A method and system for underwater scene depth assessment based on monocular vision
The feature points of underwater scenes are extracted by a monocular camera and the depth estimation model is used to solve the problem of hardware equipment reliance on underwater three-dimensional scene reconstruction in the prior art, achieving low-cost and efficient underwater depth evaluation.
Patent Information
- Application Number
- CN202210252491.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-15
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2042-03-15
AI Technical Summary
Existing underwater three-dimensional scene reconstruction methods rely on hardware equipment, especially binocular cameras and depth cameras, are costly and sensitive to light, and are limited in applications in underwater environments, making it difficult to achieve efficient depth assessment.
The monocular camera is used to evaluate the depth of underwater scenes. By acquiring and preprocessing the image data, the Moravec operator is used to extract feature points, and the depth judgment is made based on the depth estimation model, and the matching of the perspective pose parameters and scene image data of the monocular camera is established. Quasi-relationship, generate visual depth images.
It realizes the depth judgment and evaluation of underwater dynamic scenes automatically without considering the impact of camera baseline accuracy, which reduces the calculation amount and improves the application prospect of monocular cameras in underwater scene depth evaluation.
Smart Images

Figure CN114648567B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to computer vision technology, and in particular to a method and system for evaluating the depth of an underwater scene based on monocular vision. Background Art
[0002] With the continuous exploration of oceans and other waters, underwater operations, marine scientific surveys and other activities are also increasing. In order to improve the efficiency of underwater operations and achieve the purpose of marine scientific surveys, underwater operation equipment is usually used to assist operations. Due to the existence of various complex underwater scene environments, in order to meet the needs of task execution and improve people's perception of underwater scenes, it is necessary to use computer vision and image processing technologies to evaluate the position and posture of underwater operation equipment in unknown environments and reconstruct the three-dimensional structure of the scene, so as to more intuitively complete the observation of underwater scenes and complete underwater operations under visual guidance.
[0003] As computer vision and image processing technologies continue to evolve, at this stage, in the fields of computer vision and robotics, motion structure recovery (SFM) and simultaneous localization and mapping (SLAM) technologies have become research hotspots. At present, in order to achieve the reconstruction of underwater three-dimensional scenes, SLAM methods are usually based on depth cameras or binocular cameras to achieve dense reconstruction of three-dimensional scenes. However, the shortcomings of such methods are obvious, and they are overly dependent on hardware equipment, resulting in limited practical application scenarios. In particular, the binocular system using binocular cameras requires the use of a stereo camera to collect two images at the same time at each moment, which leads to high costs and the need to consider the impact of the camera baseline on the system accuracy, which is not conducive to the depth assessment of underwater scenes.
[0004] SLAM methods based on depth cameras or binocular cameras can achieve dense reconstruction of three-dimensional scenes, but the disadvantages of such methods are obvious. They are overly dependent on hardware devices, which limits the actual application scenarios. In particular, depth cameras have a limited working range and are sensitive to light, so they are only suitable for close-range indoor environments. Therefore, SLAM research based on monocular cameras has more application prospects.
[0005] The current mainstream visual odometry can be roughly divided into two categories based on hardware: monocular systems that use only one camera and binocular systems that use stereo cameras. Monocular refers to using only one camera as the only sensor, while binocular refers to using a stereo camera to collect two images at the same time at each moment. Compared with binocular, monocular visual odometry has the advantages of low cost and no need to consider the impact of camera baseline on system accuracy. Summary of the invention
[0006] Purpose of the invention: The purpose of the present invention is to provide a method and system for underwater scene depth assessment based on monocular vision, so as to complete underwater scene depth assessment without considering the influence of camera baseline on system accuracy.
[0007] Technical solution: The method for evaluating the depth of an underwater scene based on monocular vision described in the present invention comprises the following steps:
[0008] (1) Acquiring sensor data in an underwater dynamic scene, wherein the sensor data is scene image data captured by a monocular camera at different viewing angles;
[0009] (2) performing feature extraction on the video frames in the scene image data, extracting feature points in the scene image data at different viewing angles, and obtaining feature extraction results of objects captured in each video frame;
[0010] (3) According to the change of the camera's viewing angle position, the depth of the object is judged based on the feature extraction result of the object captured based on the depth estimation model to obtain the visual depth value of the object captured;
[0011] (4) Based on the different viewing angles and posture parameters of the camera, the scene image data is deeply processed so that the visual depth values of the object surfaces of the feature points in the scene image data are optimized to obtain a scene depth estimation image.
[0012] The scene image data described in step (1) is the image data of the scene captured by a monocular camera mounted on an underwater device or an underwater robot entering an underwater dynamic scene and photographed at different rotation angles.
[0013] Each video frame image in the scene image data described in step (1) corresponds to a viewing angle of the monocular camera, and feature extraction is performed on the video frames in the scene image data. When extracting feature points in the scene image data of different viewing angles, the acquired sensor data is preprocessed, and feature point detection is performed on the preprocessed scene image to extract feature points in the scene image data of each viewing angle.
[0014] The method for calculating the interest value of each pixel in the scene image data described in step (2) is specifically as follows:
[0015] (2.1) Obtaining scene image data of multiple viewpoints and postures acquired by a monocular camera, and gray-scaling the acquired scene image data;
[0016] (2.2) The Moravec operator is used to calculate the interest value of each pixel in the scene image data: taking each pixel as the center of the target pixel, the grayscale variance of each pixel and its neighboring pixels is calculated within a pixel window of a set size of n×n, and the pixels whose grayscale variance is greater than the set variance threshold are taken as feature points with obvious grayscale changes; the Moravec operator is used to calculate the grayscale variance of the neighboring pixels around each pixel in the scene image data, and the pixels whose grayscale variance meets the set variance threshold are selected as feature points of the acquired scene image data; wherein, the neighboring pixels around each pixel include the neighboring pixels of the target pixel in the horizontal, vertical and two diagonal directions.
[0017] (2.3) Pixels with interest values greater than a given empirical threshold are taken as candidate points. Within a pixel window of a set size, pixels with the largest interest value among the candidate points are taken as feature points of the window area to obtain feature points with obvious grayscale changes in the scene image data.
[0018] The training method of the depth estimation model described in step (3) is specifically as follows:
[0019] (3.1) Using a monocular camera to capture scene image data of underwater dynamic scenes;
[0020] (3.2) aligning the acquired scene image data with the view angle and pose parameters of the monocular camera to obtain a sample set of the view angle position change of the monocular camera;
[0021] (3.3) Extract feature points from the registered sample set to obtain a training set of feature extraction results and changes in the position of the monocular camera’s viewing angle;
[0022] (3.4) The depth estimation model is trained using the training set to obtain a monocular camera depth estimation model suitable for underwater dynamic scenes. The depth estimation model is used to perform depth judgment on the feature extraction results to directly obtain the visual depth value of the current scene image data.
[0023] The method for obtaining the scene depth estimation image described in step (3.4) is specifically as follows:
[0024] (3.4.1) based on the visual depth value corresponding to each pixel in the scene image data initially acquired through the registration, the visual depth value is mapped to a corresponding position in the three-channel color RGB image corresponding to the initial scene image data according to the pixel corresponding to the feature point;
[0025] (3.4.2) Using the visual depth value of each pixel in the initial scene image data, determine the visual depth value of the feature point object surface in the initial scene image data, and use the image with the determined visual depth value as the scene depth estimation image.
[0026] A monocular vision-based underwater scene depth assessment system, the system uses the above-mentioned monocular vision-based underwater scene depth assessment method to complete underwater scene depth assessment; the system includes:
[0027] Acquisition unit: used to acquire scene image data captured by a monocular camera at different viewing angles in an underwater dynamic scene;
[0028] Feature extraction unit: used to extract feature points from scene image data of different viewing angles, and obtain feature extraction results of objects captured in each video frame;
[0029] Depth judgment unit: used to perform depth judgment on the feature extraction results of the photographed object based on the depth estimation model according to the change of the camera's viewing angle position, and obtain the visual depth value of the photographed object;
[0030] Depth evaluation unit: used to perform depth processing on scene image data based on different camera view and posture parameters, so as to optimize the visual depth value of the object surface of feature points in the scene image data and obtain a scene depth estimation image.
[0031] The system further comprises:
[0032] Registration unit: registers the acquired scene image data with the view angle and pose parameters of the monocular camera, obtains a sample set of the view angle position change of the monocular camera, and establishes the corresponding relationship between the scene image data and the view angle and pose parameters of the monocular camera.
[0033] A computer storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned underwater scene depth assessment method based on monocular vision.
[0034] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned monocular vision-based underwater scene depth assessment method when executing the computer program.
[0035] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0036] 1. The present invention can automatically use a monocular camera to complete depth judgment and depth assessment of underwater dynamic scenes, establish a registration relationship between different viewing angles and posture parameters of the monocular camera and scene image data, perform depth judgment based on a depth estimation model to obtain a visual depth value, and provide a guarantee for realizing the depth assessment of underwater scenes with a monocular camera;
[0037] 2. The present invention extracts depth information of feature points of scene image data based on the change of viewing angle of a monocular camera, which greatly saves the amount of calculation and reduces the restrictions on the monocular camera for underwater scene depth assessment. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 This is a flow chart of a method for underwater scene depth assessment based on monocular vision;
[0039] Figure 2 A flowchart for generating a 3D pipeline model in a monocular vision-based underwater scene depth assessment method;
[0040] Figure 3 A flow chart of obtaining a scene depth estimation image in a method for underwater scene depth assessment based on monocular vision;
[0041] Figure 4 This is a system block diagram of an underwater scene depth assessment system based on monocular vision described in Example 1;
[0042] Figure 5 This is a system block diagram of an underwater scene depth assessment system based on monocular vision described in Example 2. DETAILED DESCRIPTION
[0043] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.
[0044] Embodiment 1:
[0045] See also Figure 1 As shown, Figure 1 A flowchart of a method for evaluating the depth of an underwater scene based on monocular vision provided by the present invention. An embodiment of the present invention provides a method for evaluating the depth of an underwater scene based on monocular vision, comprising the following steps:
[0046] S1: Acquire sensor data in an underwater dynamic scene, where the sensor data is scene image data captured by a monocular camera at different viewing angles.
[0047] In this embodiment, the scene image data is image data captured by a monocular camera mounted on an underwater device or an underwater robot entering an underwater dynamic scene and photographing the scene at different rotation angles.
[0048] Before the monocular camera takes pictures and collects data, it also includes establishing a global coordinate system for the underwater motion scene. The coordinate system is established in a gridding manner, and a coordinate linear relationship is established between pixel points imaged in the monocular camera image at the intersection of the grids, and the scene image parameters of the monocular camera in the global coordinate system of the underwater motion scene are set.
[0049] When the monocular camera is shooting and collecting, the motion trajectory of the monocular camera is calibrated according to the posture parameters of the monocular camera at different rotation angles, and the scene image data captured by the monocular camera during the rotational motion is obtained, and the image data of the sensor sequence captured is obtained.
[0050] In an embodiment of the present invention, each video frame image of the video queue in the scene image data corresponds to a viewing angle of the monocular camera.
[0051] S2: performing feature extraction on the video frames in the scene image data, extracting feature points in the scene image data of different viewing angles, and obtaining feature extraction results of the objects captured in each video frame.
[0052] It should be particularly noted that when feature extraction is performed on the video frames in the scene image data and feature points are extracted from the scene image data of different perspectives, the acquired sensor data is preprocessed, feature point detection is performed on the preprocessed scene images, and feature points in the scene image data of each perspective are extracted.
[0053] In the specific implementation of step S2 provided in the embodiment of the present invention, see Figure 2 As shown, the method for extracting feature points from scene image data of different viewing angles includes:
[0054] S201, acquiring scene image data of multiple viewpoints and postures acquired by a monocular camera, and performing grayscale processing on the acquired scene image data;
[0055] S202, using the Moravec operator to calculate the interest value of each pixel in the scene image data;
[0056] S203, taking pixels whose interest values are greater than a given empirical threshold as candidate points, and taking pixels with the largest interest value among the candidate points as feature points of the region within a pixel window of a set size, to obtain feature points with obvious grayscale changes in the scene image data.
[0057] The method for calculating the interest value of each pixel in the scene image data is as follows: taking each pixel as the center of the target pixel, calculating the grayscale variance of each pixel and its neighboring pixels within a pixel window of a set size of n×n, and taking the pixel points whose grayscale variance is greater than the set variance threshold as feature points with obvious grayscale changes.
[0058] When the Moravec operator is used to extract feature points in scene image data, the Moravec operator calculates the grayscale variance of neighboring pixels around each pixel in the scene image data, and selects the pixel point whose grayscale variance meets the set variance threshold as the feature point of the acquired scene image data;
[0059] The neighboring pixels around each pixel include the neighboring pixels of the target pixel in the horizontal, vertical, and two diagonal directions.
[0060] When performing the extraction operation, the Moravec operator calculates the grayscale variance of the grayscale changes between each pixel in the scene image to be processed and its adjacent pixels in the directions of 0°, 45°, 90° and 135°, that is, the grayscale variance of each pixel in the scene image is calculated from four directions: horizontal, vertical and two diagonal lines, and the pixel points whose grayscale variance meets the variance threshold are selected as the feature points of the scene image to be processed.
[0061] The interest value of the pixel point is represented by the minimum grayscale variance of each pixel point in four directions, and the point with the maximum interest value is selected as the feature point with obvious grayscale change in the local area of the scene image to be processed.
[0062] When calculating the interest value of each pixel, the pixel to be calculated is taken as the target pixel, the target pixel is taken as the center, and the image window of a set size is taken as the window area. The sum of the squares of similar grayscale differences of the target pixel in the four directions of horizontal, vertical, and two diagonal lines in the window area centered on the target pixel is calculated, and the smallest value is taken as the interest value of the target pixel.
[0063] Then, according to the given variance threshold, pixels larger than the variance threshold are taken as candidate points, and the candidate points with the largest interest value in the window area are taken as feature points of the window area. In this way, the detected feature points have a lower repetition rate, and the feature point recognition process is simple and fast.
[0064] S3: According to the change of the camera's viewing angle position, a depth judgment is performed on the feature extraction result of the photographed object based on the depth estimation model to obtain a visual depth value of the photographed object.
[0065] In this embodiment, the training method of the depth estimation model is specifically as follows:
[0066] Using a monocular camera to capture and collect scene image data in underwater dynamic scenes;
[0067] The acquired scene image data is registered with the view angle and pose parameters of the monocular camera to obtain a sample set of the view angle position change of the monocular camera;
[0068] Extract feature points from the registered sample set to obtain a training set of feature extraction results and changes in the position of the monocular camera's viewing angle;
[0069] The depth estimation model is trained using the training set to obtain a monocular camera depth estimation model suitable for underwater dynamic scenes. The depth estimation model is used to perform depth judgment on the feature extraction result to directly obtain the visual depth value of the current scene image data.
[0070] In this embodiment, since the scene image data initially acquired is a sequence of images captured, based on the feature point identification of each sequence of images, the position information of the captured object in the corresponding sequence of images is obtained, and two or more images containing the feature point in a continuous sequence are extracted to obtain the visual depth value of the captured object.
[0071] When obtaining the position information of the captured object of the feature point in the corresponding sequence of images, each sequence of images is also transformed by Hough to obtain the position information of the feature point in the image.
[0072] In an embodiment of the present invention, the scene image data is aligned with the viewing angle and posture parameters of the monocular camera. When different viewing angle and posture parameters are obtained, the video frames captured by the monocular camera are extracted from the scene images with different viewing angle and posture parameters and a training set is established to train a depth estimation model. When the scene image data at a specific viewing angle position is input into the depth estimation model, the visual depth value of the captured object can be obtained, and a depth estimation model based on the viewing angle and posture of the monocular camera is established.
[0073] S4: Performing depth processing on the scene image data based on different camera view and posture parameters to optimize the visual depth value of the object surface of the feature points in the scene image data and obtain a scene depth estimation image.
[0074] Figure 3 A flow chart of obtaining a scene depth estimation image in a method for underwater scene depth assessment based on monocular vision provided by an embodiment of the present invention is shown.
[0075] It should be noted that the specific implementation method of step S4 provided in the embodiment of the present invention is:
[0076] S401, based on the visual depth value corresponding to each pixel in the scene image data initially obtained through registration, map the visual depth value to a corresponding position in a three-channel color RGB image corresponding to the initial scene image data according to the pixel corresponding to the feature point.
[0077] S402 , using the visual depth value of each pixel in the initial scene image data, determining the visual depth value of the object plane of the feature point in the initial scene image data, and using the image with the determined visual depth value as the scene depth estimation image.
[0078] Specifically, in one embodiment, a method for evaluating the depth of an underwater scene based on monocular vision is as follows:
[0079] The monocular camera is installed on an underwater device or an underwater robot to obtain the shooting position and shooting direction under different viewing angle and posture parameters. The underwater device or underwater robot drives the monocular camera into the underwater dynamic scene. When the shooting position is determined, when shooting with different viewing angles and postures, the monocular camera takes pictures at a certain rotation angle and frame rate, and the scene image data of continuous video frames arranged according to the shooting angle and time sequence can be obtained. Feature extraction is performed on the scene image data video frame, feature points are extracted in each video frame, and the feature extraction results of the object captured in each video frame are obtained. According to the set rotation trajectory of the monocular camera and the corresponding relationship between the monocular camera image established in the global coordinate system, two or more images containing the feature points in the continuous sequence are extracted to obtain the visual depth value of the captured object, and the trained depth estimation model is used to perform depth judgment on the feature extraction results of the captured object to obtain the visual depth value of the captured object.
[0080] Finally, in the global coordinate system, according to the registration relationship between the camera's different viewing angles and posture parameters and the scene image data established above, the visual depth value is mapped to the corresponding position in the three-channel color RGB image corresponding to the initial scene image data, and the visual depth value of the object surface of the feature point in the initial scene image data is determined by using the visual depth value of each pixel in the initial scene image data, and the image with the determined visual depth value is used as the scene depth estimation image.
[0081] Embodiment 2:
[0082] like Figure 4 As shown, in a preferred embodiment of the present invention, a monocular vision-based underwater scene depth assessment system includes an acquisition unit 100, a feature extraction unit 200, a depth judgment unit 300 and a depth assessment unit 400. Among them:
[0083] The acquisition unit 100 is used to acquire scene image data captured by a monocular camera at different viewing angles in an underwater dynamic scene.
[0084] In this embodiment, each video frame image in the scene image data captured by the acquisition unit 100 corresponds to a viewing angle of the monocular camera, and the image data in the scene is captured at different rotation viewing angles.
[0085] The feature extraction unit 200 is used to extract feature points from scene image data of different viewing angles to obtain feature extraction results of objects captured in each video frame.
[0086] In this embodiment, the feature extraction unit 200 uses the grayscale variance of adjacent pixels in four directions, namely horizontally, vertically, and in two diagonals, of each pixel in the scene image data of the Moravec operator, and selects pixels whose grayscale variance meets the variance threshold as feature points of the scene image to be processed. The detected feature points have the advantages of low repetition rate and simple and fast recognition process.
[0087] The depth judgment unit 300 is used to perform depth judgment on the feature extraction results of the photographed object based on the depth estimation model according to the change of the camera's viewing angle position, and obtain the visual depth value of the photographed object. In this embodiment, in the training stage of the depth judgment unit 300, after aligning the acquired scene image data with the viewing angle posture parameters of the monocular camera and extracting feature points, a training set of feature extraction results and the change of the viewing angle position of the monocular camera is obtained. The training set containing the viewing angle posture parameters is used to train the depth estimation model. After the viewing angle scene image captured by the monocular camera is input into the depth estimation model, the visual depth value of the current scene image data can be directly obtained, saving a lot of calculation processes.
[0088] The depth evaluation unit 400 is used to perform depth processing on the scene image data based on different camera view and posture parameters, so as to optimize the visual depth value of the object surface of the feature points in the scene image data and obtain a scene depth estimation image.
[0089] In this embodiment, the depth assessment unit 400 maps the visual depth value corresponding to each pixel in the scene image data initially acquired through the registration to the corresponding position in the three-channel color RGB image corresponding to the initial scene image data according to the pixel corresponding to the feature point, and determines the visual depth value of the object plane of the feature point in the initial scene image data by using the visual depth value of each pixel in the initial scene image data, and uses the image with the determined visual depth value as the scene depth estimation image.
[0090] Figure 5 The structural block diagram of the underwater scene depth assessment system based on monocular vision provided by another embodiment of the present invention is shown. The underwater scene depth assessment system based on monocular vision also includes a registration unit 500, which registers the acquired scene image data with the view angle posture parameters of the monocular camera, obtains a sample set of the view angle position change of the monocular camera, and establishes a corresponding relationship between the scene image data and the view angle posture parameters of the monocular camera.
[0091] In this embodiment, the underwater scene depth assessment system based on monocular vision adopts the steps of the underwater scene depth assessment method based on monocular vision as mentioned above during execution. Therefore, the operation process of the underwater scene depth assessment system based on monocular vision will not be introduced in detail in this embodiment.
Claims
1. A method for underwater scene depth assessment based on monocular vision, It is characterized in that The following steps are involved: (1) Acquiring sensor data in an underwater dynamic scene, wherein the sensor data is scene image data captured by a monocular camera at different viewing angles; (2) performing feature extraction on the video frames in the scene image data, extracting feature points in the scene image data from different perspectives, and obtaining feature extraction results of the objects captured in each video frame; The method for extracting feature points from scene image data of different viewing angles comprises: (2.0.1) Obtain scene image data of multiple viewpoints and postures acquired by a monocular camera, and grayscale the acquired scene image data; (2.0.2) Use the Moravec operator to calculate the interest value of each pixel in the scene image data; (2.0.3) Pixels with interest values greater than a given empirical threshold are taken as candidate points, and within a pixel point window of a set size, pixels with the largest interest value among the candidate points are taken as feature points of the area, thereby obtaining feature points with obvious grayscale changes in the scene image data; (3) According to the change of the camera's viewing angle position, a depth judgment is performed on the feature extraction result of the photographed object based on the depth estimation model to obtain a visual depth value of the photographed object; the training method of the depth estimation model is specifically as follows: (3.1) Using a monocular camera to capture scene image data of underwater dynamic scenes; (3.2) Aligning the acquired scene image data with the view angle and pose parameters of the monocular camera to obtain a sample set of the view angle position change of the monocular camera; (3.3) Extract feature points from the registered sample set to obtain a training set of feature extraction results and changes in the position of the monocular camera’s viewing angle; (3.4) Using the training set to train a depth estimation model, obtaining a monocular camera depth estimation model suitable for underwater dynamic scenes, using the depth estimation model to perform depth judgment on the feature extraction results, and directly obtaining a visual depth value of the current scene image data; (4) performing depth processing on the scene image data based on different camera view and posture parameters, so as to optimize the visual depth value of the feature point object surface in the scene image data, and obtain a scene depth estimation image; the method for obtaining the scene depth estimation image is specifically as follows: (4.1) based on the visual depth value corresponding to each pixel in the scene image data initially acquired through the registration, mapping the visual depth value to a corresponding position in the three-channel color RGB image corresponding to the initial scene image data according to the pixel corresponding to the feature point; (4.2) Using the visual depth value of each pixel in the initial scene image data, determine the visual depth value of the feature point object surface in the initial scene image data, and use the image with the determined visual depth value as the scene depth estimation image.
2. The method for underwater scene depth assessment based on monocular vision according to claim 1, It is characterized in that The scene image data described in step (1) is the image data of the scene captured by a monocular camera mounted on an underwater device or an underwater robot entering an underwater dynamic scene and photographed at different rotation angles.
3. The method for underwater scene depth assessment based on monocular vision according to claim 1, It is characterized in that Each video frame image in the scene image data described in step (1) corresponds to a viewing angle of the monocular camera, and feature extraction is performed on the video frames in the scene image data. When extracting feature points in the scene image data of different viewing angles, the acquired sensor data is preprocessed, and feature point detection is performed on the preprocessed scene image to extract feature points in the scene image data of each viewing angle.
4. The method for underwater scene depth assessment based on monocular vision according to claim 1, It is characterized in that The method for calculating the interest value of each pixel in the scene image data in step (2) is specifically as follows: (2.1) Obtaining scene image data of multiple viewpoints and postures acquired by a monocular camera, and gray-scaling the acquired scene image data; (2.2) Use the Moravec operator to calculate the interest value of each pixel in the scene image data: take each pixel as the center of the target pixel, calculate the grayscale variance of each pixel and its neighboring pixels in a pixel point window of a set size of n×n, and take the pixel points whose grayscale variance is greater than the set variance threshold as the feature points with obvious grayscale changes; The Moravec operator calculates the grayscale variance of the neighboring pixels around each pixel in the scene image data, and selects the pixel points whose grayscale variance meets the set variance threshold as the feature points of the acquired scene image data; wherein, the neighboring pixels around each pixel include the neighboring pixels of the target pixel in the horizontal, vertical and two diagonal directions; (2.3) Pixels with interest values greater than a given empirical threshold are taken as candidate points. Within a pixel window of a set size, pixels with the largest interest value among the candidate points are taken as feature points of the window area to obtain feature points with obvious grayscale changes in the scene image data.
5. An underwater scene depth assessment system based on monocular vision, It is characterized in that The system uses the underwater scene depth assessment method based on monocular vision as described in any one of claims 1 to 4 to complete the underwater scene depth assessment; the system includes: Acquisition unit: used to acquire scene image data captured by a monocular camera at different viewing angles in an underwater dynamic scene; Feature extraction unit: used to extract feature points from scene image data of different viewing angles, and obtain feature extraction results of objects captured in each video frame; Depth judgment unit: used to perform depth judgment on the feature extraction results of the photographed object based on the depth estimation model according to the change of the camera's viewing angle position, and obtain the visual depth value of the photographed object; Depth evaluation unit: used to perform depth processing on scene image data based on different camera view and posture parameters, so as to optimize the visual depth value of the object surface of feature points in the scene image data and obtain a scene depth estimation image.
6. The underwater scene depth assessment system based on monocular vision according to claim 5, It is characterized in that The system further comprises: Registration unit: registers the acquired scene image data with the view angle and pose parameters of the monocular camera, obtains a sample set of the view angle position change of the monocular camera, and establishes the corresponding relationship between the scene image data and the view angle and pose parameters of the monocular camera.
7. A computer storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the method for underwater scene depth assessment based on monocular vision as described in any one of claims 1 to 4 is implemented.
8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the computer program, the method for underwater scene depth assessment based on monocular vision as described in any one of claims 1 to 4 is implemented.
