Deep-sea hydrothermal plume motion characteristics and temperature field inversion method based on video analysis

Abstract

The application discloses a deep-sea hydrothermal plume motion characteristic and temperature field inversion method based on video analysis, and relates to the technical field of deep-sea observation and marine fluid visualization measurement. After video frame cutting preprocessing, the application constructs a double-layer partition of outer layer detection of new particles and inner layer tracking of existing particles, and is supplemented by three-level progressive remedial tracking and confidence management, so that stable and continuous tracking of particle trajectories in a complex flow field is realized, and a two-dimensional average flow field is obtained through gridding statistics; a reference scale calibration and three-dimensional attitude angle correction are introduced, a monocular relative depth map is converted into an absolute depth reference map; two types of flow velocity-temperature empirical models, i.e., a dispersion type power function and a focused type inverse tangent function, are constructed, and two-dimensional temperature field distribution is obtained by substituting the flow field inversion. The application does not need additional sensing equipment, and synchronously outputs the flow field, the absolute depth and the temperature field, thereby providing integrated low-cost analysis support for deep-sea hydrothermal plume research.

Description

Technical Field

[0001] This invention relates to the field of deep-sea observation and ocean fluid visualization measurement technology, specifically to a method for retrieving the motion characteristics and temperature field of deep-sea hydrothermal plumes based on video analysis. Background Technology

[0002] Deep-sea hydrothermal plumes are an important product of the interaction between the Earth's crust and the ocean. After being released from seafloor vents, high-temperature hydrothermal fluids mix with the surrounding cold seawater to form a unique plume-like fluid structure with buoyancy characteristics. Hydrothermal plumes are not only rich in thermal energy and chemical substances, but also have a crucial impact on the energy and material cycles of deep-sea ecosystems. Therefore, accurately capturing their motion characteristics is of great significance to deep-sea research.

[0003] Current research on deep-sea hydrothermal plumes mainly falls into three categories: sensor observation, laboratory simulation experiments, and numerical simulation. Sensor observation typically uses manned submersibles, autonomous underwater vehicles (AUVs), or remotely operated vehicles (ROVs) in conjunction with various measurement sensors to directly acquire information such as temperature and chemical composition near the vent region. However, due to limitations such as operational time, equipment cost, and the complex deep-sea environment, this method is difficult to achieve long-term continuous dynamic monitoring of the plume and cannot comprehensively present the overall flow field characteristics of the plume as well as the velocity changes in peripheral local areas.

[0004] Laboratory physical simulations, through the artificial construction of controllable jet flumes, study the rising, mixing, and diffusion processes of hydrothermal plumes. This allows for relatively detailed observation of the complete process of local vortex formation and turbulent diffusion, providing important reference value for plume formation mechanisms. However, laboratories cannot simultaneously replicate the natural conditions of the deep-sea environment, such as high pressure, low temperature, complex background currents, and irregular topography, thus failing to reflect the true motion characteristics of natural deep-sea hydrothermal plumes.

[0005] While numerical simulations can provide continuous spatiotemporal information to reveal plume motion characteristics and energy transport processes, their results are highly dependent on boundary condition settings, physical model assumptions, and numerical accuracy, and are still relatively idealized compared to the complex environment of the real seabed.

[0006] Particle tracking and velocimetry (PTV) based on video images treats suspended particles in a plume as natural tracer particles. By tracking the same particle in consecutive video frames, its trajectory is obtained, and its velocity and direction are calculated. This allows for the simultaneous acquisition of flow information from multiple locations. However, the basic PTV method suffers from a high false recognition rate, easy trajectory interruption, and poor tracking stability under complex flow field conditions.

[0007] Furthermore, conventional PTV is limited to two-dimensional image planar analysis, making it difficult to obtain depth information and effectively characterize the three-dimensional spatial structure of plumes. Although monocular depth estimation techniques can recover the spatial hierarchy of a scene from a single image at a relatively low cost, due to the lack of a real distance benchmark, their depth results only reflect relative relationships and cannot establish a correspondence with real spatial distances.

[0008] Therefore, existing technologies still have the following technical problems: basic PTV lacks a stable tracking mechanism adapted to the complex flow field of plumes; monocular depth estimation lacks absolute scale calibration methods; and a quantitative mapping relationship between the velocity field and the temperature field has not yet been established, making it impossible to directly invert temperature distribution based on video images. These technical problems urgently need to be solved.

[0009] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention provides a method for retrieving the motion characteristics and temperature field of deep-sea hydrothermal plumes based on video analysis.

[0011] The method includes the following steps: Step S1: Acquire deep-sea hydrothermal plume video image data, perform frame segmentation on the video image data to obtain a video frame sequence, and preprocess the video frame sequence to obtain a preprocessed video frame sequence. Step S2: In response to the selected multiple positioning endpoints, a custom mask is generated, which determines the analysis area and the two-layer responsibility partition of the outer responsibility area and the inner responsibility area; wherein, the outer responsibility area is used to detect newly emerging particles and record their short-term motion trajectory, and the inner responsibility area is used to track particles with existing historical trajectory information and perform remedial search tracking after tracking loss occurs. The particle tracking velocimetry method is used to identify and track particles in the preprocessed video frame sequence to obtain particle trajectory information. Based on the particle trajectory information, the analysis area is divided into grids, and the particles passing through each grid are counted and averaged over time to obtain the two-dimensional average flow field. The remedial search and tracking performed by the inner responsibility area includes the following three-level detection strategy: First-level detection: Predict the current frame position based on the previous frame position and velocity in the particle historical trajectory information, and set up a first search window around the predicted position to detect particles; If the first-level detection fails, the second-level detection is performed: the frame difference and brightness detection thresholds are adjusted, and a fan-shaped search area is constructed based on historical trajectory information. Particle detection is performed within the fan-shaped search area in a range larger than the first search window. If the second-level detection fails, the third-level detection is performed: the search window is adjusted with the predicted position as the center, and the frame difference and brightness detection thresholds are adjusted to detect particles. The candidate point is selected as the matching result based on a comprehensive judgment of the distance between the candidate point and the predicted position, the distance between the candidate point and the position of the previous frame, the consistency of the motion direction, and the reverse motion penalty. Step S3: In response to the scale calibration operation of the reference object, obtain the pixel length and true length of the reference object, and calculate the pixel ratio; use a monocular depth estimation model to perform relative depth estimation on the images in the video frame sequence to obtain a relative depth map; in response to the set three-dimensional pose angle of the reference object, combine the pixel ratio and the camera focal length to perform absolute scale transformation on the relative depth map to generate an absolute depth reference map with actual distance scale. Step S4: In response to the preset target nozzle type, substitute the two-dimensional average flow field obtained in step S2 into the velocity-temperature empirical inversion model corresponding to the target nozzle type to obtain the two-dimensional temperature field distribution of the target region; the velocity-temperature empirical inversion model includes a power function model for a dispersion nozzle and an arctangent function model for a focusing nozzle.

[0012] Preferably, the preprocessing in step S1 includes at least one of image grayscale conversion, Gaussian filtering noise smoothing, threshold segmentation, and mask control.

[0013] Preferably, the inner responsibility area sets a confidence level for each particle trajectory, with the initial confidence level being a preset value; when any level of detection is successful, the confidence level is increased; when all detections fail, the confidence level is decreased; when the confidence level is lower than a first threshold or the number of consecutive lost frames exceeds a second threshold, the lifecycle of the particle trajectory is terminated.

[0014] Preferably, the absolute scale transformation in step S3 specifically includes: Based on the principle of pinhole imaging, the ideal reference distance is calculated using the camera focal length, the true length of the reference object, and the pixel length of the reference object in the image. A rotation matrix is ​​constructed based on the three-dimensional attitude angle of the reference object set by the user. The projection ratio of the major axis of the reference object in the camera image plane is calculated. The attitude of the ideal reference distance is corrected using the projection ratio to obtain the corrected reference distance. Using the corrected reference distance as a conversion benchmark, the relative depth map is scaled to obtain the absolute depth reference map.

[0015] Preferably, the absolute scale conversion is performed according to the following formula: In response to the scale calibration operation on the reference object, according to the formula Calculate pixel ratio ,in The actual length of the reference object; x 1, y 1) and ( x 2, y 2) These are the pixel coordinates of the two ends of the reference object selected in the image. x 1. y 1 represents the x-coordinate and y-coordinate of the first positioning point in the image coordinate system, respectively. x 2. y 2 are the x and y coordinates of the second positioning point in the image coordinate system, respectively, and the Euclidean distance calculated from the coordinates of the two positioning points is the reference pixel length; Set the angles of rotation of the reference plane around the X, Y, and Z axes. , , ,according to , , Construct rotation matrix ,in, This represents the rotation angle of the reference plane about the X-axis of the camera coordinate system. This represents the rotation angle of the reference plane about the Y-axis of the camera coordinate system. R represents the rotation angle of the reference plane around the Z-axis of the camera coordinate system, where R is the rotation angle according to the preset rotation sequence. , , The resulting attitude rotation matrix represents the magnitude of the object's rotation. The major axis of the reference object is assumed to be the Y-axis, therefore its local direction vector is... Through the attitude rotation matrix Transform the direction vector to the camera coordinate system , , , These represent the components of the major axis of the reference object in the X, Y, and Z directions of the camera coordinate system, respectively. According to the formula Calculate the projection ratio of the major axis of the reference object onto the camera's imaging plane. The projected length of the reference object in the y-direction of the camera's imaging plane. ; According to the formula Calculate the absolute distance of the midpoint of the reference object. Where f is the camera focal length, The pixel length between the two ends of the reference object; The absolute distance from the midpoint of the reference object Using the relative depth map as a conversion benchmark, a scale transformation is performed to obtain an absolute depth reference map.

[0016] Preferably, the method for constructing the flow velocity-temperature empirical inversion model in step S4 includes: Construct a temperature-velocity dataset, which includes nozzle temperature, plume velocity, and nozzle morphology parameters; Based on the nozzle morphology, temperature characteristics, and jet flow state, hydrothermal nozzles are classified into two categories: dispersive nozzles and focusing nozzles. The temperature-velocity relationship of the two types of nozzles was fitted separately: a power function model was used for the dispersion nozzle, and an arctangent function model was used for the focusing nozzle.

[0017] Preferably, the custom mask in step S2 is determined by selecting four positioning endpoints, the analysis area is defined by the mask, and the outer responsibility area and the inner responsibility area are divided by the inner and outer boundaries of the mask.

[0018] The present invention provides a method for retrieving the motion characteristics and temperature field of deep-sea hydrothermal plumes based on video analysis, which has the following beneficial effects: (1) In terms of particle tracking, a two-layer partitioning system is constructed, consisting of an outer responsibility region and an inner responsibility region. The outer layer is responsible for detecting and recording newly emerging particles in a short period of time, while the inner layer focuses on the continuous tracking of existing particles, thus reducing the occurrence of misidentification and repeated trajectory establishment. At the same time, the inner layer adopts a three-level progressive remedial search and tracking strategy, combined with a confidence management mechanism to dynamically regulate the life cycle of particle trajectories. This improves the problem of easy trajectory interruption and poor tracking stability of the basic particle tracking velocimetry method under complex flow field conditions, and can output more complete and reliable particle trajectory information.

[0019] (2) In terms of depth estimation, this invention, based on the relative depth map output by the monocular depth estimation model, combines the reference object scale calibration and three-dimensional attitude angle correction mechanism. Through pixel ratio calculation, pinhole imaging principle and attitude angle rotation matrix construction, the projection ratio of the long axis of the reference object in the imaging plane is calculated and the attitude is corrected. Finally, an absolute depth reference map based on reference scale correction is generated, which effectively improves the problem that the traditional monocular depth estimation results lack real scale constraints and are difficult to directly correspond to the actual spatial distance.

[0020] (3) In terms of temperature field inversion, this invention constructs a dataset including nozzle temperature, flow velocity and morphological parameters, establishes empirical relationship models of flow velocity-temperature for different types of hydrothermal vents, and uses power function model and arctangent function model for fitting respectively, and couples them with the two-dimensional average flow field output by particle velocity measurement, realizing the estimation of hydrothermal plume temperature parameters from video flow field information, and expanding the application scope of video flow field analysis technology in the study of temperature characteristics of deep-sea hydrothermal plumes. Attached Figure Description

[0021] Figure 1 These are comparison images of image preprocessing results; among them, Figure 1 (a) in the image is the image before Gaussian filtering; Figure 1 (b) in the figure shows the effect after Gaussian filtering.

[0022] Figure 2 This is a graph showing the average velocity-time variation trend of particle tracking.

[0023] Figure 3 This is a flowchart of the improved PTV particle tracking algorithm.

[0024] Figure 4 This is a two-dimensional average flow field vector diagram.

[0025] Figure 5 This is a diagram illustrating the user's attitude angle adjustment.

[0026] Figure 6 This is a graph showing the velocity variation over time for some high-quality particles; among them, Figure 6 (a) in the text refers to particle number 1999; Figure 6 (b) in the text refers to particle number 3200; Figure 6 (c) in the text represents particle number 3306; Figure 6 (d) in the figure represents particle number 4766.

[0027] Figure 7 This is a flowchart of the absolute depth estimation process.

[0028] Figure 8 This is an example of an absolute depth map result.

[0029] Figure 9 The graph shows the fitting results of the temperature-flow velocity power function for a dispersion-type nozzle.

[0030] Figure 10 The graph shows the fitting results of the arctangent function of temperature-flow rate for a focusing nozzle.

[0031] Figure 11 This is a two-dimensional temperature field distribution map based on velocity field inversion.

[0032] Figure 12 This is a flowchart illustrating the overall processing of the present invention. Detailed Implementation

[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] The video data primarily originates from multiple DIVE observation missions conducted by a remotely operated vehicle (ROV) between 2015 and 2022. This data covers multiple hydrothermal vents and their surrounding areas, with specific locations including JADE, Dragon, PAC-Fenway, Desmos, and PAC-Satanic, providing representative samples for analyzing the dynamic differences of plumes under various environmental conditions. Each video segment meticulously records parameters such as the corresponding DIVE mission number, recording time, temperature, heading, bottom contact distance, and depth, providing multi-dimensional reference for the comprehensive analysis of plume movement characteristics.

[0035] A method for retrieving deep-sea hydrothermal plume motion characteristics and temperature field based on video analysis, with the overall flowchart as follows: Figure 12 As shown.

[0036] Step S1: The input video image data is segmented into frames to obtain a video frame sequence, followed by preprocessing such as image grayscale conversion, noise smoothing, threshold segmentation, and mask control. First, the frame sequence is processed to grayscale, and the result is as follows: Figure 1 As shown in (a) above, the calculation is simplified and the brightness characteristics are highlighted; then, Gaussian filtering is used for noise smoothing, and the result is as follows. Figure 1 As shown in (b) above, this method aims to suppress the effects of abnormal water disturbance and uneven illumination. The aforementioned image preprocessing provides clear and stable input images for subsequent analysis and calculations.

[0037] Step S2: Select four positioning endpoints to determine the mask, which defines the analysis area and the inner and outer dual-layer responsibility zones. The outer layer is mainly responsible for identifying newly emerging particles, while the inner layer is only responsible for tracking particles with existing historical trajectory information and performing remedial search and tracking after tracking loss occurs.

[0038] After completing the masking settings and reference scale calibration, adjust the preset parameters such as the start time point, maximum velocity threshold, and particle detection threshold. You can also manually input a more accurate pixel ratio. After clicking the run button, the waiting log area will display "Particle tracking complete," and you can preview screenshots of the particle tracking result video on this page. The average velocity-time variation trend of particle tracking is shown below. Figure 2 As shown.

[0039] Detailed calculation process as follows Figure 3 As shown. After completing the mask setup and reference scale calibration, a two-layer responsibility region is first created, with the inner and outer layers undertaking different processing tasks. The outer layer is mainly responsible for detecting newly appearing particles: first, the particle trajectory set is initialized, and then candidate particle detection is performed on the outer coverage area of ​​the image frame; all particles detected in the first frame have new trajectories created. From the second frame onwards, the MNN method is used to match the detected candidate particles with existing historical trajectories. If a match is successful, the current candidate particle is merged into the existing trajectory; if a match fails, it is determined to be a newly appearing particle, and a new trajectory is created for it. Subsequently, the trajectory set is updated uniformly, and the next frame is read for iterative processing.

[0040] The inner layer is only responsible for tracking existing particle trajectories and performing remedial processing after tracking loss. First, it reads the historical trajectory information of candidate particles entering the inner layer. Each trajectory is accompanied by a confidence score, initially set to 0.6. Each particle entering the inner layer is initially considered a reliable particle and will not disappear into the inner layer region; it will only be temporarily lost due to environmental factors such as brightness changes or occlusion. Therefore, the inner layer sets up three detection functions: Local, Rescue, and Force, which detect candidate points sequentially. If Local tracking fails, Rescue and Force perform remedial tracking in sequence; if any one of the three detections succeeds, the confidence score of the current trajectory increases by 0.12, with a maximum value of 1.00. If all three fail, the confidence score decreases, and the trajectory that cannot be detected temporarily due to environmental factors is suspended until the next frame. The first four consecutive frames of loss are the hold period, during which the confidence score decreases by 0.063 per frame; after the hold period, the confidence score decreases by 0.18 for each failed detection frame. When the confidence level drops below 0.05 or the number of consecutive lost frames reaches 40, the lifecycle of the current particle trajectory is terminated.

[0041] The remedial search and tracing performed by the inner responsibility area includes the following three-level detection strategy: Level 1 Detection: Based on the particle's historical trajectory information, the current frame position is predicted, and a first search window is set around the predicted position to detect particles. The first search window has a radius of 30 pixels and a size of 61×61 pixels. Particles with a grayscale difference of 12 between adjacent frames and a current frame brightness of 110 are detected.

[0042] If the first-level detection fails, the second-level detection is performed: the detection conditions are adjusted so that the grayscale difference between adjacent frames reaches 10 and the brightness of the current frame reaches 99, and a fan-shaped search area with an effective radius of 120 pixels and a half angle of 45° is constructed based on historical trajectory information. Particle detection is performed within the fan-shaped search area with a range larger than the first search window.

[0043] If the second-level detection fails, the third-level detection is executed: with the predicted position as the center, the search window is expanded from a radius of 30 pixels to a radius of 80 pixels relative to the first search window, that is, a search window with a size of 161×161 pixels is set, and the detection conditions are further adjusted to be that the gray level difference between adjacent frames reaches 9 and the brightness of the current frame reaches 90. The candidate point is selected as the matching result based on a comprehensive judgment of the distance between the candidate point and the predicted position, the distance with the previous frame position, the consistency of the motion direction, and the reverse motion penalty.

[0044] Local is the first-level detection layer within the inner region. It predicts the position of particles in the current frame based on the previous frame's position and velocity from historical trajectory information, and generates a small circular window at the predicted position for particle detection. Its search range is relatively small, suitable for particles with relatively stable motion and minimal changes in motion direction and brightness. Rescue is a remedial strategy after Local failures. It appropriately relaxes the detection threshold and constructs a fan-shaped region based on historical trajectory information to search for candidate particles with a detection range larger than Local. Force, after Rescue failures, establishes an expanded search window centered on the trajectory prediction position and appropriately relaxes candidate detection thresholds such as frame difference and brightness to improve the probability of trajectory retrieval after weak signals or short-term occlusion. To avoid mismatches, Force does not directly accept arbitrary candidate points. Instead, it comprehensively judges the distance between the candidate point and the predicted position, the distance from the previous frame position, the consistency of the motion direction, and the reverse motion penalty, ultimately selecting a reasonable candidate point as the matching result.

[0045] After detection, the particle trajectory set is updated uniformly and the next frame is processed. After all frames are processed, all particle information is statistically analyzed and the results are output; simultaneously, the analysis area is divided into grids, and the particle samples within each grid are statistically averaged to obtain... Figure 4 The two-dimensional flow field results are shown. The velocity variations of some high-quality particles are as follows: Figure 6 As shown, Figure 6 The velocity change trends of the four particles indicate that the present invention can continuously track particles.

[0046] Step S3: The user performs reference scale calibration. The user can drag two reference points to mark the coordinates of the two ends of the reference object. Determine the pixel length of the reference object in the image, and input the actual length of the reference object. The pixel ratio s is calculated as shown in equation (1): (1).

[0047] Users use the adjustment bar to set the attitude angle, such as... Figure 5 As shown in the diagram, the absolute depth estimation process is as follows: Figure 7 As shown.

[0048] This gives us the rotation angles in the X, Y, and Z directions. This represents the rotation angle of the reference plane about the X-axis of the camera coordinate system. This represents the rotation angle of the reference plane about the Y-axis of the camera coordinate system. This represents the rotation angle of the reference plane about the Z-axis of the camera coordinate system. It allows calculation of the tilt state of the reference object in three-dimensional space and the construction of a rotation matrix. The system is set so that all three rotation angles can only be integers, and , The range is from -90 degrees to 90 degrees. The range is from -180 degrees to 180 degrees, representing the attitude transformation relationship from the local coordinate system of the reference object to the camera coordinate system, as shown in the following formula: ; ; ; (2).

[0049] The program defaults to the reference point ( Figure 5 The major axis of the reference plane is initially along the Y-axis, therefore the local direction vector of the reference object is... (Used to describe the orientation of the reference object without rotation) is: (3).

[0050] By rotation matrix Transform the direction vector to the camera coordinate system (that is, rotate the reference plane around the X, Y, and Z axes respectively). As shown below: (4).

[0051] The result after rotation is shown below: (5).

[0052] in, , , These represent the components of the reference object's major axis in the X, Y, and Z directions of the camera coordinate system (each value ranges from -1 to 1). After the attitude angle rotation, the reference object's major axis no longer only follows the original local Y-axis direction, but also has projection components in all three directions in the camera coordinate system.

[0053] This does not represent the actual position or length of the reference object, but rather the unit direction vector of the reference object's major axis in the camera coordinate system after attitude rotation. Because the original direction vector... For a unit vector, the rotation matrix only changes its direction and not its length, therefore the rotated vector... It remains a unit vector of length 1.

[0054] (6).

[0055] In the system definition, the camera's shooting direction is defined as the X-axis, and the image imaging plane corresponds to the YZ plane. Therefore... Actually perpendicular to the imaging plane, and The component represents the component of the reference object's major axis within the camera's imaging plane. To calculate the visible proportion of the reference object's major axis in the image plane, it is necessary to convert the three-dimensional direction vector... Projected onto the YZ plane. The length of the projection vector. This can be obtained from the Pythagorean theorem: (7).

[0056] Since the original direction vector has a length of 1, the projection length itself is equal to the projection scale. As shown in the formula below: (8).

[0057] It indicates the proportion of the reference object along its long axis that falls within the camera's imaging plane.

[0058] use The true length of the reference object Multiplying them together gives the projected length of the reference object. As shown in the formula below: (9).

[0059] Using the pinhole imaging principle in conjunction with the user-input camera focal length User-defined pixel length The absolute distance to the midpoint of the reference object can be obtained through calculation. As shown in the formula below: (10).

[0060] By fusing this result with the relative depth map output by the Depth Anything V2 model, the absolute depth map of the target can be obtained, as shown in the following figure. Figure 8As shown, this allows us to obtain the actual depth information corresponding to each location within the study area, providing a depth reference for the spatial scale quantification and subsequent analysis of the study area.

[0061] Step S4: The user presets the nozzle type to either diffuse or focused, based on the results of empirical inversion, as follows: Figure 9 , Figure 10 As shown.

[0062] Figure 9 and Figure 10 The English labels in the lower right corner represent the foreign authors' names of the corresponding source documents, used to distinguish data groups from different sources; the year is the publication year of the corresponding document, and n is the number of samples used. For example, Sarrazin 1999 (n=14) means that 14 data samples from Sarrazin's document published in 1999 were used.

[0063] After the user confirms the parameters, an inversion calculation is performed based on the output flow velocity field results, and the final temperature field results are as follows: Figure 11 As shown.

[0064] Figure 11 The figure shows a two-dimensional temperature field distribution map based on velocity field inversion. The horizontal and vertical axes represent the spatial location of the image. The results clearly show the spatial distribution, temperature gradient changes and diffusion range of the high-temperature region in the core of the hydrothermal plume. This demonstrates that the present invention can effectively obtain temperature distribution information of the target area based on video images without the need to deploy sensors.

[0065] In summary, this invention constructs a complete video analysis framework for deep-sea hydrothermal plumes based on deep-sea scientific expedition video samples. This framework uses real-world deep-sea video as its core input and processes data through three core steps: particle motion tracking, spatial depth estimation, and velocity-temperature empirical inversion. It can extract multi-dimensional observational information such as hydrothermal plume motion characteristics, spatial depth structure, and temperature distribution from video data without the need for additional complex sensing equipment.

[0066] In terms of output results, the improved PTV particle velocity measurement module can output complete particle trajectories, local velocity statistics, and two-dimensional average flow field results within the analysis area, providing quantitative support for the analysis of hydrothermal plume motion characteristics; the depth estimation module can generate a depth map with actual physical distance reference significance based on the monocular relative depth estimation results, combined with reference object scale and attitude angle correction, providing assistance for the spatial structure analysis of the target area; the velocity-temperature empirical inversion module further couples the velocity field output results, converts the velocity information into temperature estimation results, and finally obtains the two-dimensional temperature field distribution of the target area.

[0067] This invention integrates video particle velocity measurement, monocular depth estimation, and empirical temperature inversion into a unified analysis process. Only the original video and a few necessary calibration parameters are required to obtain comprehensive analysis results such as velocity fields, depth reference maps, and temperature fields, providing data support for identifying the motion patterns of deep-sea hydrothermal plumes, analyzing their spatial structure, inverting temperature fields, and conducting subsequent numerical simulation studies.

[0068] The specific embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not constitute a limitation on the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made without creative effort based on the technical solutions of the present invention should fall within the scope of protection of the present invention.

Claims

1. A method for retrieving motion characteristics and temperature field of deep-sea hydrothermal plumes based on video analysis, characterized in that, Includes the following steps: Step S1: Acquire deep-sea hydrothermal plume video image data, perform frame segmentation on the video image data to obtain a video frame sequence, and preprocess the video frame sequence to obtain a preprocessed video frame sequence. Step S2: In response to the selected multiple positioning endpoints, a custom mask is generated, which determines the analysis area and the two-layer responsibility partition of the outer responsibility area and the inner responsibility area; wherein, the outer responsibility area is used to detect newly emerging particles and record their short-term motion trajectory, and the inner responsibility area is used to track particles with existing historical trajectory information and perform remedial search tracking after tracking loss occurs. The particle tracking velocimetry method is used to identify and track particles in the preprocessed video frame sequence to obtain particle trajectory information. Based on the particle trajectory information, the analysis area is divided into grids, and the particles passing through each grid are counted and averaged over time to obtain the two-dimensional average flow field. The remedial search and tracking performed by the inner responsibility region includes the following three-level detection strategy: Level 1 detection: Predict the current frame position based on the previous frame position and velocity information of the particle's historical trajectory, and set up a first search window around the predicted position to detect particles; If the first-level detection fails, the second-level detection is performed: the frame difference and brightness detection thresholds are adjusted, and a fan-shaped search area is constructed based on historical trajectory information. Particle detection is performed within the fan-shaped search area in a range larger than the first search window. If the second-level detection fails, the third-level detection is performed: the search window is adjusted with the predicted position as the center, and the frame difference and brightness detection thresholds are adjusted to detect particles. The candidate point is selected as the matching result based on a comprehensive judgment of the distance between the candidate point and the predicted position, the distance between the candidate point and the position of the previous frame, the consistency of the motion direction, and the reverse motion penalty. Step S3: In response to the scale calibration operation of the reference object, obtain the pixel length and true length of the reference object, and calculate the pixel ratio; use a monocular depth estimation model to perform relative depth estimation on the images in the video frame sequence to obtain a relative depth map; in response to the set three-dimensional pose angle of the reference object, combine the pixel ratio and the camera focal length to perform absolute scale transformation on the relative depth map to generate an absolute depth reference map with actual distance scale. Step S4: In response to the preset target nozzle type, substitute the two-dimensional average flow field obtained in step S2 into the velocity-temperature empirical inversion model corresponding to the target nozzle type to obtain the two-dimensional temperature field distribution of the target region; the velocity-temperature empirical inversion model includes a power function model for a dispersion nozzle and an arctangent function model for a focusing nozzle.

2. The method according to claim 1, characterized in that, The preprocessing described in step S1 includes at least one of image grayscale conversion, Gaussian filtering noise smoothing, threshold segmentation, and mask control.

3. The method according to claim 1, characterized in that, The inner responsibility area sets a confidence level for each particle trajectory, with the initial confidence level being a preset value. When any level of detection is successful, the confidence level is increased; when all detections fail, the confidence level is decreased; when the confidence level is lower than a first threshold or the number of consecutive lost frames exceeds a second threshold, the lifecycle of the particle trajectory is terminated.

4. The method according to claim 1, characterized in that, The absolute scale transformation mentioned in step S3 specifically includes: Based on the principle of pinhole imaging, the ideal reference distance is calculated using the camera focal length, the true length of the reference object, and the pixel length of the reference object in the image. A rotation matrix is ​​constructed based on the set three-dimensional attitude angle of the reference object. The projection ratio of the major axis of the reference object in the camera image plane is calculated. The ideal reference distance is then corrected using the projection ratio to obtain the corrected reference distance. Using the corrected reference distance as a conversion benchmark, the relative depth map is scaled to obtain the absolute depth reference map.

5. The method according to claim 4, characterized in that, The absolute scale transformation is performed according to the following formula: In response to the scale calibration operation on the reference object, according to the formula Calculate pixel ratio ,in The actual length of the reference object; x 1, y 1) and ( x 2, y 2) These are the pixel coordinates of the two ends of the reference object selected in the image. x 1. y 1 represents the x-coordinate and y-coordinate of the first positioning point in the image coordinate system, respectively. x 2. y 2 are the x and y coordinates of the second positioning point in the image coordinate system, respectively, and the Euclidean distance calculated from the coordinates of the two positioning points is the reference pixel length; Set the angles of rotation of the reference plane around the X, Y, and Z axes. , , ,according to , , Construct rotation matrix ,in, This represents the rotation angle of the reference plane about the X-axis of the camera coordinate system. This represents the rotation angle of the reference plane about the Y-axis of the camera coordinate system. R represents the rotation angle of the reference plane around the Z-axis of the camera coordinate system, where R is the rotation angle according to the preset rotation sequence. , , The resulting attitude rotation matrix represents the magnitude of the object's rotation. The major axis of the reference object is assumed to be the Y-axis, therefore its local direction vector is... Through the attitude rotation matrix Transform the direction vector to the camera coordinate system , , , These represent the components of the major axis of the reference object in the X, Y, and Z directions of the camera coordinate system, respectively. According to the formula Calculate the projection ratio of the major axis of the reference object onto the camera's imaging plane. The projected length of the reference object in the y-direction of the camera's imaging plane. ; According to the formula Calculate the absolute distance of the midpoint of the reference object. Where f is the camera focal length, The pixel length between the two ends of the reference object; The absolute distance from the midpoint of the reference object Using the relative depth map as a conversion benchmark, a scale transformation is performed to obtain an absolute depth reference map.

6. The method according to claim 1, characterized in that, The method for constructing the flow velocity-temperature empirical inversion model in step S4 includes: Construct a temperature-velocity dataset, which includes nozzle temperature, plume velocity, and nozzle morphology parameters; Based on the nozzle morphology, temperature characteristics, and jet flow state, hydrothermal nozzles are classified into two categories: dispersive nozzles and focusing nozzles. The temperature-velocity relationship of the two types of nozzles was fitted separately: a power function model was used for the dispersion nozzle, and an arctangent function model was used for the focusing nozzle.

7. The method according to claim 1, characterized in that, The custom mask mentioned in step S2 is determined by selecting four positioning endpoints. The analysis area is defined by the mask, and the outer responsibility area and the inner responsibility area are divided by the inner and outer boundaries of the mask.

Images