A ship berth water depth detection and early warning method based on unmanned aerial vehicle image recognition

Abstract

The present application relates to the technical field of ship berth draft detection and early warning, and specifically discloses a ship berth draft detection and early warning method based on unmanned aerial vehicle image recognition. The present application acquires a ship draft line video through an unmanned aerial vehicle; calls an image segmentation network to determine a ship body and water surface segmentation line, and calls a character OCR module to recognize water gauge characters; recognizes the intersection of the water gauge character center point connecting line and the segmentation line to obtain a draft point; performs semantic correction to obtain the draft depth, and calculates the average draft depth of the water gauge in a calm state; corrects and tilts the average draft depth of multiple water gauges to provide real-time ship load and tilt early warning. The present application can accurately recognize the draft depth of large ships parked in a port, provide real-time ship load and tilt early warning, improve port operation management efficiency, reduce ship loading risks, has very significant practical value, has higher recognition accuracy, can overcome the environmental influence of character corrosion, bad weather and other complex scenes, and accurately obtain the ship draft depth.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of ship berth water depth detection and early warning, and particularly relates to a ship berth water depth detection and early warning method based on unmanned aerial vehicle image recognition. BACKGROUND

[0002] Ship water depth detection is one of the daily management work of the port. Large ship waterline data is an important index for evaluating ship load and navigation safety. The waterline reading for measuring the load of a cargo ship is generally required to be accurate to within 5 cm.

[0003] Common ship water depth detection methods include: (1) visual inspection: the management method currently used by many ports has limitations such as inconvenient operation, time-consuming and labor-intensive, long detection period, and detection results affected by subjective judgment of workers; (2) installation of detection devices: generally, a water level detection device is installed on the ship body. This method can eliminate the need for on-site detection, but it requires the installation of additional equipment, and the accuracy cannot be guaranteed, and standardized data exchange between different ships and ports is difficult to achieve; (3) image detection: such as patent CN112598001A, a ship water gauge reading automatic identification method based on multi-model fusion, which uses machine vision method to read the ship water depth, and is more flexible and practical than previous methods.

[0004] Image recognition water gauge detection is the development trend of future port management operations. Generally, fixed monitoring cameras or unmanned aerial vehicles are used to obtain ship water gauge images, and character OCR and waterline detection methods are used to identify water gauge scales. However, the existing technology still has the following defects: in the actual environment, due to long-term corrosion of seawater, water gauge characters are prone to rust, and there are complex scenes such as seawater reflection, waves, and reflections, so general scene OCR recognition algorithms are prone to false detection and missed detection, and the accuracy is low. SUMMARY

[0005] The purpose of the embodiment of the application is to provide a ship berth water depth detection and early warning method based on unmanned aerial vehicle image recognition, which aims to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the embodiment of the application provides the following technical scheme:

[0007] A ship berth water depth detection and early warning method based on unmanned aerial vehicle image recognition, the method specifically comprises the following steps:

[0008] acquiring a ship waterline video by an unmanned aerial vehicle, and uploading the ship waterline video to a video analysis server;

[0009] The ship waterline video is processed by the video analysis server to construct two training sets of image segmentation and character detection, and train a waterline image segmentation model and a character OCR module network model;

[0010] The waterline image segmentation model is called to determine the division line between the ship body and the water surface, and the character OCR module network model is called to identify the water gauge characters;

[0011] The water gauge characters are fitted with center points to obtain a water gauge character center point line, and the intersection point between the water gauge character center point line and the division line is identified to obtain the water point;

[0012] The water gauge characters are subjected to semantic correction and intensive training to obtain accurate water depths, and the accurate water depths are subjected to sea wave fitting to calculate the average water depth of the water gauge in a calm state;

[0013] The average water depths of multiple water gauges are corrected and tilted to provide early warning.

[0014] As a further limitation of the technical scheme of the embodiment of the application, the water gauge character center point fitting to obtain a water gauge character center point line and the identification of the intersection point between the water gauge character center point line and the division line to obtain the water point specifically comprises the following steps:

[0015] Obtain the character target;

[0016] The center point pixel coordinates (x i ,y i ) of the character target are used to calculate the scale division, wherein the x coordinate is used to determine whether the character belongs to a depth value or a scale division value, the y coordinate is used for depth center point straight line fitting and physical depth calculation, and a segmented fitting method is used to fit the water gauge character center point line;

[0017] The intersection point between the water gauge character center point line and the division line is identified to obtain the water point.

[0018] As a further limitation of the technical scheme of the embodiment of the application, the calculation process of using the center point pixel coordinates (x i ,y i ) of the character target to calculate the scale division, wherein the x coordinate is used to determine whether the character belongs to a depth value or a scale division value, the y coordinate is used for depth center point straight line fitting and physical depth calculation, and a segmented fitting method is used to fit the water gauge character center point line is as follows:

[0019] The characters in the character target are sorted and stored in sequence S according to the y coordinates from small to large;

[0020] Find the numeric characters parallel to character M from sequence S, combine them into the water gauge depth value according to the x coordinate and move them to the depth sequence D. D contains at least two depth tuples. Each tuple consists of 1 to 2 numeric characters sorted by x coordinate, indicating the water gauge depth.

[0021] According to the sequence S, the center point connecting line L is segmented and fitted to establish the corresponding sequence of pixel coordinate y and physical coordinate , and calculate the physical depth corresponding to each pixel value under the current fitting line;

[0022] Read the water point in the center point connecting line L, and calculate the sequence according to the pixel distance of the character closest to the water point in the sequence S. The corresponding physical draft.

[0023] As a further limitation of the technical solution of the embodiment of the present invention, the rules for performing semantic correction and reinforcement training on the draft gauge characters to obtain accurate draft depth are:

[0024] Depth information is mutually verified. When there are more than two water gauge depth characters in the image, the y coordinate of the depth center point is proportionally associated with the depth scale, and the difference between two consecutive depth values ​​is 1m;

[0025] The depth character number is flush with the bottom scale of the character M. When the absolute value difference of the y coordinate is < (Pick When the pixel value is mapped to a physical distance of 0.5 cm, the digital characters and the character M are determined to be the water gauge depth, and the x-axis coordinates of the three characters are used for depth reading and verification;

[0026] The scale characters appear alternately in the order of M, 8, 6, 4, and 2, and the y-coordinate distances of their center points are proportionally related to each other, so as to perform error detection correction and missed detection filling.

[0027] The detection result is obtained, and when the detection result does not match the physical draft depth, the recognition error is determined, correction is automatically performed, the manual verification phase is automatically entered, and the corrected data increment is added to the training set.

[0028] As a further limitation of the technical solution of the embodiment of the present invention, the draft is subjected to wave fitting to calculate the average draft under a calm water gauge state:

[0029] Using the one-dimensional wave equation: y ( x , t ) = A cos[ ω ( t − x μ ) + φ ] ,in, For the front The draft depth at that moment, is the current draft depth, 、 、 、 are the wave equation parameters;

[0030] The curve fitting function curve_fit in the Python machine learning library SciPy was used to fit the equation with the draft depth detection data of the continuous frame image to obtain the parameters. The average draft depth was obtained by fitting the peak and trough data of the curve.

[0031] As a further limitation of the technical solution of the embodiment of the present invention, in the correction of the average draft of multiple water gauges and the tilt warning:

[0032] Calculate the right side draft and the right side fore-aft inclination angle;

[0033] Calculate the left side draft depth and the left side fore-aft inclination angle;

[0034] Calculate the corrected draft and left and right heel angles;

[0035] When the left-right tilt angle, the right front-back tilt angle or the left front-back tilt angle is greater than a warning threshold, a tilt warning is issued.

[0036] As a further limitation of the technical solution of the embodiment of the present invention, in the calculation of the right side draft depth and the right side front-back tilt angle:

[0037] The tangent of the right front-to-back inclination angle is ;

[0038] According to the triangle similarity principle, the relationship between h1 and h2 is = ,and ,and , where l1 and l2 are the distances from the water gauges r1 and r2 to the center point c of the ship respectively;

[0039] The right front-to-back tilt angle is solved as: .

[0040] As a further limitation of the technical solution of the embodiment of the present invention, in the correction of the average draft of multiple water gauges and the tilt warning:

[0041] The coordinates of the water gauge position on the left are (x L_r1 ,y L_r1 )、(x L_r2 ,y L_r1 ), the coordinates of the water gauge position on the right are (x R_r1 ,y R_r1 ), (x R_r2 ,y R_r2 ), where x is the horizontal distance from the water gauge to the bow, and y is the draft;

[0042] When reading the water gauge water depth, such as y L_r1 , y L_r1 , y R_r1 , y R_r2 When the water depth readings are inconsistent, it is determined that the ship body has front-to-back tilt or left-to-right tilt.

[0043] Compared with the prior art, the beneficial effects of the present application are:

[0044] 1. The present application can accurately identify the water depth of large ships parked in the port, and provide real-time ship load and tilt warning, which has very significant practical value for improving port operation management efficiency and reducing ship loading risk.

[0045] 2. The waterline identification scheme using character recognition and image segmentation adopted by the present application has high recognition accuracy, can overcome the influence of complex scenes such as character corrosion and bad weather, and can accurately obtain the water depth of the ship in the wave scene. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application.

[0047] Fig. 1 The waterline character information and reading schematic diagram provided by the embodiment of the present application is shown.

[0048] Fig. 2 The waterline segmentation and character OCR effect diagram provided by the embodiment of the present application is shown.

[0049] Fig. 3 The schematic diagram of sea wave fluctuation fitting curve provided by the embodiment of the present application is shown.

[0050] Fig. 4 The large ship water gauge position top view provided by the embodiment of the present application is shown.

[0051] Fig. 5 The first schematic diagram of single-side water tilt correction provided by the embodiment of the present application is shown.

[0052] Fig. 6 The second schematic diagram of single-side water tilt correction provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of the present application more clear, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0054] It can be understood that the image recognition water gauge detection is the development trend of future port management operation, generally using fixed monitoring camera or unmanned aerial vehicle to obtain ship water gauge image, and recognizing water gauge scale through character OCR and waterline detection method. However, the existing technology still has the following defects: in the actual environment, the general scene OCR recognition algorithm is prone to misjudgment and omission, and the accuracy is low due to the rust of water gauge characters caused by long-term seawater erosion, and complex scenes such as seawater reflection, sea waves and reflection.

[0055] To solve the above problems, the embodiment of the application acquires ship waterline video by unmanned aerial vehicle; calls image segmentation network to determine the segmentation line of ship body and water surface, and calls character OCR module to recognize water gauge characters; recognizes the intersection of water gauge character center point connecting line and segmentation line to obtain waterline; performs semantic correction to obtain water depth, and calculates the average water depth of water gauge in calm state; corrects and tilts the average water depth of multiple water gauges to give early warning. It can accurately recognize the water depth of large ships parked in the port, provide real-time ship load and tilt early warning, has very significant practical value for improving port operation management efficiency and reducing ship loading risk, has higher recognition accuracy, can overcome the influence of complex scenes such as character rust and bad weather, and accurately obtain the water depth of the ship.

[0056] Specifically, in one preferred embodiment provided by the application, a ship berth water depth detection and early warning method based on unmanned aerial vehicle image recognition, the method specifically comprises the following steps:

[0057] Acquire ship waterline video by unmanned aerial vehicle, and upload the ship waterline video to video analysis server;

[0058] Process the ship waterline video by video analysis server, construct two training sets of image segmentation and character detection, train waterline image segmentation model and character OCR module network model;

[0059] Call waterline image segmentation model to determine the segmentation line of ship body and water surface, and call character OCR module network model to recognize water gauge characters;

[0060] Fit the center point of the water gauge characters to obtain water gauge character center point connecting line, recognize the intersection of the water gauge character center point connecting line and the segmentation line to obtain water point;

[0061] Perform semantic correction and reinforcement training on the water gauge characters to obtain accurate water depth, and perform sea wave fitting on the accurate water depth to calculate the average water depth of water gauge in calm state;

[0062] Correct and tilt the average water depth of multiple water gauges to give early warning.

[0063] In the embodiment of the present application, the implementation process of the ship berth water depth detection and early warning method based on unmanned aerial vehicle image recognition is as follows: (1) using an unmanned aerial vehicle to collect a ship waterline image, uploading the video to a video analysis server through an operator and a management system, and transmitting a ship ID identifier or a water detection configuration parameter; (2) calling a waterline image segmentation model to determine the division line between the ship body and the water surface, and calling a character OCR module network model to identify the water gauge character; (3) fitting the water gauge character center point to obtain the water gauge character center line, and obtaining the intersection point of the waterline and the water gauge center line, i.e. the water depth point; (4) obtaining the water depth through character recognition result and semantic correction; fitting the water depth of continuous frames with sea waves to calculate the water depth of the water gauge in a calm state; (5) combining the ship information and the multiple water gauge depths to give the ship water depth information and early warning.

[0064] Specifically, in the water gauge character and water depth point detection, the PaddleOCR character recognition framework and the BiSeNetV2 semantic segmentation network are used to jointly realize the water gauge character and water depth point detection. First, 50 training pictures are selected from 50 videos by frame skipping, and 50 frames of pictures are extracted from each video, a total of 2500 pictures are obtained. The 2500 pictures are randomly rotated by 5 degrees, 10 degrees and 15 degrees to obtain 2500 enhanced samples, which are combined with the original samples, a total of 5000 pictures. The combined sample set is randomly operated to increase noise, adjust brightness and adjust contrast to obtain 5000 enhanced samples. All samples are combined, a total of 10000 samples. In order to improve the labeling efficiency, the initial training model is used for automatic labeling and manual adjustment and iterative optimization. First, 100 samples are labeled, the initial model inti_model is trained, the initial model inti_model is introduced into the automatic labeling program, and the remaining 400 samples are automatically labeled and manually corrected. The optimized model finetune_model is retrained, the optimized model finetune_model is introduced into the automatic labeling program again, 1000 samples are automatically labeled and manually corrected, and finally the final_model is trained, the remaining model is labeled and optimized, and a complete training set is obtained. The training samples are allocated in the ratio of 6:3:1 for training: testing: validation. In the initial stage of actual application, the manual verification method is adopted, the misdetected and missed samples are added to the training set, and the model is continuously iteratively optimized. cos[The waterline character information and reading schematic diagram provided by the embodiment of the present application is shown, the draft depth is marked in meters, only four even characters 2, 4, 6 and 8 appear between two depth characters “M” to represent the sub-scale, the character height is 10 CM, the spacing between characters is 10 CM, the cropped English character dictionary set {0, 2, 4, 6, 8, 1, 3, 5, 7, 9, M} is used as a marking label in consideration of the actual distribution of the characters of the water gauge, and the character recognition model uses PaddleOCR single character label for marking, wherein the four characters 2, 4, 6 and 8 are used to mark the sub-scale of the water gauge, and the ship and water body segmentation model uses BiSeNetV2 semantic network for marking, including two parts of the ship body and the water body; the scale is calculated using the center point pixel coordinates (xi, yi) of the character target, wherein the x coordinate is used to judge whether the character belongs to the depth value or the sub-scale value, the y coordinate is used for straight line fitting of the depth center point and physical depth calculation, and a segmented fitting method is used to fit the center point connecting line of the water gauge. The calculation process is as follows: 1) the characters in the character target are sorted in ascending order of y coordinates and stored in sequence S; 2) find the number characters parallel to the character M from the sequence S, combine them into water gauge depth values according to the x coordinates, and move them to the depth sequence D, D contains at least two depth tuples, each tuple is composed of 1 to 2 number characters sorted by x coordinates, representing the water gauge depth of a scale is how many meters; 3) segment fitting of the center point connecting line L according to the sequence S, establishing the sequence of pixel coordinates y and physical coordinates , and calculating the physical depth corresponding to each pixel value under the current fitted straight line; 4) reading the draft point in the center point connecting line L, calculating the physical draft depth corresponding to the sequence according to the pixel distance of the nearest character of the draft point in the sequence S, such as ω The waterline segmentation and character OCR effect diagram provided by the embodiment of the present application is shown.

[0065] Specifically, the water gauge characters are semantically corrected, the character set {0, 2, 4, 6, 8, 1, 3, 5, 7, 9, M} and the context relationship thereof are considered, and the check and correction rules are introduced as follows: (1) mutual verification of depth information, when there are more than two water gauge depth characters in the image, the y coordinates of the depth center points are in equal ratio relation with the depth scale, and the difference between two consecutive depth values is 1 m; (2) the depth character number is flush with the bottom scale of the character M, when the absolute value difference of the y coordinates is less than 5 pixels, the character is corrected to be the character M. (take (the pixel value mapped to the physical distance of 0.5cm) is used to determine that the digital characters and the character M are the water gauge depth, and the x-axis coordinates of the three characters are used for depth reading and verification; (3) The scale characters appear alternately in the order of M, 8, 6, 4, and 2, and the y-coordinate distances of their center points are proportionally related. This information can be used to correct false positives and fill in missed detections. For example, if a detection sequence is {(8,51), (8,119), (4,201), (2,266), (m,338)}, based on the distribution of the y-coordinates of the center points, it can be determined that (8,119) is an incorrectly recognized digital character and should be corrected to (6,119); 4) The water body segmentation area is generally located below the hull, and the connected diameter of the segmentation area is greater than 1 / 2 of the image width. When the detection result does not meet the requirements, the system determines that the recognition is wrong, automatically corrects it, and automatically enters the manual verification stage. The corrected data increment is added to the training set to continuously improve the coverage of the proofreading library and improve the accuracy of the model.

[0066] Specifically, the draft depth is fitted with waves, and the average draft depth under the calm water gauge is calculated. A one-dimensional wave equation is introduced to solve the problem of draft depth interference caused by wave fluctuations on continuous frame images:

[0067] y ( x , t ) = A μ φ ( t − x Fig. 1 ) + Fig. 2 ]

[0068] in, For the front The draft depth at that moment, is the current draft depth, 、 、 、 The parameters of the wave equation are obtained by fitting the equation using the curve_fit function of the Python machine learning library SciPy and the draft depth detection data of the continuous frame image. The average draft of the draft point is obtained by fitting the peak and trough data of the curve, as shown in the following example: cos[ A schematic diagram of a wave fluctuation fitting curve provided by an embodiment of the present invention is shown, wherein the red curve is the fitted fluctuation curve and the green straight line is the average draft under the fluctuation condition.

[0069] Specifically, in the process of correcting the average draft of multiple water gauges and providing tilt warning, according to the draft detection values ​​of multiple water gauges of large ships, the ship draft correction and tilt warning are given, such as ω The diagram shows a top view of the water gauge position of a large ship provided by an embodiment of the present invention. Two water gauges are distributed on the left and right sides of the hull. The coordinates of the water gauge position on the left side are (x L_r1 ,y L_r1 ),(x L_r2 ,y L_r1 ), the right side is (xR_r1 y R_r1 , R_r2 y R_r2 , L_r1 wherein x is the horizontal distance from the water gauge to the bow of the ship, y is the draft depth, when the image recognition module reads the water gauge draft depth, such as y L_r1 , R_r1 , R_r2 when the water gauge depth readings are inconsistent, it can be determined that the ship body has front-to-back or left-to-right inclination, and the specific calculation of the draft data of the ship in the calm state is as follows:

[0070] Taking the right side water gauge depth as an example, as shown in the first schematic diagram for single-side water inclination correction provided by the embodiment of the present application, μ the second schematic diagram for single-side water inclination correction provided by the embodiment of the present application is shown, and the tangent value of the front-to-back inclination angle of the ship body is φ Fig. 3 Fig. 4 Fig. 5 Fig. 6 ; ;

[0071] According to the principle of similar triangles, h1 and h2 have the following relationship: = and

[0072] wherein l1 and l2 are the distances from the water gauges r1 and r2 to the center point c of the ship, respectively:

[0073]

[0074] The right side front-to-back direction inclination angle can be obtained, and the right side water gauge draft depth is yR_c=yr1'=yr2', and by the same method, the left side front-to-back direction inclination angle can be obtained, and the draft depth, wherein yL_c=yl1'=yl2', at this time, if the left and right draft depths yL_c and yR_c of the ship are not equal, the left and right direction inclination angle of the ship body can be obtained, wherein W is the ship width, and the final draft depth of the ship is When the front-to-back direction or left-to-right direction inclination angle of the ship body is greater than the early warning threshold, an inclination early warning is issued.

[0075] The technical features of the above-described embodiments can be combined in any manner, and to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0076] ​The above embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but should not be understood as a limitation on the patent scope of the present application. It should be noted that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

[0077] The above merely describes the preferred embodiments of the present application and should not be used to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for detecting and warning the draft depth of a ship based on drone image recognition, characterized in that: The method specifically comprises the following steps: Acquire a video of the ship's waterline through a drone, and upload the video to a video analysis server; The video analysis server processes the ship's waterline video, constructs two training sets for image segmentation and character detection, and trains a waterline image segmentation model and a character OCR module network model; The waterline image segmentation model is called to determine the dividing line between the hull and the water surface, and the character OCR module network model is called to recognize the water gauge characters; Performing center point fitting on the water gauge character to obtain a line connecting the center points of the water gauge character, identifying an intersection of the line connecting the center points of the water gauge character and the dividing line to obtain a draft point; Perform semantic correction and reinforcement training on the draft gauge characters to obtain accurate draft, perform wave fitting on the accurate draft, and calculate the average draft of the draft gauge in a calm state; The average draft depth of multiple water gauges is corrected and tilt warning is performed, the center point fitting of the water gauge characters is performed to obtain the water gauge character center point connection line, and the intersection of the water gauge character center point connection line and the dividing line is identified to obtain the draft point. Specifically, the following steps are included: obtaining the character target; using the center point pixel coordinates of the character target (x i ,y i ) calculates the sub-scale, wherein the x coordinate is used to determine whether the character belongs to the depth value or the sub-scale value, and the y coordinate is used for the depth center point straight line fitting and physical depth calculation, and the segmented fitting method is used to fit the water ruler character center point connection line; identify the intersection of the water ruler character center point connection line and the segmentation line to obtain the draft point, and the pixel coordinates of the center point of the character target (x i ,y i ) calculates the sub-scale, where the x coordinate is used to determine whether the character belongs to the depth value or the sub-scale value, and the y coordinate is used for depth center point straight line fitting and physical depth calculation, and the calculation process of fitting the center point connection line of the water ruler character using the segmented fitting method is as follows: sort the characters in the character target by y coordinate from small to large and store them in the sequence S; find the digital characters parallel to the character M from the sequence S, combine them into the water ruler depth value according to the x coordinate and move them to the depth sequence D, D contains at least two depth tuples, each tuple consists of 1 to 2 digital characters sorted by x coordinate, representing the water ruler depth; segmentally fit the center point connection line L according to the sequence S, establish the pixel coordinate y and the physical coordinate corresponding sequence S′, and calculate the physical depth corresponding to each pixel value under the current fitting line; read the draft point in the center point connection line L, and calculate the corresponding physical draft point in the sequence S′ according to the pixel distance to the nearest character in the sequence S. Water depth, the semantic correction and intensive training of the water gauge characters are carried out to obtain the rules for accurate draft depth: the depth information is verified with each other, when there are more than two water gauge depth characters in the image, the y coordinate of the depth center point is proportionally associated with the depth scale, and the difference between two consecutive depth values ​​is 1m; the depth character number is flush with the bottom scale of the character M, and when the y coordinate absolute value difference is <ε (ε is taken as the pixel value mapped to the physical distance of 0.5cm), the digital character and the character M are determined to be the water gauge depth, and the x-axis coordinates of the three characters are used for depth reading and verification; the scale characters appear alternately in the order of M, 8, 6, 4, and 2, and the y coordinate distances of their center points are proportionally associated, and error detection calibration and missed detection filling are performed; the detection result is obtained, and when the detection result does not match the physical draft, the recognition error is determined, and correction is automatically performed, and the manual verification link is automatically entered, and the corrected data increment is added to the training set.

2. The method for detecting and warning the ship's draft depth based on drone image recognition according to claim 1 is characterized in that: The draft is subjected to wave fitting to calculate the average draft in a calm state: Using the one-dimensional wave equation: Among them, x is the draft depth at the previous t moments, y is the draft depth at the current moment, A, ω, μ, are the wave equation parameters; The curve fitting function curve_fit of the Python machine learning library SciPy and the draft depth detection data of the continuous frame image are used to fit the equation to obtain the parameters. The average draft depth is obtained by fitting the peak and trough data of the curve.

3. The method for detecting and warning the ship's draft depth based on drone image recognition according to claim 1 is characterized in that: In the correction of the average draft of multiple water gauges and the tilt warning: Calculate the right side draft and the right side fore-aft inclination angle; Calculate the left side draft depth and the left side fore-aft inclination angle; Calculate the corrected draft and left and right heel angles; When the left-right tilt angle, the right front-back tilt angle or the left front-back tilt angle is greater than a warning threshold, a tilt warning is issued.

4. The method for detecting and warning the ship's draft depth based on drone image recognition according to claim 3 is characterized in that: In the calculation of the right side draft depth and the right side fore-aft tilt angle: The tangent of the right front-to-back inclination angle is According to the triangle similarity principle, the relationship between h1 and h2 is And h1+h2=abs(y R_r2 -y R_r1 ),and Where l1 and l2 are the distances from the water gauges r1 and r2 to the center point c of the ship respectively; The right front-to-back tilt angle is solved as: α = arctan(tan(α)).

5. The method for detecting and warning the ship's draft depth based on drone image recognition according to claim 4 is characterized in that: In the correction of the average draft of multiple water gauges and the tilt warning: The coordinates of the water gauge position on the left are (x L_r1 ,y L_r1 ), (x L_r2 ,y L_r1 ), the coordinates of the water gauge position on the right are (x R_r1 ,y R_r1 ), (x R_r2 ,y R_r2 ), where x is the horizontal distance from the water gauge to the bow, and y is the draft; After reading the draft depth, if y L_r1 、y L_r1 、y R_r1 ,y R_r2 When the draft depth readings are inconsistent, it is determined that the hull is tilted fore and aft or left and right.

Images