Further, preferably, in S4, the constructed rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system is expressed by the following formula: Among them, lon and lat represent the longitude and latitude coordinates of the drone's location when the current frame image is taken; h, p, and r represent the yaw angle, pitch angle, and roll angle of the drone at this time; L represents the central meridian of the Mercator projection graduation zone where the drone's location is located, and B represents the longitude of the central meridian at this time.

[0012] Further preferably, in S5, the geographic coordinates of the location of the risk target are calculated, including The rotation matrix is ​​written as ; Wherein a1, a2, a3, b1, b2, b3, c1, c2, c3 are all abbreviations of the parameters in the rotation matrix constructed from the pixel space coordinate system to the ground auxiliary coordinate system; If the pixel coordinates of the image of the identified risk target are (x, y), the focal length is recorded as F, and the altitude is recorded as H, the Gaussian plane coordinates of the drone's location can be obtained by Gaussian calculation: ( ), so the geographical coordinates of the risk target location are expressed as follows: .

[0013] The present invention also provides a system for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos, which is used to implement the above-mentioned method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos, comprising: The data acquisition module is used to pre-process the data captured by the drone to form a sample data set; The target detection model uses the sample data set to train the target detection network to obtain the optimal weight parameters; the environmental image collected by the drone in real time is used to determine the target detection model applicable to the current scene, and the trained target detection network is used to obtain the identification frame of the risk target; the identification frame contains the pixel coordinates of the risk target and its category information; The coordinate conversion module constructs a rotation matrix from the pixel space coordinate system to the ground-aided coordinate system; based on the pixel coordinates of the identified risk target, the Gaussian plane coordinates of the drone's location are obtained, and the constructed rotation matrix is ​​used to calculate the geographic coordinates of the risk target's location.

[0014] Compared with the existing technology, the intelligent identification method of railway external environment hidden dangers based on drone inspection videos disclosed in this application has built a set of efficient and intelligent inspection and management system solutions for railway external environment hidden dangers by integrating drone inspections and artificial intelligence technologies and combining relevant principles of photogrammetry. It not only realizes the intelligent construction and dynamic expansion of hidden danger sample sets driven by multi-period inspection videos, but also utilizes the integration of deep learning and high-precision positioning technology to significantly improve the accuracy of hidden danger identification and the accuracy of spatial positioning, further enhancing the efficiency of railway external environment inspection and investigation. In addition, a one-click hidden danger report is automatically generated, which greatly simplifies the information aggregation process. The iterative optimization framework based on intelligent identification and manual feedback ensures the continuous optimization of the hidden danger identification model, guarantees the accuracy and stability of the identification capability, and provides reliable intelligent support for the safe operation and maintenance of the railway external environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The present invention is a flow chart of a method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos.

[0016] Figure 2 It is a schematic diagram of labeling sample images in the present invention.

[0017] Figure 3 This is an image detected by a target detection model in one embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention is further described in detail below through the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, an intelligent identification method for railway external environment hidden dangers based on drone inspection videos provided by an embodiment of the present invention includes the following steps: S1, pre-process the data captured by the drone to form a sample data set; S2. Use the sample data set to train the target detection network to obtain the optimal weight parameters; S3, using the environmental image collected by the drone in real time, determining the target detection model applicable to the current scene, and using the trained target detection network to obtain the identification frame of the risk target; the identification frame includes the pixel coordinates of the risk target and its category information; S4, constructing a rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system; S5. According to the pixel coordinates of the identified risk target, the Gaussian plane coordinates of the location of the UAV are obtained, and the geographic coordinates of the location of the risk target are calculated using the constructed rotation matrix.

[0020] In S1, the data captured by the drone is preprocessed to form a sample data set, including: S101, classifying and extracting key frames from the drone's shooting data to form long-focus samples and short-focus samples; Based on multiple periods of UAV inspection video data, taking into account the large differences in target features at different scales, multiple periods of UAV videos are divided into three types of data according to focal length. One is represented by 24mm focal length aerial video, one is represented by 160mm focal length aerial video, and the other is represented by scenes at the bottom of bridge piers shot at a lower aerial altitude. Based on these three types of data, external environmental hazard sample sets are produced respectively.

[0021] Traversing the three types of videos, key image frames with hidden danger targets are extracted manually or automatically, and image frames that are easily confused with the targets to be identified are retained as anti-samples.

[0022] S102, annotating part of the data in the long-focus sample and the short-focus sample to obtain an annotated data set; Obtain sample data taken by the drone, classify the sample data according to focal length, and produce long-focus samples and short-focus samples respectively; Label some data in the long-focus samples and short-focus samples. Use the initial annotation set to train the recognition model and obtain the optimal weight file; Use the trained model to intelligently identify the remaining large amount of data sets to obtain labeled data sets. Figure 2 As shown in the figure, it is the result of marking different types of hidden dangers. The yellow label fcw in the figure represents the dust net; the blue label, cgw, represents the color steel tile; the red label, sld represents the plastic bag or floating object.

[0023] S103, enhancing the sample data, including enhancing the size of the sample data, enhancing the pixels of the sample data, and synchronously enhancing the sample image and the corresponding annotation file; Among them, the enhancement of the size of sample data includes the use of geometric transformation methods (such as rotation, flipping and scaling, etc.) to solve the limitations of different scales and angles of data sets based on the characteristics of drone images; Assume that the image is scaled and the image scaling ratio is ,in is the horizontal scaling ratio, is the vertical scaling ratio, and the coordinate of the upper left corner of the original annotation box is , the coordinates of the lower right corner are , then the coordinates of the upper left corner of the annotation box after the change are , the lower right corner coordinates are ; The calculation is as follows: ②Assume that the image revolves around the origin Rotated counterclockwise , the transformation formula is: ③Assume that the image is flipped around the vertical axis (i.e. flipped left and right), and the width of the image is , the transformation formula is: Assume that the image is flipped around the horizontal axis (i.e., upside down), and the image height is , the transformation formula is: The enhancement of sample data pixels includes: using pixel transformation methods (such as adjusting brightness, saturation, etc.) to take into account the limitations of different shooting conditions of data sets at different time periods.

[0024] The sample image and the corresponding XML annotation file are enhanced synchronously, including: reading the original annotation information, enhancing the image and calculating the new coordinates, updating the bounding box information in the XML annotation file, and saving the updated file. The specific steps are as follows: 1) Read the XML annotation file of the original image and obtain the original bounding box information.

[0025] 2) Enhance the original image and calculate the new coordinates of the bounding box in the enhanced image. The calculation of the new coordinates takes into account the impact of image transformation on the position of the bounding box, ensuring the consistency of the bounding box with the enhanced image.

[0026] 3) Update the bounding box information in the XML annotation file, including the file name, the upper left corner and lower right corner coordinates of the bounding box. The updated file name contains the enhanced serial number to distinguish different enhanced images.

[0027] 4) Use Python code to save the updated XML annotation file.

[0028] S104: Divide the enhanced images and label files into a training set and a validation set according to a specific ratio.

[0029] In S2, the target detection network is trained using the sample data set to obtain the optimal weight parameters; the training set is also input into the target detection network, and the relevant parameters are set before training. By evaluating the performance of the model on the validation set, the round with the best accuracy is selected as the weight file of the final model.

[0030] In S3, the environment image collected by the drone in real time is used to obtain the identification frame of the risk target using the trained target detection network; the identification frame includes the pixel coordinates of the risk target and its category information.

[0031] Use the trained weight file to identify the target image and obtain the identification box of the risk target. The identification box contains the pixel coordinates of the risk target and its category information. Figure 3 As shown, this is an aerial photo of the roof of a building. It can be seen that there are many idle items such as plastics and dust nets in the roof photo. Therefore, when performing target identification, all identified dangerous targets will be displayed with red identification frames.

[0032] When collecting drone flight data, ensure that the drone flight video and POS file are complete. The POS file contains information such as the drone's latitude and longitude, altitude, yaw angle, pitch angle, roll angle, and camera-related parameters.

[0033] Coordinate system conversion: Since the yaw angle, pitch angle, and roll angle in the POS file describe the relationship between the navigation coordinate system and the IMU coordinate system, they need to be converted into the omega, phi, and kappa angles commonly used in photogrammetry.

[0034] S4, constructing a rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system; According to the converted omega, phi, and kappa angles, combined with the camera focal length, altitude, and image pixel coordinates, the geographic coordinates of the pixel coordinates are calculated using the principle of photogrammetry. In the specific calculation process, factors such as the central meridian and latitude must be considered to construct a rotation matrix from the pixel space coordinate system to the ground-auxiliary coordinate system, and the coordinate conversion is completed by combining related formulas. The longitude and latitude coordinates of the location of the drone in a certain frame of image are lon, lat, the central meridian of the Mercator projection graduation zone where the point is located is L, the latitude is B, the yaw angle, pitch angle, and roll angle are h, p, and r, respectively, then the calculation formula for the rotation matrix R from the image space coordinate system to the ground-auxiliary coordinate system is as follows: According to the above calculation results, ; Wherein a1, a2, a3, b1, b2, b3, c1, c2, c3 are all abbreviations of the parameters in the rotation matrix constructed from the pixel space coordinate system to the ground auxiliary coordinate system; S5. According to the pixel coordinates of the identified risk target, the Gaussian plane coordinates of the drone's location are obtained, and the geographic coordinates of the risk target's location are calculated using the constructed rotation matrix R.

[0035] The pixel coordinates of the frame image are recorded as (x, y), the focal length is recorded as F, and the altitude is recorded as H. Using Gaussian calculation, the Gaussian plane coordinates of the drone's position can be obtained as ( ), combined with the above parameters, we have the following formula: .

[0036] Combining the above formulas, the geographical coordinates of the pixel coordinates (x, y) of the hidden danger target can be calculated ( ); S6. After identifying the location of the hidden danger target, establish a background adaptive identification interface according to the following process.

[0037] 1. Front-end and back-end interaction: Use the python-tornado framework to build the back-end interface. The front-end communicates with the back-end in real time through the specified interface number, receives the target recognition results and progress information returned by the back-end, and performs corresponding interface display or operations based on this information.

[0038] 2. Backend algorithm encapsulation: Encapsulate the target detection algorithm, write it into a callable and parameter-passable function type, and embed it into the Tornado framework. The Tornado framework, with its efficient asynchronous processing capabilities, can meet the needs of real-time target recognition.

[0039] It should be noted that in S3, the environmental image collected by the drone in real time is used to obtain the identification frame of the risk target using the trained target detection network; and the following steps are also included: When receiving the drone video address and POS file from the front end, first extract the POS information (including position, height, attitude, etc.) of the 10th frame of the video, and then combine the camera focal length, flight altitude and other parameters to determine the target detection model applicable to the current scene. Using Tornado's asynchronous mechanism, the information of the selected model (such as model name, identification type, etc.) is passed to the front end in real time. The specific principles for selecting the applicable model are summarized as follows: 1) Let the flying height of the drone in this frame be H, in meters, the focal length of the camera be f, in millimeters, and the pixel size be p, in µm. From this, the spatial resolution (GSD) of the drone can be calculated using the following formula: The above calculations take into account unit conversions. When calculating, simply substitute the specified units given above.

[0040] 2) During the model training process, this algorithm sets the attributes of multiple target detection models in advance, and assigns the GSD attribute to the image according to the minimum object size marked during the model making process.

[0041] 3) Compare the GSD calculated in step 1) with the GSD value attributes in multiple models, and select the model with the closest value as the optimal model for identifying the incoming data.

[0042] Target recognition and positioning: After receiving the target recognition command from the front end, the drone video is first divided into several frames according to the preset scene overlap. Then, the selected target detection model is used to infer each frame of the image, and JSON format data including the second of the frame image, the recognized type, the annotated image, the original image, and the latitude and longitude of the image target are returned. These data will be transmitted to the front end in real time for display or further processing.

[0043] 5. Progress monitoring and result output: During the target recognition process, use the python script to record the recognition progress in real time and write the progress information into the progress.txt file in the project folder. At the same time, create a new port to monitor the recognition progress in real time and display it to the front end so that users can view the task progress.

[0044] 6. Front-end display and report output: Check the hidden danger results in the front-end interface to confirm whether the hidden dangers identified intelligently are correct and whether the hidden danger target locations calculated by the positioning algorithm are accurate. After manual review, the hidden danger information is entered into the hidden danger database, and the data can be exported with one click to generate a special report containing key information such as the hidden danger date, interval, location, photos, etc., for relevant personnel to carry out high-standard rectification.

[0045] Hidden dangers are identified through the intelligent recognition technology mentioned above, and the manual review stage is carried out. The results of manual verification are incorporated into new samples, and the sample optimization model of supplementation and error correction is combined to form a closed-loop iterative optimization process. In this way, the accuracy of hidden danger identification can be effectively improved, and the intelligent recognition model can be continuously optimized to make it more suitable for the actual application scenario of hidden danger inspection in the external environment of railways. The relevant steps are as follows: 1. Intelligent identification and manual verification: Intelligent identification technology is used to identify hidden dangers in the external environment of the railway, and the identification results are output to the manual review interface. When the user conducts manual review, the system automatically records the false positives and false negatives in the model during the identification process, and saves the original image and the manually annotated image. The manually annotated samples include the location coordinates and category information of the hidden dangers.

[0046] 2. Sample optimization and data set update: According to the new or corrected sample frame coordinate set manually annotated by the user, the upper left corner coordinate (the minimum value of the xy coordinate) and the lower right corner coordinate (the maximum value of the xy coordinate) of the target annotated rectangular frame are automatically calculated, and spliced ​​into the label txt file required by the model according to specific rules. After the image is named consistent with the txt file, it is saved in the new training data set folder as a supplement and correction sample for model iteration.

[0047] 3. Model architecture adjustment and optimization: Based on the iterated data set, adjust and optimize the architecture of the intelligent recognition model, including adjusting the number of network layers, changing the size of the convolution kernel, introducing the attention mechanism, etc. Retrain and verify the model, and output the accuracy indicators after training, such as accuracy, recall rate, F1 score, etc., to evaluate the performance of the model.

[0048] 4. Model performance monitoring and backtracking: Keep an eye on the performance of the model in actual applications, track and record it. If the model effect deteriorates, such as a decrease in accuracy or an increase in false alarm rate, you can backtrack to the previous version of the model with one click to ensure the continuity and stability of the identification of hidden dangers in the railway external environment. Performance monitoring can promptly identify problems with the model and take corresponding optimization measures, such as increasing training samples and adjusting model parameters.

[0049] Obviously, the above embodiments are only examples for clear explanation, and are not intended to limit the implementation methods. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from them are still within the protection scope of the invention.

Claims

1. A method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos, characterized in that: include: S1, pre-process the data captured by the drone to form a sample data set; S2. Use the sample data set to train the target detection network to obtain the optimal weight parameters; S3, using the environmental image collected by the drone in real time, determining the target detection model applicable to the current scene, and using the trained target detection network to obtain the identification frame of the risk target; the identification frame includes the pixel coordinates of the risk target and its category information; S4, constructing a rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system; S5. According to the pixel coordinates of the identified risk target, the Gaussian plane coordinates of the location of the UAV are obtained, and the geographic coordinates of the location of the risk target are calculated using the constructed rotation matrix.

2. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 1 is characterized in that: In S1, the preprocessing of the data captured by the drone also includes: S101, classifying and extracting key frames from the drone's shooting data to form long-focus samples and short-focus samples; S102, annotating part of the data in the long-focus sample and the short-focus sample to obtain an annotated data set; S103, enhancing all sample data; S104: Divide the enhanced images and the corresponding label files into a training set and a validation set according to a specific ratio.

3. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 2 is characterized in that: In S103, the sample data is enhanced, including: enhancing the size of the sample data, enhancing the pixels of the sample data, and synchronously enhancing the sample image and the corresponding annotation file.

4. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 3 is characterized in that: The sample image and the corresponding annotation file are enhanced synchronously, and the XML annotation file of the original image is read to obtain the original bounding box information; Enhance the original image and calculate the new coordinates of the bounding box in the enhanced image; Update the bounding box information in the XML markup file, including the file name, the upper left corner and lower right corner coordinates of the bounding box; The updated file names contain the enhanced sequence number to distinguish different enhanced images; Use python code to save the updated XML markup file.

5. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 1 is characterized in that: In S3, the environmental image collected by the drone in real time is used to determine the target detection model applicable to the current scene, and the trained target detection network is used to obtain the identification frame of the risk target; the following steps are also included: S301, obtaining the drone video address and POS file transmitted in the previous section, and extracting the POS information of the 10th frame; S302, judging the target detection model applicable to the current scene according to the camera focal length and flight altitude at this time; S303, using Tornado's asynchronous mechanism, transmit the information of the selected target detection model to the front end in real time; S304: Infer each frame of the image according to the selected target detection model, and return JSON format data including the second of the frame image, the recognized type, the annotated image, the original image, and the latitude and longitude information of the image target.

6. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 1 is characterized in that: In S302, determining the target detection model applicable to the current scene includes the following steps: The flight height of the drone to obtain the current frame image is H, the camera focal length is f, and the pixel size is recorded as p. The spatial resolution GSD of the drone is calculated as follows: Image GSD attribute of the minimum object size marked in the preset attributes of each target detection model; The calculated GSD is compared with the GSD value attribute in each target detection model, and the model with the closest value is selected as the optimal model for identifying the incoming data.

7. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 1 is characterized in that: In S4, the constructed rotation matrix from the pixel space coordinate system to the ground auxiliary coordinate system is expressed by the following formula: Among them, lon and lat represent the longitude and latitude coordinates of the drone's location when the current frame image is taken; h, p, and r represent the yaw angle, pitch angle, and roll angle of the drone at this time; L represents the central meridian of the Mercator projection graduation zone where the drone's location is located, and B represents the longitude of the central meridian at this time.

8. The method for intelligently identifying hidden dangers in the railway external environment based on drone inspection videos according to claim 7 is characterized in that: In S5, the geographic coordinates of the location of the risk target are calculated, including The rotation matrix is ​​written as ; Wherein a1, a2, a3, b1, b2, b3, c1, c2, c3 are all abbreviations of the parameters in the rotation matrix constructed from the pixel space coordinate system to the ground auxiliary coordinate system; If the pixel coordinates of the image of the identified risk target are (x, y), the focal length is recorded as F, and the altitude is recorded as H, the Gaussian plane coordinates of the drone's location can be obtained by Gaussian calculation: ( ), the geographical coordinates of the location of the risk target ( ) is expressed as follows: 。 9. An intelligent identification system for railway external environment hidden dangers based on drone inspection videos, characterized in that: The method for intelligently identifying hidden dangers in the external environment of railways based on drone inspection videos as described in any one of claims 1 to 8 above comprises: The data acquisition module is used to pre-process the data captured by the drone to form a sample data set; The target detection model uses the sample data set to train the target detection network to obtain the optimal weight parameters; the environmental image collected by the drone in real time is used to determine the target detection model applicable to the current scene, and the trained target detection network is used to obtain the identification frame of the risk target; the identification frame contains the pixel coordinates of the risk target and its category information; The coordinate conversion module constructs a rotation matrix from the pixel space coordinate system to the ground-aided coordinate system; based on the pixel coordinates of the identified risk target, the Gaussian plane coordinates of the drone's location are obtained, and the constructed rotation matrix is ​​used to calculate the geographic coordinates of the risk target's location.

Images