A Method and System for Self-Assessing the Credibility of Digital Images Based on Feature Point Prediction Consistency Test

Abstract

A method and system for self-evaluation of digital image reliability based on feature point prediction consistency testing, relating to the field of sensor reliability self-evaluation. It addresses the problem that traditional reliability evaluation methods cannot directly determine the reliability of raw digital image data for navigation tasks. The method includes: receiving a digital image at the current moment from the sensor; extracting feature point information from the digital image at the current moment; calculating the potential positions of key feature points at the next moment based on the feature point information in the current digital image; and comparing the potential positions of the key feature points at the next moment with the actual digital image acquired at the next moment to complete the reliability self-evaluation. This invention is applied in the field of machine vision.

Description

Technical Field

[0001] This invention relates to the field of sensor reliability self-assessment, and in particular to a digital image reliability self-assessment method based on feature point prediction consistency test. Background Technology

[0002] Sensor reliability self-assessment refers to the process by which sensor devices evaluate the reliability of the data and information they provide in various applications, especially in situations involving critical tasks and decision-making.

[0003] Currently, sensor devices are constrained by their physical characteristics and technological limitations. These limitations may include measurement range, accuracy, resolution, and sensitivity. Regardless of self-evaluation, sensors cannot overcome these fundamental sensing limitations. Furthermore, sensor performance is time-dependent, environmental, and usage-dependent. Sensors require regular calibration and maintenance to ensure their performance. However, these processes can lead to sensor instability, resulting in data quality issues that self-evaluation often fails to capture. Sensor reliability is closely related to the environment, with different environments potentially leading to varying performance. Self-evaluation struggles to cover all possible environmental conditions, thus making it impossible to accurately assess reliability in certain situations.

[0004] Existing sensor reliability self-assessment typically involves detecting outliers, noise, and other data quality issues. However, some problems can be difficult to detect, especially when they interact with other factors in complex ways, leading to erroneous self-assessment results. Self-assessing sensor reliability often requires significant computational and data processing. This can be impractical for embedded systems or resource-constrained environments and may result in latency or resource bottlenecks.

[0005] Furthermore, most existing sensor data evaluation methods are geared towards numerical sensors. In terms of the reliability evaluation of digital image sensor data, they are still limited to applications such as image quality assessment and ambiguity estimation, with very few digital image reliability evaluation methods for navigation and positioning applications.

[0006] Traditional reliability assessment methods for navigation systems based on digital images often estimate the navigation results directly. However, navigation results are affected by various factors, such as the fusion of data from other sensors or the use of novel filtering methods. Therefore, traditional reliability assessment methods cannot directly determine the reliability of the original digital image data for the navigation task. Summary of the Invention

[0007] This invention addresses the problem that traditional reliability assessment methods cannot directly determine the reliability of raw digital image data for navigation tasks. It proposes a self-assessment method for digital image reliability based on feature point prediction consistency testing. The method includes:

[0008] Based on the digital image received by the sensor at the current moment;

[0009] Extract feature point information from the digital image at the current moment;

[0010] Calculate the potential location of key feature points at the next moment based on the feature point information in the current digital image;

[0011] The credibility self-assessment is completed by comparing the potential location of the key feature points at the next moment with the actual digital image acquired at the next moment.

[0012] Furthermore, a preferred method is also provided, wherein the extraction of feature point information from the digital image at the current moment specifically involves:

[0013] The point cloud information of the digital image at the current moment is extracted using the point cloud rotation invariant network PRIN;

[0014] The extracted point cloud information is rotation-invariant feature.

[0015] Furthermore, a preferred method is also provided, wherein the method for extracting point cloud information as rotation-invariant point features is as follows:

[0016] In the absence of input data augmentation point cloud information, sparse points of the point cloud information are used as input;

[0017] Density-aware adaptive sampling is used to convert sparse points into spherical voxel signals.

[0018] The spherical voxel signal is passed through a spherical voxel convolutional layer to generate a feature on each spherical voxel grid, and rotation-invariant point features are extracted by point resampling.

[0019] The rotation-invariant point features are output through a fully connected layer.

[0020] Furthermore, a preferred embodiment is also provided, wherein calculating the potential location of key feature points at the next moment based on feature point information in the current digital image includes:

[0021] Based on the camera pose of the previous frame Obtain the transformation matrix of the world coordinate system relative to the camera coordinate system and the transformation matrix of the camera coordinate system relative to the world coordinate system for the camera pose in the previous frame;

[0022] The current frame is obtained based on the uniformly accelerated motion of the camera pose. transformation matrix and Relative relationships between the transformation matrices of frames ;

[0023] According to the current frame transformation matrix and Relative relationships between the transformation matrices of frames The camera pose of the current frame is estimated using the transformation matrix of the camera coordinate system relative to the world coordinate system. ;

[0024] Based on the estimated current frame pose Corresponding to the stereo feature points of this frame World coordinates, which obtain the camera's 3D coordinates relative to the current frame;

[0025] Calculate the potential location of key feature points at the next moment based on the stereo coordinates of the current frame.

[0026] Furthermore, a preferred embodiment is also provided, wherein the method further includes generating the same data continuously at different sampling times when the sensor malfunctions.

[0027] Furthermore, a preferred embodiment is also provided, wherein the method further includes using the RANSAC sampling algorithm to select 50 feature points in the digital image at the current moment to predict the position of 50 feature points at the next moment, and comparing them with the digital image at the next moment acquired by the sensor.

[0028] Furthermore, a preferred method is provided, wherein comparing the digital image acquired by the sensor at the next moment includes: the image actually acquired at the next moment... , Compare within the specified range.

[0029] Based on the same inventive concept, this invention also proposes a digital image credibility self-assessment system based on feature point prediction consistency verification, the system comprising:

[0030] The current moment image acquisition unit is used to receive the digital image at the current moment from the sensor;

[0031] The feature point extraction unit is used to extract feature point information from the digital image at the current moment;

[0032] The calculation unit is used to calculate the potential location of key feature points at the next moment based on the feature point information in the current digital image;

[0033] The evaluation unit is used to compare the potential location of the key feature points at the next moment with the actual digital image acquired at the next moment to complete the self-evaluation of credibility.

[0034] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium, characterized in that the computer-readable storage medium is used to store a computer program, the computer program executing the digital image credibility self-assessment method based on feature point prediction consistency test as described above.

[0035] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the digital image credibility self-assessment method based on feature point prediction consistency test as described in any one of the preceding claims.

[0036] The advantages of this invention are:

[0037] This invention solves the problem that traditional credibility assessment methods cannot directly determine the credibility of raw digital image data for navigation tasks.

[0038] The digital image reliability self-assessment method based on feature point prediction consistency verification described in this invention introduces feature point information into sensor reliability self-assessment, enabling it to detect problems that are difficult to find using traditional methods. Traditional methods primarily focus on outliers and noise, failing to directly determine the reliability of the original digital image data for navigation tasks. This invention addresses the issue of detecting the reliability of the original digital image data for navigation tasks using feature point information, improving the comprehensiveness of reliability assessment for navigation tasks. By calculating the latent position based on feature points and comparing it with the actual image, this method can more accurately assess the reliability of image data. This helps improve data reliability, especially in applications with stringent requirements for high-quality image data, such as medical imaging or autonomous driving. The method described in this invention allows for real-time reliability self-assessment because it uses image data from the current and next time moments, along with feature point information. This means that problems can be detected and addressed promptly, thereby improving system robustness. The method described in this invention is not only applicable to specific fields such as computer vision or autonomous driving but can also be used in various digital image acquisition and sensor applications. Whether in industrial applications, the medical field, or scientific research, all can benefit from the reliability self-assessment capabilities of this method.

[0039] The method described in this invention improves the reliability of digital image sensor data, ensuring the accuracy and reliability of acquired image data. This is crucial for many applications, including computer vision, machine learning, autonomous driving, and medical imaging, as these fields heavily rely on high-quality image data. Automated reliability assessment reduces the impact of erroneous data, improving system robustness and reliability. Attached Figure Description

[0040] Figure 1 The flowchart below shows the digital image credibility self-assessment method based on feature point prediction consistency test as described in Implementation Method 1.

[0041] Figure 2 This is a schematic diagram of the point cloud rotation invariant network described in Implementation Method 3;

[0042] Figure 3 This is a schematic diagram of the RANSAC sampling algorithm described in Implementation Method Eleven. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0044] Implementation Method 1, see [link] Figure 1 This embodiment describes a digital image reliability self-assessment method based on feature point prediction consistency testing. The method includes:

[0045] Based on the digital image received by the sensor at the current moment;

[0046] Extract feature point information from the digital image at the current moment;

[0047] Calculate the potential location of key feature points at the next moment based on the feature point information in the current digital image;

[0048] The credibility self-assessment is completed by comparing the potential location of the key feature points at the next moment with the actual digital image acquired at the next moment.

[0049] The digital image reliability self-assessment method based on feature point prediction consistency verification described in this embodiment introduces feature point information into sensor reliability self-assessment, enabling it to detect problems that are difficult to find using traditional methods. Traditional methods primarily focus on outliers and noise, failing to directly determine the reliability of the original digital image data for navigation tasks. This invention addresses the issue of detecting the reliability of original digital image data for navigation tasks using feature point information, improving the comprehensiveness of reliability assessment for navigation tasks. By calculating the potential location based on feature points and comparing it with the actual image, this method can more accurately assess the reliability of image data. This helps improve data reliability, especially in applications with stringent requirements for high-quality image data, such as medical imaging or autonomous driving. The method described in this embodiment allows for real-time reliability self-assessment because it uses image data from the current and next time moments, along with feature point information. This means that problems can be detected and addressed promptly, thereby improving system robustness. The method described in this embodiment is not only applicable to specific fields such as computer vision or autonomous driving but can also be used in various digital image acquisition and sensor applications. Whether in industrial applications, the medical field, or scientific research, all can benefit from the reliability self-assessment capabilities of this method.

[0050] The method described in this embodiment improves the reliability of digital image sensor data, ensuring that the acquired image data is accurate and reliable. This is crucial for many applications, including computer vision, machine learning, autonomous driving, and medical imaging, as these fields heavily rely on high-quality image data. Automated reliability assessment reduces the impact of erroneous data, improving the robustness and reliability of the system.

[0051] Implementation Method Two: This implementation method further defines the digital image credibility self-evaluation method based on feature point prediction consistency verification described in Implementation Method One. Specifically, extracting feature point information from the digital image at the current moment involves:

[0052] The point cloud information of the digital image at the current moment is extracted using the point cloud rotation invariant network PRIN;

[0053] The extracted point cloud information is rotation-invariant feature.

[0054] This implementation utilizes a point cloud rotation-invariant network, enabling the extraction of rotationally invariant features. This means that even image data acquired at different rotation angles retains consistent extracted features, thereby increasing the stability of data reliability self-evaluation. Point cloud information typically contains high-level image features, such as object shape, texture, and structure. Therefore, using point cloud information as feature point information can provide more useful information about image content, helping to more accurately assess the reliability of image data. This method is applicable to various image types, including 3D point cloud data, and is thus suitable for fields requiring spatial information processing, such as machine vision, machine learning, and automation.

[0055] This implementation increases the accuracy and stability of digital image reliability self-evaluation, especially when processing 3D point cloud data. Extraction of rotation-invariant features ensures that image data retains similar characteristics under rotation, transformation, or other changes, which is crucial for various applications. By using point cloud information as feature point information, the features and structures in the image can be better described and evaluated.

[0056] This implementation uses a rotation-invariant point cloud network (PRIN) to process the current digital image, converting it into point cloud data. PRIN is a neural network specifically designed for extracting point cloud data; it is rotation-invariant, ensuring that the point cloud data remains unchanged under rotation. PRIN extracts rotation-invariant features from the point cloud data. These features can include the shape, density, and color information of the point cloud, and are unaffected by viewpoint rotation. The extracted rotation-invariant features are used as feature point information. This information is used for subsequent reliability self-evaluation to detect data anomalies or problems.

[0057] Implementation Method 3, see below Figure 2 This embodiment describes a further limitation on the digital image reliability self-assessment method based on feature point prediction consistency verification described in Embodiment 2. The method for extracting rotation-invariant point cloud information is as follows:

[0058] In the absence of input data augmentation point cloud information, sparse points of the point cloud information are used as input;

[0059] Density-aware adaptive sampling is used to convert sparse points into spherical voxel signals.

[0060] The spherical voxel signal is passed through a spherical voxel convolutional layer to generate a feature on each spherical voxel grid, and rotation-invariant point features are extracted by point resampling.

[0061] The rotation-invariant point features are output through a fully connected layer.

[0062] The method described in this embodiment can extract point features with rotation invariance. Even when point cloud information is acquired at different rotation angles, the extracted features remain consistent, thereby increasing the stability of data reliability self-evaluation. This method is applicable to point cloud information without input data augmentation. This makes this method usable in various situations without additional data processing steps. High-level features of point cloud information can be effectively extracted through spherical voxel signals and spherical voxel convolutional layers. These features can include important information such as the shape and density of the point cloud.

[0063] This implementation extracts rotation-invariant point features for subsequent digital image credibility self-evaluation. Extracting rotation-invariant point features ensures that point cloud data retains similar characteristics under rotation, transformation, or other changes, which is crucial for various applications. Specifically, the process includes: First, using sparse points from the point cloud as input. This means using only a subset of points in the point cloud for feature extraction, rather than all the point cloud data. Then, density-aware adaptive sampling converts the sparse points into spherical voxel signals. This step helps preserve important information in the point cloud while reducing the impact of noise. Next, a spherical voxel convolutional layer is applied to each spherical voxel grid to generate a feature. This step helps extract higher-level features from the spherical voxel signals. Finally, point resampling extracts rotation-invariant point features. This step ensures that the extracted features remain consistent even at different rotation angles. Finally, a fully connected layer outputs the rotation-invariant point features for subsequent credibility self-evaluation.

[0064] Implementation Method Four: This implementation method further defines the digital image reliability self-assessment method based on feature point prediction consistency verification described in Implementation Method One. The step of calculating the potential location of key feature points at the next moment based on feature point information in the current digital image includes:

[0065] Based on the camera pose of the previous frame Obtain the transformation matrix of the world coordinate system relative to the camera coordinate system and the transformation matrix of the camera coordinate system relative to the world coordinate system for the camera pose in the previous frame;

[0066] The current frame is obtained based on the uniformly accelerated motion of the camera pose. transformation matrix and Relative relationships between the transformation matrices of frames ;

[0067] According to the current frame transformation matrix and Relative relationships between the transformation matrices of frames The camera pose of the current frame is estimated using the transformation matrix of the camera coordinate system relative to the world coordinate system. ;

[0068] Based on the estimated current frame pose Corresponding to the stereo feature points of this frame World coordinates, which obtain the camera's 3D coordinates relative to the current frame;

[0069] Calculate the potential location of key feature points at the next moment based on the stereo coordinates of the current frame.

[0070] This implementation estimates the potential locations of key feature points in the next time step based on the camera pose and uniform acceleration motion model of the previous frame through continuous prediction. This can help improve the robustness of image processing algorithms because it allows the system to make reasonable predictions between consecutive frames. The method described in this implementation allows for real-time calculation of the locations of key feature points in the next time step because it is based on information from the previous frame and a model of camera motion. This is very useful for applications requiring rapid response. By estimating the camera pose and the world coordinates of stereo feature points, problems caused by feature point tracking errors can be reduced. This helps improve the accuracy of digital image reliability self-assessment.

[0071] This implementation calculates the potential locations of key feature points at the next moment based on information from the previous frame and a model of camera motion. The purpose of this is to provide more information for self-assessment of digital image reliability while reducing potential problems with feature point tracking errors.

[0072] This implementation first obtains the transformation matrix of the world coordinate system relative to the camera coordinate system and the transformation matrix of the camera coordinate system relative to the world coordinate system, based on the camera pose of the previous frame. Through uniformly accelerated motion of the camera pose, the relative relationship between the transformation matrix of the current frame and the transformation matrix of the previous frame is obtained. This takes into account the camera's acceleration to more accurately estimate changes in camera pose. Based on the relative relationship between the previous and current frames and the camera pose transformation matrix of the previous frame, the camera pose of the current frame is estimated. This includes position and pose information. Based on the estimated current frame pose and the world coordinates corresponding to the stereo feature points in that frame, the stereo coordinates of the camera relative to the current frame are calculated. This provides the three-dimensional coordinates of each key feature point in the current frame. Finally, based on the stereo coordinates of the current frame, the potential positions of the key feature points at the next moment can be calculated, taking into account changes in camera pose. This implementation is based on continuous prediction, using information from the previous frame and a model of camera motion to estimate the potential positions of key feature points at the next moment. This can provide more accurate feature point positions for digital image reliability self-evaluation and reduce potential problems with feature tracking errors.

[0073] Implementation Method 5: This implementation method further defines the digital image reliability self-assessment method based on feature point prediction consistency test described in Implementation Method 1. The method also includes generating the same data continuously at different sampling times when the sensor malfunctions.

[0074] In practical applications, sensors may encounter malfunctions, noise, or interference, leading to a deterioration in the quality of digital image acquisition data. By continuously generating the same data at different sampling times, these problems can be identified and corrected, thereby improving the robustness of the self-assessment of reliability. This implementation takes into account the issue of continuously generating the same data, ensuring the consistency of image data across different time points. By comparing the continuously generated data, it is possible to detect whether the sensor has malfunctioned. If the data from multiple sampling times are consistent, the sensor may be functioning normally; however, if the data differs significantly, it may indicate a sensor malfunction or abnormal condition.

[0075] When a sensor is functioning normally, the continuously generated data should be similar because they capture the same scene. The principle is that under normal circumstances, this data should be highly consistent, thus allowing for the prediction and comparison of feature point locations. When a sensor malfunctions, data consistency can be affected, as the malfunction may lead to a decrease in data quality or the introduction of errors. By comparing continuously generated data, this inconsistency can be detected, alerting the user or automatically taking corrective measures, such as re-estimating the location of feature points based on multiple samples to improve the accuracy of the reliability self-assessment.

[0076] Implementation Method Six: This implementation method further defines the self-evaluation method for digital image reliability based on feature point prediction consistency test described in Implementation Method One. The method further includes using the RANSAC sampling algorithm to select 50 feature points in the digital image at the current time to predict the positions of 50 feature points at the next time, and comparing them with the digital image at the next time acquired by the sensor.

[0077] This implementation enhances the robustness of digital image reliability self-assessment by using RANSAC. By selecting only a subset of feature points for prediction and comparison, computational complexity is reduced, making the system more efficient. Selecting 50 feature points instead of all feature points reduces computational resource consumption while maintaining the accuracy of reliability self-assessment. The RANSAC algorithm selects 50 feature points at the current time step, and then uses these feature points to predict the positions of 50 feature points at the next time step. The purpose of this step is to establish a correlation between the current and next time step feature points to achieve continuous feature tracking. The predicted feature point positions are compared with the actual feature points in the digital image acquired by the sensor at the next time step. The purpose of this comparison is to verify whether the predicted feature points are consistent with the actual image data, thereby assessing the reliability of the digital image data.

[0078] Implementation Method Seven: This implementation method further defines the digital image reliability self-assessment method based on feature point prediction consistency verification described in Implementation Method Six. The comparison with the digital image acquired by the sensor at the next moment includes: the actual image acquired at the next moment... , Compare within the specified range.

[0079] in, Predict the width coordinates for feature points. This is the width retrieval threshold. Predict the height coordinates for feature points. This is a high-level retrieval threshold. The range set in this implementation is closer to the actual situation, enhancing robustness.

[0080] Implementation Method 8: The digital image credibility self-assessment system based on feature point prediction consistency test described in this implementation method includes:

[0081] The current moment image acquisition unit is used to receive the digital image at the current moment from the sensor;

[0082] The feature point extraction unit is used to extract feature point information from the digital image at the current moment;

[0083] The calculation unit is used to calculate the potential location of key feature points at the next moment based on the feature point information in the current digital image;

[0084] The evaluation unit is used to compare the potential location of the key feature points at the next moment with the actual digital image acquired at the next moment to complete the self-evaluation of credibility.

[0085] Implementation Method Nine: A computer-readable storage medium according to this implementation method, the computer-readable storage medium being used to store a computer program, the computer program executing the digital image credibility self-assessment method based on feature point prediction consistency test as described in any one of Implementation Methods One to Seven.

[0086] Implementation Method 10: A computer device according to this implementation method includes a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the digital image credibility self-assessment method based on feature point prediction consistency test according to any one of Implementation Methods 1 to 7.

[0087] Implementation Method Eleven: This implementation method provides a specific example of the digital image credibility self-assessment method based on feature point prediction consistency test described in Implementation Method One, and also serves to explain Implementation Methods Two to Seven. Specifically:

[0088] For the current frame The camera pose of the previous frame It is known. The transformation matrix of the world coordinate system relative to the camera coordinate system in the previous frame is set as follows: , and The transformation matrix of the camera coordinate system relative to the world coordinate system is: , and The world coordinate system can be obtained. Compared to Transformation of the world coordinate system:

[0089]

[0090]

[0091] Based on the uniformly accelerated motion model, the current frame can be considered as... The transformation matrix relative to The frames have the following relationship:

[0092]

[0093] The camera pose of the current frame can be estimated from the above:

[0094]

[0095] Based on the estimated current frame pose Corresponding to the stereo feature points of this frame World coordinates This allows us to obtain the camera's 3D coordinates relative to the current frame. :

[0096]

[0097]

[0098] It is a transformation matrix The rotated part, the mapped coordinates of the feature points As shown below:

[0099]

[0100] in, It is the camera's focal length. , These are the coordinates of the center pixel of the image.

[0101] Calculate the potential location of key feature points at the next moment based on the feature point information in the current digital image, and compare it with the actual situation to conduct a reliability analysis.

[0102] For point cloud extraction, a Point-wise Rotation Invariant Network (PRIN) is used to extract rotation-invariant features from the point cloud. The network structure is as follows: Figure 2 As shown, without input data augmentation of point cloud information, the network takes sparse points as input and uses density-aware adaptive sampling to transform the signal into a spherical voxel grid. The spherical voxel signal is passed through a spherical voxel convolutional layer, generating a feature on each spherical voxel grid. Features of any point can be extracted through point resampling.

[0103] In practical applications, the RANSAC sampling algorithm is used to select 50 feature points for prediction. The algorithm principle is as follows: Figure 3 As shown. In the actual acquisition of images , Retrieve feature points within the range. If more than 10 feature points fail to match, then... Add 1. Generally speaking, and Take 1 / 10 of the image width and height. Additionally, when the number of feature points in the image is less than 20, or when the preceding and following images are completely identical, Add 1.

[0104] Calculate the probability of anomalies in sensor data by combining the above two scenarios:

[0105]

[0106] in, for Real-time credibility , As a preset constant, The number of feature point matching failures. This represents the number of times there are insufficient feature points.

[0107] Although preferred embodiments of this disclosure have been described, those skilled in the art, upon learning the basic inventive concept, can make further changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of this disclosure. Clearly, those skilled in the art can make various alterations and variations to this disclosure without departing from its spirit and scope. Thus, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

[0108] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0109] The technical solutions provided by the present invention have been described in further detail above with reference to the accompanying drawings in order to highlight their advantages and benefits, and are not intended to limit the present invention. Any modifications, combinations, improvements and equivalent substitutions of the present invention based on the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for self-evaluation of the reliability of a digital image based on the consistency check of the prediction of characteristic points, characterized in that, The method includes: The digital image at the current moment is obtained from the sensor; Extract feature point information from the digital image at the current moment; Calculate the potential location of key feature points at the next moment based on the feature point information in the digital image at the current moment; The credibility self-assessment is completed by comparing the potential location of the key feature points at the next moment with the actual digital image acquired at the next moment. The method further includes using the RANSAC sampling algorithm to select 50 feature points in the digital image at the current moment to predict the position of 50 feature points at the next moment, and comparing them with the digital image at the next moment acquired by the sensor; The next time digital image acquired by the sensor is compared, including: comparing the next time actual acquisition image in the range of , ​ If more than 10 feature points fail to match, then Add 1; and Take 1 / 10 of the width and height of the image; in addition, when the number of feature points in the image is less than 20, or the front and back images are completely consistent, Add 1; Calculate the probability of anomalies in sensor data by combining the above two scenarios: in, for Real-time credibility , As a preset constant, The number of feature point matching failures. This represents the number of times there are insufficient feature points.

2. The digital image credibility self-assessment method based on feature point prediction consistency test according to claim 1, characterized in that, The extraction of feature point information from the digital image at the current moment specifically involves: The point cloud information of the digital image at the current moment is extracted using the point cloud rotation invariant network PRIN; The extracted point cloud information is rotation-invariant feature.

3. The digital image credibility self-assessment method based on feature point prediction consistency test according to claim 2, characterized in that, The method for extracting rotation-invariant point cloud information is as follows: In the absence of input data augmentation point cloud information, sparse points of the point cloud information are used as input; Density-aware adaptive sampling is used to convert sparse points into spherical voxel signals. The spherical voxel signal is passed through a spherical voxel convolutional layer to generate a feature on each spherical voxel grid, and rotation-invariant point features are extracted by point resampling. The rotation-invariant point features are output through a fully connected layer.

4. The digital image credibility self-assessment method based on feature point prediction consistency test according to claim 1, characterized in that, The step of calculating the potential location of key feature points at the next moment based on feature point information in the digital image at the current moment includes: Based on the camera pose of the previous frame Obtain the transformation matrix of the world coordinate system relative to the camera coordinate system and the transformation matrix of the camera coordinate system relative to the world coordinate system for the camera pose in the previous frame; The current frame is obtained based on the uniformly accelerated motion of the camera pose. transformation matrix and Relative relationships between the transformation matrices of frames ; According to the current frame transformation matrix and Relative relationships between the transformation matrices of frames The camera pose of the current frame is estimated using the transformation matrix of the camera coordinate system relative to the world coordinate system. ; Based on the estimated current frame pose Corresponding to the stereo feature points of this frame World coordinates, which obtain the camera's 3D coordinates relative to the current frame; Calculate the potential location of key feature points at the next moment based on the stereo coordinates of the current frame.

5. The digital image credibility self-assessment method based on feature point prediction consistency test according to claim 1, characterized in that, The method also includes generating the same data continuously at different sampling times when the sensor malfunctions.

6. A digital image credibility self-assessment system based on feature point prediction consistency verification, characterized in that, The system is implemented based on the method of claim 1, and the system includes: The current moment image acquisition unit is used to acquire a digital image of the current moment based on the sensor. The feature point extraction unit is used to extract feature point information from the digital image at the current moment; The calculation unit is used to calculate the potential location of key feature points at the next moment based on the feature point information in the digital image at the current moment; The evaluation unit is used to compare the potential location of the key feature points at the next moment with the actual digital image acquired at the next moment to complete the self-evaluation of credibility.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program that executes the digital image credibility self-assessment method based on feature point prediction consistency test as described in any one of claims 1-5.

8. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the digital image credibility self-assessment method based on feature point prediction consistency test as described in any one of claims 1-5.

Images