Method for filtering a video stream

Abstract

The invention relates to a method for filtering a video stream by discarding a second frame which temporally follows a first frame in the video stream, the method comprising: - detecting a local gradient in the first frame, - detecting the local gradient from the first frame in the second frame, and - discarding one of the two frames if a temporal gradient between the local gradient in the first frame and the local gradient in the second frame satisfies a predetermined condition.

Description

[0001] Method for filtering a video stream

[0002] The present invention relates to a method for filtering a video stream, a control device for carrying out the method and a system for interpolating a 3D model from a temporal sequence of 2D images using the control device.

[0003] WO 2016 / 005412Al describes a photogrammetric method for creating a 3D model from a video stream based on a temporal sequence of 2D images. This method selects images that exhibit significant overlap and offer different viewing angles. These criteria ensure, firstly, that there are enough common features for photogrammetric analysis and, secondly, that there is sufficient distance between the image acquisition points. This is crucial for achieving clear differences in the viewing angles. However, the precise algorithmic implementation of this image selection is not explained in detail.

[0004] The purpose of the invention is to improve monitoring.

[0005] The problem is solved by the features of the independent claims. Preferred embodiments are the subject of the dependent claims. According to one aspect of the invention, a method for filtering a video stream by discarding a second frame following a first frame in the video stream comprises the steps of detecting a spatial gradient in the first frame, detecting the spatial gradient from the first frame in the second frame, and discarding one of the two frames if a temporal gradient between the spatial gradient in the first frame and the spatial gradient in the second frame satisfies a predetermined condition.

[0006] The described method is based on the analysis of spatial and temporal gradients in successive frames. This approach aims to identify and remove frames that do not provide significant new information compared to the previous frame, which is important for creating an effective 3D model using photogrammetry.

[0007] First, local gradients, such as edges or similar structures, are detected in each frame. These local gradients represent significant features or changes in the scene, such as the outlines of objects. Next, it is checked whether these gradients are present in the following frame and how they have changed over time. If the temporal gradient—that is, the change in these features over time—meets a predetermined condition, one of the frames is discarded. For example, a frame might be discarded if the change in the gradients is below a certain threshold, indicating that the scene has not changed significantly.

[0008] The advantages of this method lie primarily in increased efficiency and reduced redundancy in the video footage. By removing frames that offer no additional or only minor new information, the amount of data required to create the 3D model is reduced. This can lead to faster processing and a reduction in the required computing resources. Furthermore, the quality of the resulting 3D model can be improved, as the focus is placed on frames with significant changes, enabling a more accurate and detailed reconstruction. Although the described method is motivated by an application in photogrammetry, its application is not limited to photogrammetry.

[0009] In a further development of the specified method, the predetermined condition is adjustable. Adjustable in this context refers to the fact that the predetermined condition used to filter the video streams is variable. This means that the parameters of this condition can either be adjusted manually by a user or automatically modified by mechanisms such as filters, neural networks, or the like.

[0010] The ability to manually adjust this condition allows users to tailor the filtering to their specific needs. For example, users can increase or decrease the filter's sensitivity to respond to different movement patterns or lighting conditions. This makes the process highly customizable and versatile.

[0011] Automatic adaptability offers the advantage of real-time adjustment. In dynamic environments, the predetermined condition can be automatically adjusted to achieve optimal results. This can help reduce unwanted artifacts and increase the efficiency of video stream filtering.

[0012] In addition to adaptability, scalability also offers the advantage of increased efficiency. By optimizing the conditions, the process can utilize resources more efficiently and improve filtering accuracy. This leads to overall improved performance and broader applicability of the process in various situations.

[0013] Overall, the ability to divide the predetermined condition allows for flexible adaptation to different requirements and environments, which significantly increases the effectiveness and applicability of the procedure.

[0014] In a preferred refinement of the described method, the predetermined condition is a threshold that must be exceeded to discard one of the two frames. This achieves several advantages. If the predetermined condition has a low threshold, it means that only frames that differ significantly from each other are retained. This leads to a substantial reduction in redundancy in the video stream. In scenarios where many consecutive frames change only slightly, this can help optimize storage space and data transmission. Conversely, discarding frames that change only slightly improves the quality of the remaining video stream. The remaining frames are likely to be more relevant and contain more significant information.This is particularly important in applications such as video compression or video streaming, where high video quality is desired. In applications that aim to analyze motion or changes in videos, low thresholds can also help ensure that only relevant movements are captured. This increases the efficiency of subsequent analysis procedures and reduces processing time. In photogrammetry, as explained earlier, sufficient dissimilarity between frames is also crucial for depth perception. A low threshold ensures that the selected frames exhibit significant differences in viewing angles, leading to more accurate 3D reconstructions. Depending on the requirements and the specific characteristics of the videos, users can adjust the threshold. This allows the method to be optimized and flexibly deployed for various scenarios and applications.Overall, using a threshold as a predetermined condition helps to filter the video stream more efficiently and in a more targeted way, which is advantageous in various application areas, be it video compression, motion detection or 3D reconstruction.

[0015] In a further development of the described method, the frames for detecting the local gradients are transformed into a feature space with at least one feature variable. In image processing, a feature space is a mathematical space in which each image or frame is represented by features or properties. These features can be derived from the pixel values ​​of the image and serve to extract and characterize information about the image. Feature spaces can take various forms, including Fourier space, HOG space (Histogram of Oriented Gradients), color space (e.g., RGB or HSV), and many others, depending on the specific requirements of the image processing task. This approach offers several advantages. For one, feature spaces represent complex information in a more compact form.This leads to a reduction in the dimensionality of the data, making processing and analysis more efficient. Selecting a suitable feature space also allows for the highlighting of important visual properties of an image and the minimization of unimportant details. This can help reduce noise and improve the robustness of algorithms. A well-chosen feature space can make the features of an image more robust against various transformations such as rotation, scaling, and changes in lighting. This is crucial for the reliability of image processing systems. Patterns and structures can be more easily recognized in feature spaces. This is particularly important in applications such as object detection, face recognition, or texture recognition. Certain image processing tasks, such as edge or texture detection, can be performed more efficiently in feature spaces.This can reduce computing power and memory requirements. The selection of the right feature space depends on the specific application and the tasks to be solved. Each feature space has its own advantages and disadvantages, and choosing the optimal space often requires experience and experimentation. Overall, however, the use of feature spaces enables a powerful and flexible method for image analysis and processing.

[0016] In a further development of the described method, the feature variable of the feature space is a variable describing a pattern, specifically a texture variable of a texture space as the feature space. The terms "pattern variable," "pattern space," "texture space," and "texture variable" are important concepts in image processing and pattern recognition and are defined as follows: A pattern variable is an abstract representation of a visual pattern or structure in an image. It can be used to describe certain characteristic properties of a pattern, such as shapes, textures, colors, or other visual features. Pattern variables are mathematical representations that allow patterns to be analyzed and compared. A pattern space is the space in which pattern variables exist. This space can take various forms, depending on the patterns to be described.In a pattern space, pattern variables are used to represent and analyze patterns. For example, a pattern space for faces might contain all possible variations of facial features. A texture space is a special type of pattern space that focuses on representing textures in images. Textures are patterns defined by the arrangement of structures or pixels in the image. A texture space therefore contains pattern variables specifically designed to characterize textures. A texture variable is a specific pattern variable in a texture space. It describes a particular property or characteristic of a texture in an image. Examples of texture variables might be the density of the grain in a textured image or the direction of the repeating patterns in a fabric pattern.

[0017] Using a texture space with texture variables offers several advantages. First, texture variables in a texture space allow for a more precise and detailed description of textures in images. This enables finer differentiation between various textures. A texture space with texture variables facilitates the analysis of textures in 2D images. This is particularly important in applications such as quality control, material identification, and medical imaging. Furthermore, texture variables allow for the classification of textures into different categories. This is helpful in automating tasks such as surface defect detection or material identification. In certain applications where textures play a crucial role, such as...In satellite image analysis or geology, the use of texture variables can increase the accuracy of pattern recognition and identification.

[0018] Especially in the context of photogrammetry, the use of pattern variables, particularly texture variables, offers additional advantages. In photogrammetry, the accurate reconstruction of three-dimensional objects from two-dimensional images is crucial. Using texture variables in pattern space allows for a more precise characterization of surface textures, which in turn leads to more accurate 3D reconstruction. When creating 3D models from images, it is essential to accurately describe surface properties.

[0019] Texture variables can therefore help to represent the texture and details of surfaces more realistically, which increases the accuracy of model creation in photogrammetry. The use of texture variables also allows for better differentiation between various surface types and textures in the captured images. This is particularly important when it comes to identifying and separating different parts of an object or terrain. In photogrammetric applications, such as terrain mapping or the creation of digital terrain models, the use of texture variables can also significantly improve the efficiency and accuracy of texture mapping. Finally, in surface inspection, for example in the construction industry or archaeological research, texture variables enable a more detailed analysis of surface conditions and structures.

[0020] Overall, the use of texture variables and the application of a corresponding pattern space contribute to increasing the performance, especially of photogrammetry, by providing more accurate and precise information about surface structures and textures. This leads to improved 3D reconstruction and more accurate acquisition of spatial data.

[0021] In a further development of the described method, the pattern, and in particular the texture, is a parametric curve. A pattern or texture in the form of a parametric curve refers to the representation of textures and patterns on surfaces using mathematical parameters that describe a curve shape. This parameterization can be achieved, for example, with the so-called Hoggs transform and allows for the systematic capture of complex textures and patterns. This offers several advantages that positively impact the fundamental concept of the targeted selection of 2D images prior to 3D reconstruction. The parametric curves enable an extremely precise representation of patterns and textures. This allows for better capture of the finest surface details in the 2D images, which improves the quality of the generated 3D reconstruction.Because parametric curves can be easily scaled and adapted, they are also extremely flexible. This is particularly important when different objects or surfaces with varying textures need to be captured. This flexibility allows for adaptation to different scenarios, simplifying the selection of the right 2D images. Furthermore, using parametric curves ensures high consistency and repeatability in texture analysis. This is crucial for guaranteeing consistent results when selecting 2D images for 3D reconstruction. Consistent images lead to more reliable 3D models. Parametric curves can also be mathematically analyzed to extract specific features and properties of the textures. This enables automated image processing and analysis, making the selection process more efficient.Compared to storing image data, parametric curves generally require less storage space. This is advantageous because it reduces storage and transmission costs. More efficient data handling is highly beneficial in photogrammetric applications. The precise representation of surface textures using parametric curves ultimately contributes significantly to the quality of the 3D reconstruction. This is the core aspect of the underlying concept, as the selection of 2D images serves to create accurate 3D models. Overall, using parametric curves as patterns or textures greatly simplifies the selection of the correct 2D images prior to 3D reconstruction. It improves the quality, accuracy, and efficiency of the entire process and enables more precise capture of surface details, ultimately leading to better 3D models.

[0022] In a preferred advanced training, the parametric curve is a straight line, preferably in Hessian normal form, used as a pattern or texture for filtering 2D images. This offers several advantages for image filtering. Parametric straight lines in Hessian normal form are mathematically easy to describe and recognize. This greatly facilitates pattern recognition, as straight lines are common in images and easily identifiable. Straight lines also offer a degree of robustness against image noise. Using a straight line as a pattern reduces the likelihood of random noise being interpreted as a significant pattern. This contributes to improved image filtering. Hessian normal form describes straight lines using clear parameters such as the distance to the origin and the angle to the x-axis.This allows for precise control of the filtering and targeted selection of images exhibiting the desired line patterns. The Hessian normal form is particularly well-suited for representing horizontal lines, as it enables the detection of lines parallel to the x-axis (e.g., x=0). This is beneficial in many applications, such as horizon or horizontal structure detection. Using parametric lines as patterns allows image information to be reduced to its essentials. This can help minimize the amount of data and increase the efficiency of image filtering. Line patterns can be combined with other features and patterns in images to create complex filtering criteria. This opens up the possibility of selectively extracting specific image features.Overall, the use of parametric lines, especially in the Hessian normal form, facilitates image filtering, particularly the selection of 2D images exhibiting the desired line patterns. This contributes to improving the quality of 3D reconstruction, as it allows for a precise selection of images that are crucial for photogrammetry.

[0023] In a further development of the described procedure, a distribution of the characteristic variable, or at least a parameter describing the characteristic variable, is determined to reject one of the two frames, with the corresponding frame being rejected depending on the predetermined condition. Using a distribution function or a histogram of the characteristic variable (or a parameter describing the characteristic variable) to decide which of the two frames to reject offers several advantages. For example, a distribution function provides a statistical basis for the decision. This means that the choice of which frame to reject is based on a comprehensive analysis of the characteristic variable, thus enabling objective decisions. The parameters of the distribution function can be adjusted to make the decision criteria flexible.This allows the sensitivity of frame selection to be varied according to the application's requirements. Distribution functions can reflect the uncertainty in the data. This allows frames that might not clearly fall into the desired category to be evaluated and selected or discarded accordingly. If necessary, other relevant information can be incorporated into the distribution function. This enables a holistic view and selection of frames that goes beyond the feature variable alone. The use of a statistical foundation leads to consistent and reproducible results. This is crucial in applications such as photogrammetry to ensure reliable 3D reconstructions. The distribution function makes it possible to efficiently reduce data by selecting frames that best represent the desired features or properties.This can reduce the amount of data and increase processing efficiency. The distribution function can help identify and appropriately handle outliers or unexpected variations in the data. This is important for maintaining the quality of the 3D reconstruction. Overall, using a distribution function as the basis for deciding whether to discard a frame provides a robust and data-driven method for frame selection. This contributes to improving the accuracy and quality of 3D reconstruction in photogrammetric applications.

[0024] Ideally, a Hogg transform would serve as a feature space that simultaneously yields a distribution function. However, other potential candidates for feature spaces that produce a distribution function exist, such as color histograms, which are often used in image processing to capture the distribution of color information in an image. These histograms could serve as feature spaces and represent the distribution of color values ​​across frames. Instead of a single texture variable, statistical measures could be used to characterize the texture information in an image. This could include texture homogeneity, contrast, and other statistical properties. In videos, the motion vectors between successive frames could serve as a feature space. The distribution of these vectors could provide insights into the motion dynamics within the video stream.The distribution of gradient directions in an image could be used as a feature space. This could be useful for distinguishing images with different structures. Applying a Fourier or wavelet transform to image regions could lead to feature spaces that describe the distribution of frequencies in the frames. This could be helpful in distinguishing images with different frequency components. In applications requiring depth information, feature spaces that represent the distribution of depth values ​​in the frames could be used. This could be useful in 3D reconstruction. As already explained, the choice of the appropriate feature space depends on the specific requirements of the application and the data to be analyzed.Depending on which features are of interest in the frames, a suitable feature space can be selected that represents the relevant information in the form of a distribution function.

[0025] According to another aspect of the invention, a control device is set up to carry out one of the specified procedures.

[0026] In a further development of the specified device, the device comprises a memory and a processor. The specified method is stored in the memory in the form of a computer program, and the processor is provided for executing the method when the computer program is loaded from memory into the processor.

[0027] According to another aspect of the invention, a computer program comprises program code means to perform all steps of the specified method when the computer program is executed on an electronic device or one of the specified devices.

[0028] According to another aspect of the invention, a computer program product includes program code that is stored on a computer-readable data carrier and that, when executed on a data processing device, performs the specified method.

[0029] According to another aspect of the invention, a system for constructing a 3D model from a video stream from a temporal sequence of 2D images comprises a receiving interface for receiving the video stream, one of the specified control devices for filtering the video stream, and a photogrammetry device for generating the 3D model from the filtered video stream.

[0030] The properties, features, and advantages of this invention described above, as well as the manner in which they are achieved, will become clearer in connection with the following description of the exemplary embodiments, which are explained in more detail in conjunction with the drawings. The drawings show:

[0031] Fig. 1 shows a structural view of the essential components of photogrammetry software.

[0032] The figures use identical technical elements with the same reference symbols and describe them only once. The figures are purely schematic and, above all, do not represent the actual geometric relationships.

[0033] Reference is made to Fig. 1, which shows in a structural view the essential components of a photogrammetry software 2.

[0034] First, a feature detection module 4 receives frames 6, or 2D images, derived from a video stream 8 shown in Fig. 2. These frames 6 serve as the basis for identifying distinctive points. The output of the feature detection module 4 consists of the detected distinctive points 10 in the frames 6. Typical distinctive points can be corners, edges, or other visually prominent features. The detected distinctive points 10 are then passed to a feature matching module 12, which uses an algorithm to calculate mapping information 14 from the distinctive points 10 between the different frames. This mapping information 14 describes the mapping of the distinctive points 10 from the feature matching module 12. This mapping information 14 enables the determination of a spatial relationship between the frames 6.The feature-matching module 12 is followed by an image alignment and registration module 16, which uses the mapping information 14 to determine spatially aligned and registered images 18. These registered images 18 are able to provide consistent information about the positions of the distinctive points 10. Subsequently, a triangulation and 3D reconstruction module 20 receives the spatially aligned and registered images 18 as well as camera parameters 22 and determines at least one 3D mesh model 24 from them. This mesh model describes the 3D positions 24 of the distinctive points 10 and reconstructs the three-dimensional structure of an object or scene shown in the video stream 8 in the form of a point cloud. Finally, a texturing and rendering module 26 receives the 3D mesh model 24, the frames 6, and the mapping information 14 and uses this information to generate a textured 3D model 28.The texturing is therefore based on frames 6 and is carried out on the basis of the mapping information 14 in order to create a realistic appearance of the three-dimensional structure of the object or scene shown with the video stream 8.

[0035] The structure shown in Fig. 1 represents the basic process in photogrammetry, in which information from the frames 6 of the video stream is converted into a complete 3D model 28 of an object or scene and is fundamentally known, for example, from WO 2016 / 005 412 A1. To improve the results in the construction of the 3D model 28, the frames 6 should be pre-filtered as explained in WO 2016 / 005 412 A1. Before discussing a corresponding pre-filter 30, explained in Fig. 3, in more detail, the position of such a pre-filter 30 will first be shown with reference to Fig. 2.

[0036] The starting point is a framebuffer (not shown), typically provided by a camera or other image source. This framebuffer contains the aforementioned continuous video stream 8, which consists of successive individual frames (not referenced in Fig. 2).

[0037] The pre-filter 30 is the first component of the process and plays a crucial role in selecting the frames 6 from the video stream 8 that are used for photogrammetry. The pre-filter 30 operates using at least one input parameter 32, which serves as a controller to select the frames 6 and discard others.

[0038] For this purpose, the pre-filter 30 analyzes the video stream 8 from the framebuffer and, using input parameter 32, identifies which frames 6 are relevant for the photogrammetry software 2 in a manner to be described later. Frames 6 that meet the criteria of the pre-filter 30 are selected for the subsequent processing step, while others are discarded.

[0039] After the pre-filter 30 has selected the relevant frames 6, these are converted into a suitable format by a formatter 34, which is optimal for transfer to the photogrammetry software 2. A commonly used format is ZIP, which compresses the frames 6 and combines them into a single file. The formatter 34 can also add additional metadata relevant to the photogrammetry software 2. This includes, among other things, information about the image sequence, camera parameters, and other relevant information required for subsequent 3D reconstruction.

[0040] In Fig. 2, the input and output of formatter 34 are labeled with the same reference symbol and are referred to below as frames. This is purely for clarity and brevity. In reality, formatted frames are passed to the photogrammetry software 2, and the photogrammetry software 2 actually requires a module to reformat the formatted frames.

[0041] Reference is made below to Fig. 3, which describes in more detail the pre-filter 30 for filtering the video stream 8 by discarding individual frames in the video stream 8. The filtering starts with a local gradient detection in the video stream 8. An edge detection algorithm 36 is used for this purpose, which in this embodiment is based on Canny's algorithm. With the edge detection algorithm 36, developed by John F. Canny in 1986, clear and precise edges in the frames of the video stream 8 can be identified.

[0042] The Canny edge detection algorithm 36 begins by smoothing the frames in video stream 8 to reduce noise. This is done by applying a Gaussian blur filter to each frame. The Gaussian filter introduces a slight blur to the image and helps to eliminate minor irregularities. In the next step, the gradients of the frames in video stream 8 are calculated. The gradient at a point in the image indicates the direction of the maximum change in brightness at that point. The Canny edge detection algorithm 36 uses the derivatives in the horizontal and vertical directions (x and y directions) for this purpose. Two gradient images are created: one for the x direction (horizontal gradient) and one for the y direction (vertical gradient). These images contain information about the change in brightness in the horizontal and vertical directions.

[0043] Based on the gradients in the x and y directions, the gradient direction at each point in the respective frame of the video stream 8 is calculated. This is done using the arctangent function, which measures the direction in degrees. The gradient strength at each point is also calculated. This is done by applying the Pythagorean theorem to the gradients in the x and y directions. The gradient strength indicates how much the brightness varies at that point in the image. Next, edge points are searched for using the gradient direction. The Canny edge detection algorithm 36 traverses each point in the gradient image and checks whether the gradient value at that point is a local maximum in its direction. If the gradient value at a point is greater than the gradient values ​​in its vicinity in the direction of the gradient, that point is retained. Otherwise, it is suppressed.This step ensures that only the sharpest edge points in the image are retained. To further refine the edges and identify connected edge structures, two thresholds are used: a high threshold and a low threshold. Pixels with gradient strengths above the high threshold are marked as strong edges and reliably retained as edge points. Pixels with gradient strengths between the low and high thresholds are temporarily marked and reviewed in a subsequent step. Finally, in a joining step, pixels between strong edge points and temporarily marked pixels are considered part of the edge if they are adjacent. This allows for the detection of connected edge structures.

[0044] The Canny edge detection algorithm 36 produces binary frames 38 in which the detected edges are represented as white lines on a black background and is characterized by its ability to identify noise-free and clear edges. The white lines in the binary frames 38 therefore represent

[0045] The Canny edge detection algorithm 36 is followed by a temporal gradient detection algorithm that allows the detection of a temporal gradient between the binary frames 38, so that, based on the input parameters 32, a decision can be made as to whether one of two consecutive frames in the video stream 8 can be discarded. This detection algorithm is implemented based on a Hoggs transform algorithm 40.

[0046] The task of the Hoggs transformation algorithm 40 is to detect any temporal changes between the binary frames 38. This allows the decision, depending on the input parameters 32, as to whether one of two temporally consecutive frames should be discarded.

[0047] The Hoggs transform is a mathematical method used to detect and analyze temporal changes in images, specifically in binary frames. Its fundamental approach is based on the observation that moving objects or changes in a scene appear as patterns of brightness variations over time. A specific parametric curve is defined to capture and quantify these changes. To perform the Hoggs transform, a parametric curve is first defined. This curve describes the expected shape or pattern of the temporal changes to be detected. The choice of parametric curve depends on the type of changes to be detected. Typically, a straight line in Hesse normal form is chosen as the parametric curve.

[0048] A histogram is then created in each binary frame. The histogram shows how the parameters of the parametric curve are distributed in that frame. It captures the differences in brightness or changes in intensity relative to the parametric curve.

[0049] The temporal evolution of the histogram is analyzed over several consecutive frames. This allows for the identification of patterns and trends in the changes of the parametric curve parameters. Based on this temporal analysis, a gradient or derivative of the parametric curve parameters over time is calculated. This gradient reflects the rate or magnitude of the temporal changes in the images.

[0050] Finally, a predefined condition is applied to the calculated gradient. This condition can be a threshold value derived from the input parameters 32 or a specific mathematical function, which can also be defined via the input parameters 32. If the gradient meets this condition, it is recognized whether significant temporal changes are present or not, so that at least one of several consecutive frames in the video stream 8 is filtered out and not output as frame 6.

[0051] The Hoggs transform aims to precisely and quantitatively capture changes in images. It allows for the straightforward identification of moving objects or other dynamic phenomena in a video stream. The choice of the parametric curve and the conditions for gradient calculation are crucial for adapting the transformation to the specific type of changes being analyzed. In photogrammetry, the Hoggs transform enables the targeted selection of frames exhibiting significant changes. This contributes to reducing redundancy and noise, which in turn improves the accuracy and efficiency of 3D reconstruction.

[0052] The Hogg transformation helps to visualize movements, fluctuations, or temporal changes in images. It analyzes the temporal development of features or patterns in the images and can thus help to extract relevant information.

[0053] Overall, the pre-filter 30 enables precise control of the selection of frames in the video stream 8 to ensure that only relevant ones are included.

[0054] Images 6 are used for the subsequent photogrammetry software 2. This contributes to improving the quality and efficiency of the photogrammetry process.

Claims

Claims 1. A method for filtering a video stream by discarding a second frame that follows a first frame in the video stream, comprising: - Detection of a local gradient in the first frame, - Detecting the local gradient from the first frame in the second frame, - Discard one of the two frames if a temporal gradient between the local gradient in the first frame and the local gradient in the second frame satisfies a predetermined condition.

2. The method of claim 1, wherein the predetermined condition is adjustable.

3. The method of claim 1, wherein the predetermined condition is a threshold value that must be undercut to discard one of the two frames.

4. Method according to one of the above features, wherein the frames are transformed into a feature space with at least one feature variable to detect the local gradients.

5. Method according to claim 3, wherein the feature variable of the feature space is a variable describing a pattern, in particular a texture variable of a texture space as a feature space.

6. Method according to claim 5, wherein the pattern, in particular the texture, is a parametric curve.

7. Method according to claim 6, wherein the parametric curve is a straight line, preferably in Hessian normal form.

8. Method according to any one of the preceding claims 4 to 7, wherein to reject one of the two frames a distribution of the feature variable or at least one parameter describing the feature variable is determined and the corresponding frame is rejected depending on the predetermined condition.

9. Control device for carrying out a method according to one of the preceding claims.

10. System for constructing a 3D model from a video stream from a temporal sequence of 2D images, comprising: - a receiving interface for receiving the video stream, - a control device according to claim 9 for filtering the video stream, and - a photogrammetry device for generating the 3D model from the filtered video stream.

Images