Analysis of motion in a video data stream

Abstract

A method (100) for analyzing a video data stream, the video data stream comprising a sequence (1) of images (2a-2f), the method having the steps of: • representing (110) each individual image (2a-2f) of the sequence (1) as an overlay (3a-3f) providing a function of location-dependent contributions (4a-4h) to the individual image (2a-2f); • generating (120) a representation (6a-6f) of the individual image (2a-2f) in a work space (6) from parameters (5a-5f) characterizing the overlay (3a-3f); • for at least one of the parameters (5a-5f), finding (130) at least one time schedule (7a-7f) such that in combination with the time schedule (7a-7f), a proper prediction (2a#-2f#) of a subsequent image (2a-2f) in the sequence (1) from an earlier image (2a-2f) in the sequence (1) is obtained; and • evaluating (140) a motion (8) in the sequence (1) of images (2a-2f) from the at least one time schedule (7a-7f).

Description

Technical Field

[0001] The present invention relates to a method for analyzing motion, such as the motion of objects, in a video data stream comprising a sequence of images. Background Technology

[0002] The at least partially automated guidance of vehicles or robots in operational settings or even in public road traffic requires continuous monitoring of the vehicle or robot's environment. Specifically, this environmental monitoring utilizes video cameras as a source of information, for example.

[0003] For planning the future behavior of a vehicle or robot, such as its trajectory, changes in the environment in which it operates are crucial. For example, if an object moves within the scene, it triggers an incentive for the vehicle or robot to adjust its own behavior. For instance, if someone walks from a hiding place between parked vehicles into the driveway, evasive maneuvers or emergency braking may be necessary. Summary of the Invention

[0004] This invention provides a method for analyzing video data streams. The video data stream comprises a sequence of images. Images can be recorded using any technology. For example, in addition to visible light cameras, thermal imaging cameras, ultrasonic sensors, radar sensors, and / or lidar sensors can be used. That is, individual images can also be generated in a multimodal manner.

[0005] Within this method, each individual image is represented as a superposition of functions providing a location-related contribution to that individual image. Representatives of the individual images in the workspace are generated from the parameters characterizing the superposition. Thus, the superimposed parameters obtain a location reference. That is, the level of abstraction of the parameters is significantly lower than, for example, entries in a feature map provided by a convolutional neural network, where location references can only be constructed indirectly via so-called "receptive fields." In other words, the representations in the workspace themselves have meanings that are more easily understood than, for example, representations in the feature map space.

[0006] Therefore, a sequence of representatives in the workspace is formed from a sequence of individual images. These representatives can be created using any known method. For example, individual images can be used with a trained machine learning model, such as Flash3D.

[0007] Now, for at least one of the parameters, a time schedule is derived such that, by combining the time schedule, appropriate predictions for subsequent images in the sequence are obtained from earlier images in the sequence. Motion in the sequence of images is evaluated from at least one time schedule.

[0008] It has been recognized that motion can be derived more directly from the time schedule of parameters than, for example, from the luminous flux describing changes between individual images in the original pixel space. This is because, as previously stated, the superimposed parameters already contain a location reference, and therefore the changes in these parameters already possess a reference to the content related to motion.

[0009] In particular, in scenarios with multiple moving object instances, these moving object instances can be identified and distinguished from each other. This is feasible even without knowing the categories to which the object instances belong. In this respect, human perception of a scene with multiple object instances can be simulated. This perception is also based on the ability to identify and distinguish the moving object instances even without knowing what they are specifically.

[0010] In particular, it can also identify rare types of object instances, where neural networks or other conventional object detectors may not have been trained for such rare types of object instances. For example, uncommon loads, such as furniture, are easily lost on highways due to a lack of secure loading. Based on the timeline of parameters, it can now, for example, identify when such a load detaches from a preceding vehicle.

[0011] Typically, only a portion of an image region in an image sequence is affected by motion. A time-series representation of parameters that change during motion is a compact expression of the motion (“sparse flow”), requiring significantly less memory and computational power compared to light flux. This is particularly advantageous for applications localized to vehicles or robots, where the amount of hardware that can be carried and the energy consumption are typically very limited.

[0012] Furthermore, evaluating motion via a time schedule of parameters is more robust than assessing the presence of noise and only weakly textured image regions. To this end, correlation functions that provide location-related contributions can be sorted and filtered according to scale level. Thus, for example, stable and accurate identification of motion in relevant regions of the vehicle environment can be ensured, particularly during real-world vehicle operation.

[0013] Finally, the flow, expressed as a timeline of parameters, is also clear and easy to understand, especially regarding the location reference of the function that provides a location-related contribution to individual images. It is readily apparent that specific regions of the flow image contain entirely different or absolutely identifiable erroneous directions.

[0014] In a particularly advantageous design, the timeline includes: • Use a candidate timeline to generate predictions for at least one subsequent image in the sequence from earlier images in the sequence. • Compare the predictions with the actual subsequent images in the sequence, and • Based on the deviations identified in the comparison, the changes to the candidate timelines are determined, and these changes are expected to reduce the deviations.

[0015] This means that the candidate timeline can be progressively optimized until the accuracy of subsequent image predictions in the sequence obtained using the candidate timeline is sufficiently high. Thus, a candidate timeline with at least one superimposed parameter can sufficiently accurately describe scene changes.

[0016] The optimization also does not require "labeling" object instances with category classification or other prior knowledge. Instead, it only requires information already present in the sequence of individual images. By eliminating "labeling," significant cost savings are achieved, and strong subjective elements are avoided.

[0017] For example, an empty timetable can be used to initialize the optimization. However, in many cases, it would be more meaningful to initialize the derived timetable using previous, for example, individual images.

[0018] Bias can be determined using any metric, such as mean squared error. Additionally, other terms can be used to accumulate bias, such as comparing image statistics or specifically “penalizing” the occurrence of particular artifacts.

[0019] In a particularly advantageous design, a parameterization scheme is chosen for the time schedule, which has a set of free time schedule parameters. That is, the superimposed parameters are time-dependent, and this time dependence is characterized by the time schedule parameters. The time schedule parameters can be optimized using any optimization method to minimize the deviation between predictions in the sequence and subsequent actual images. For example, the time schedule parameters can be continuous, allowing for the derivation of proposed changes to reduce the expected deviation from existing deviations. If the time schedule parameters take discrete values ​​(e.g., integer values), a search space expanded by the time schedule parameters can also be searched, for example, according to a pre-defined scheme.

[0020] In another particularly advantageous design, the correlation between the overlay and the free time schedule parameters is differentiable. Therefore, in particular, it is possible, for example, to select parameters that are differentiable to the overlay based on the given function, and then make these parameters time-dependent using a time schedule that is subsequently differentiable. For example, it is also possible to selectively choose functions that contribute to the location-relatedness of individual images, and to selectively identify the relationship between these functions and the overlay, such that there exists a parameter to which the overlay result is differentiable. Then, these parameters are made time-dependent again using a time schedule that is differentiable. From the deviation between the predicted side of subsequent images in the sequence and the actual subsequent images in the sequence, it can be calculated how strongly which free time schedule parameters contribute to the deviation. Then, suggestions for changing the time schedule parameters are derived, which are expected to reduce the deviation. This is somewhat analogous to backpropagating the value of a cost function (loss function) to parameters (such as weights) characterizing the behavior of the neural network, where the performance of the neural network is evaluated using the cost function.

[0021] In another particularly advantageous design, at least one superimposed parameter includes the rate of change of another superimposed parameter. The timescale of this parameter contains a velocity field, which is far more intuitive than, for example, "luminous flux" describing changes between pixel images. For instance, if the superimposed parameters include position in spatial coordinates x, y, and z, the rates dx, dy, and dz of change of those spatial coordinates can be added as additional parameters. The timescale of these rates dx, dy, and dz then also produces a timescale of spatial coordinates x, y, and z.

[0022] Specifically, the stacking parameters used to reconstruct the given individual images can be excluded from the time schedule. In particular, the time schedule can then be extended, for example, only to parameters that are themselves added to determine the velocity of motion, such as changes in other parameters. That is, the representation of the individual images themselves remains untouched, and the possibility of predicting future individual images or interpolating intermediate images between existing individual images is simply added additively. There is an important difference from known methods, such as “4D Gaussian scattering” (where location coordinates x, y, z and their temporal evolution dx, dy, dz are re-discussed): such complete optimization sacrifices the reconstruction variation of a given individual image in order to better detect dynamics in the sequence of individual images. The method proposed here detects the dynamics under the boundary condition that the reconstruction of individual images remains invariant.

[0023] In another particularly advantageous design, the timetable for at least one parameter includes a linear evolution of that parameter over at least segments of time. This evolution can be obtained, for example, using linear optimization methods. For instance, gradient descent and / or methods based on solving systems of linear equations can be used.

[0024] In another particularly advantageous design, a sequence of images is chosen in which individual images follow each other at a rate of 10 Hz or higher. Thus, the evolution between two following individual images is not too large in magnitude, and the temporal evolution can be linearized.

[0025] In another particularly advantageous design, the parameters characterizing the superposition include: • Parameters that characterize the behavior of each function. • Parameters characterizing the type and / or intensity of the effect of each function on the image produced by superposition, and • Parameters that characterize the relative weights of multiple functions.

[0026] Therefore, for example, specific parameters can characterize the extent to which a function is translated, rotated, or compressed along one or more coordinate axes. For instance, the type and / or intensity of the influence of individual functions can be determined by parameters used to confirm the location-related contributions of functions to color and / or occlusion forces in the superposition. Parameters characterizing the relative weights of multiple functions relative to each other can, for example, be coefficients of linear combinations or other aggregates.

[0027] In a particularly advantageous design, at least one distribution function is chosen as a function to provide a contribution to the location-related aspects of individual images, the distribution function associating a measure of probability with each location in the individual image. This contribution can be interpreted particularly well and can also be motivated. Therefore, the representation composed of such contributions itself has an implication that can be further evaluated, for example, particularly well by a downstream, task-specific neural network (task network).

[0028] An example of such a distribution function is the probability density function of the Gaussian distribution, often simply referred to as the "Gaussian function". For instance, this function can be characterized as follows: • The three parameters of spatial displacement in the three coordinate directions of Cartesian space. • The three parameters for scaling in the three coordinate directions, • Four parameters used to determine the orientation of the function in space. • Three parameters are used to specify the color, and the contribution of the function in the superposition is represented by the color, where the color is one of the three additive primary colors: red, green, and blue.

[0029] • (Optionally additionally) The velocity vector for translation and / or rotation.

[0030] All these parameters are in the independent variables of the sine, cosine, and exponential functions. Therefore, the Gaussian function can be differentiated with respect to these parameters without any problems.

[0031] In another particularly advantageous design, motion evaluation includes filtering out motions consistent with the motion of the camera used to record the sequence of images. In this way, motions not caused by camera motion can be precisely identified. For example, if a vehicle or robot is moving in traffic and carrying a camera, almost every pixel of the recorded image changes solely due to this motion. That is, the image is filled with light flux. However, for further planning of the vehicle or robot's behavior, it is important that the object moves for its own motivation and thus intersects with the vehicle or robot's trajectory. For example, if a vehicle enters an intersection area from a crossroads, or a pedestrian moves from a hiding place between stopped vehicles into the lane, this motion can be well distinguished from the motion of the vehicle with the camera. This also applies if the vehicle ahead suddenly brakes, because there is a significant speed difference between the motion of the vehicle with the camera and the speed of the vehicle ahead. As a result, a more rapid response can be made to this external motion that is different from its own motion. For example, evasive maneuvers or emergency braking can be taken.

[0032] In another particularly advantageous design, motion evaluation includes: identifying object instances based on their motion as shown in a sequence of images, and / or distinguishing the object instances from one another. Specifically, for example, formations exhibiting consistent changes in location-related overlay parameters in a representation are evaluated as moving object instances. Objects that are close to each other can also be well distinguished, provided they move in different ways. This would be the case, for example, among pedestrians with different intentions.

[0033] Specifically, clustering can be performed on image components related to the time schedule of the parameters. The resulting clusters can be evaluated as belonging to distinct object instances. Clustering can be performed using any method and is completely independent of the categories to which the object instances might belong. Any method can be used for clustering, such as k-means clustering, DBSCAN, or mean shift.

[0034] Therefore, in another particularly advantageous design, at least one change in the candidate time schedule and / or at least one change in the superposition parameter can represent a change in position and / or orientation in space.

[0035] In another particularly advantageous design, motion evaluation involves interpolating intermediate states of a scene shown in a sequence of images between two separate images. By determining how the superposition parameters change over time from one separate image to the next, any point in time between the records of the first and second separate images can be identified. The timescale of the parameters can then be evaluated for said time point, and the superposition of functions for said time point can provide the sought intermediate state. Therefore, if motion is understood by determining the timescale, arbitrary instantaneous records of said motion can be created. In contrast to generating such intermediate images via a generative model, it is ensured that the obtained instantaneous records are geometrically correct and do not contain strongly distorted versions of the objects shown in a given separate image.

[0036] Similarly, for example, new perspectives on a scene can be created by superimposing the displacements in the location-dependent independent variables of the function.

[0037] In another particularly advantageous design, control signals are generated from the evaluated motion. These control signals are used to control vehicles, driver assistance systems, robots, systems for quality control, systems for monitoring areas, and / or systems for medical imaging. Due to more reliable motion recognition via a time schedule of parameters, the probability of a match between the responses of the individually controlled engineering systems to the control signals and sequences of individual images is increased.

[0038] This method can be implemented entirely or partially by a computer. Therefore, the invention also relates to a computer program having machine-readable instructions that, when executed on one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the described method. In this sense, control devices for vehicles and embedded systems for engineering equipment (which are also capable of executing machine-readable instructions) can also be considered computers. Computing instances can be, for example, virtual machines, containers, or serverless execution environments, which can be provided, in particular, in the cloud.

[0039] The present invention also relates to a machine-readable data carrier and / or downloadable product having the computer program. The downloadable product is a digital product that can be transmitted via a data network and downloaded by a data network user, such as a product that can be sold in an online store for immediate download.

[0040] In addition, one or more computers and / or computing instances may be equipped with computer programs, machine-readable data carriers, or downloadable products. Attached Figure Description

[0041] Hereinafter, with reference to the accompanying drawings, other measures to improve the invention are shown in more detail together with the description of preferred embodiments of the invention.

[0042] Example in: Figure 1 An embodiment of a method 100 for analyzing video data streams is shown; Figure 2 A schematic diagram showing the motion 8 obtained according to method 100 is shown; Figure 3 This illustrates an application example of separating motion using method 100. Detailed Implementation

[0043] Figure 1 This is a schematic flowchart of one embodiment of a method 100 for analyzing a video data stream. The video data stream includes a sequence of images 2a-2f.

[0044] In step 110, each individual image 2a-2f of sequence 1 is represented as a superposition 3a-3f of functions that provide location-related contributions to the individual images 2a-2f.

[0045] According to block 111, a sequence 1 of images can be selected, in which individual images 2a-2f follow each other at a rate of 10Hz or higher.

[0046] According to block 112, at least one distribution function can be selected as a function to provide a location-related contribution 4a-4h to the individual images 2a-2f, wherein the distribution function associates a measure of probability with each location in the individual images 2a-2f.

[0047] According to block 112a, at least one probability density function of a Gaussian distribution can be selected as the distribution function.

[0048] The parameters 5a-5f are superimposed on the 3a-3f. For example, the parameters 5a-5f can be included in the independent variables of a function that provides the location-related contributions 4a-4h to the individual images 2a-2f. Alternatively, or in conjunction with this, the parameters 5a-5f can, for example, identify coefficients by which the location-related contributions 4a-4h of multiple functions are aggregated. In step 120, representatives 6a-6f of the individual images 2a-2f in workspace 6 are generated from the parameters 5a-5f.

[0049] According to block 122, for example, the parameters 5a-5f characterizing the superposition of 3a-3f can therefore specifically include: • Parameters 5a-5f characterize the behavior of each function. • Parameters 5a-5f characterize the type and / or intensity of the effect of each function on the image generated by superimposing 3a-3f, and • Parameters 5a-5f characterize the relative weights of multiple functions.

[0050] In step 130, at least one time schedule 7a-7f is obtained for at least one of the parameters 5a-5f, such that appropriate predictions 2a#-2f# of subsequent images 2a-2f in sequence 1 are derived from the earlier images 2a-2f in sequence 1 in combination with the time schedule 7a-7f.

[0051] The determination of the time schedule 7a-7f may include, for example: • Based on block 131, using candidate time schedules 7a#-7f#, from the earlier images 2a-2f in sequence 1, the predicted values ​​2a#-2f# for at least one subsequent image 2a-2f in sequence 1 are obtained. • According to block 132, the predicted 2a#-2f# is compared with the actual subsequent images 2a-2f in sequence 1, and • According to block 133, based on the deviation Δ confirmed in the comparison, the changes 7a*-7f* of the candidate time schedule 7a#-7f# are calculated, and the changes are expected to reduce the deviation Δ.

[0052] According to block 134, a parameterization scheme for the set of free time schedule parameters can be selected for time schedules 7a-7f. Here, for example, in particular, according to block 134a, the correlation between the superpositions 3a-3f and the free time schedule parameters can be made differentiable.

[0053] According to block 135, the time schedule 7a-7f may include the linear evolution of at least a segment of the parameter 5a-5f over time for at least one parameter 5a-5f.

[0054] In step 140, motion 8 in sequence 1 of images 2a-2f is evaluated from at least one time schedule 7a-7f.

[0055] According to block 141, for example, the evaluation of motion 8 may particularly include filtering out motions consistent with the motion of the camera used to record images 2a-2f in sequence 1. As previously explained, for example, the camera may be mounted on a vehicle or robot such that the entire image changes continuously due to the movement of the vehicle or robot itself.

[0056] According to block 142, for example, the evaluation of motion 8 may in particular include: identifying the object instances and / or distinguishing them from one another based on the motion of the object instances shown in sequence 1 of images 2a-2f.

[0057] Here, for example, the image components in terms of the time schedule parameters can be clustered according to block 142b. According to block 142b, the clusters obtained here can then be evaluated as belonging to distinct object instances.

[0058] According to block 143, the evaluation of motion 8 may include: interpolating the intermediate state of the scene shown in sequence 1 of images 2a-2f to the two separate images 2a-2f.

[0059] exist Figure 1 In the example shown, in step 150, a control signal 150a is generated from the evaluated motion 8. In step 160, the control signal 150a is used to control the vehicle 50, the driver assistance system 51, the robot 60, the quality control system 70, the monitoring area system 80, and / or the medical imaging system 90.

[0060] In one example, 6a-6f can be represented by parameters 5a-5f, which can be expressed as a vector. The vector G describes the three-dimensional Gaussian distribution function, where... - Explain the coverage. - Describe the mean (center) of the distribution function. - It is the covariance matrix, and - It is the radiance (directional color) of each Gaussian component.

[0061] Therefore, the distribution function can be described as follows: .

[0062] Colors associated with the distribution function are typically described using spherical harmonic functions, such that G... .in, It is the direction of observation, and Yl m It is a spherical harmonic function of different orders m and degrees l.

[0063] The following color functions represent the occlusion force and radiation field of G: - Covering force at a point in three-dimensional space: : .

[0064] Radiance along the ν direction at -x: .

[0065] The field is rendered into image J by integrating the radiance along the line of sight using the emission-absorption equation: in, It is a ray that spreads from the center x0 of the camera along the −ν direction toward pixel u.

[0066] For image reconstruction using Gaussian scattering, it is important to make an effective approximation of the integral. To this end, a differentiable rendering function is established. The rendering function takes a representative vector G and a viewpoint π as input and outputs an estimate of the image visible from the viewpoint. .

[0067] Now, time correlation can be added to the representative G. t For example, add them in the form of the rates of change of coordinates x, y, and z, dx, dy, and dz:

[0068] .

[0069] Now, in order to obtain the "Gaussian flux" ΔG between individual images recorded at time points t and t+Δ, the following optimization problem can be solved: Where I_t+Δt is the image at time t+Δt, |.| is any distance function in the image space (e.g., L1, L2), GS() is the rendering function described above, and G_t is the representative of time t, which is obtained, for example, by using Flash3D: G_t=Flash3D(I_t).

[0070] Figure 2 The processing steps are illustrated using an exemplary sequence 1 of six individual images 2a-2f. Each individual image 2a-2f is represented as a superposition of functions 3a-3f, which provide location-related contributions 4a-4h to the corresponding individual image 2a-2f. The superpositions 3a-3f are characterized by parameters 5a-5f, respectively.

[0071] Starting with a first individual image, such as 2a, and a second individual image, such as 2b, a time schedule, such as 7a, can be derived, which guides the calculation from parameter 5a of the first individual image 2a to parameter 5b of the second individual image 2b. This time schedule, 7a, can be combined with the first individual image, 2a, to form a prediction of the second individual image, 2b#. The time schedule 7a can be optimized to ensure that the predicted value 2b# matches the actual individual image 2b as closely as possible.

[0072] Similarly, processing can be performed using individual images from any different pairs. From the derived time schedules 7a-7e, motion 8 in sequence 1 of individual images 2a-2f can be evaluated.

[0073] Figure 3The following scenario illustrates the implementation of method 100: distinguishing between important and unimportant movements. Scenario 10 is shown, featuring a street 11 on which a vehicle 12 to be controlled moves, and on which other vehicles 13a-13g pass. Other vehicles 13a-13f are stopped at the edge of road 11, while other vehicle 13g approaches the vehicle 12 in the opposite lane. Additionally, a pedestrian 14 enters lane 11 through the gap between other vehicles 13c and 13d.

[0074] A camera mounted on the vehicle 12 records images that are constantly changing due to the vehicle 12's own movement. That is, when two separate images 2a-2f that follow each other are compared, the entire image is filled with motion. Using the method presented herein, all motion consistent with the vehicle 12's own movement can be filtered out. The movement of external vehicles 13a-13f and street 11 relative to the vehicle 12 is not surprising, but rather predictable due to the vehicle 12's own movement. Only the movement 8 of external objects is important for possible adjustments to the vehicle's future behavior; this movement is not always predictable due to the vehicle 12's own movement, but rather caused by external intent.

[0075] First, consider the approaching vehicle 13g. Although the approaching vehicle 13g is heading towards the vehicle 12, this is happening in the opposite lane designated for this purpose on street 11. This in itself does not require any adjustment to the planned behavior of the vehicle 12, but rather the vehicle needs to continue straight in its lane.

[0076] The behavior of pedestrian 14 is significantly more critical because the pedestrian steps onto street 11 and then into the lane where vehicle 12 is traveling. Therefore, the movement of pedestrian 14 causes vehicle 12 to take either evasive or braking maneuvers. However, evasion is not an option because the right edge of the street in front of vehicle 12 is occupied by oncoming vehicles 13b and 13c, and oncoming vehicle 13g is moving in the opposite lane. Therefore, the only correct response for vehicle 12 is to brake fully. The proposed method 100 achieves this by filtering out the irrelevant and predictably motion of oncoming objects, enabling a solution to be found faster than, for example, by analyzing the light flux of an image where all areas are in motion.

Claims

1. A method (100) for analyzing a video data stream, the video data stream comprising a sequence (1) of images (2a-2f), the method comprising the following steps: • Each individual image (2a-2f) of the sequence (1) is represented (110) as a superposition (3a-3f) of functions that provide the location-related contribution (4a-4h) to the individual image (2a-2f). • Generate (120) a representation (6a-6f) of the individual image (2a-2f) in the workspace (6) from the parameters (5a-5f) that characterize the superposition (3a-3f). • For at least one of the parameters (5a-5f), find (130) at least one time schedule (7a-7f) such that, in conjunction with the time schedule (7a-7f), an appropriate prediction (2a#-2f#) for subsequent images (2a-2f) in the sequence (1) is obtained from the earlier images (2a-2f) in the sequence (1); and • The motion (8) in the sequence (1) of the images (2a-2f) is evaluated (140) from the at least one time schedule (7a-7f).

2. The method (100) according to claim 1, wherein determining the time schedule (7a-7f) comprises: • Using the candidate time schedule (7a#-7f#), the prediction (2a#-2f#) for at least one subsequent image (2a-2f) in the sequence (1) is obtained (131) from the earlier images (2a-2f) in the sequence (1). • Compare the predictions (2a#-2f#) with the actual subsequent images (2a-2f) in the sequence (1) (132), and • Based on the deviation (Δ) confirmed in the comparison, calculate (133) the change (7a*-7f*) of the candidate time schedule (7a#-7f#), which is expected to reduce the deviation (Δ).

3. The method (100) according to any one of claims 1 to 2, wherein a parameterization scheme having a set of free time schedule parameters is selected (134) for the time schedule (7a-7f).

4. The method (100) according to claim 3, wherein the correlation of the superposition (3a-3f) with the free time schedule parameter is differentiable (134a).

5. The method (100) according to any one of claims 1 to 4, wherein at least one parameter (5a-5f) of the superposition (3a-3f) includes (121) the rate of change of the additional parameter (5a-5f).

6. The method (100) according to any one of claims 2 to 5, wherein at least one variation (7a*-7f*) of the candidate time schedule (7a#-7f#) and / or at least one variation of the parameters (5a-5f) of the superposition (3a-3f) represent a change in position and / or orientation in space.

7. The method (100) according to any one of claims 1 to 6, wherein the time schedule (7a-7f) of at least one parameter (5a-5f) comprises (135) the linear evolution of the parameter (5a-5f) over at least a segment of time.

8. The method (100) according to any one of claims 1 to 7, wherein (111) a sequence (1) of images is selected, wherein the individual images (2a-2f) follow each other at a rate of 10 Hz or higher.

9. The method (100) according to any one of claims 1 to 8, wherein the parameters (5a-5f) characterizing the superposition (3a-3f) include (122). • Parameters characterizing the behavior of each function (5a-5f). • Parameters (5a-5f) characterizing the type and / or intensity of the influence of each function on the image produced by the superposition (3a-3f), and • Parameters (5a-5f) characterizing the relative weights of multiple functions to each other.

10. The method (100) according to any one of claims 1 to 9, wherein at least one distribution function is selected (112) as a function providing a location-related contribution (4a-4h) to the individual images (2a-2f), the distribution function associating a measure of probability with each location in the individual images (2a-2f).

11. The method (100) of claim 10, wherein at least one probability density function of (112a) Gaussian distribution is selected as the distribution function.

12. The method (100) according to any one of claims 1 to 11, wherein the evaluation of motion (8) comprises: Filter out (1) the motion that is consistent with the motion of the camera used to record the sequence (1) of images (2a-2f).

13. The method (100) according to any one of claims 1 to 12, wherein the evaluation of motion (8) comprises: Identify the object instances based on the motion of the object instances shown in the sequence (1) of the images (2a-2f), and / or distinguish the object instances from one another (142).

14. The method (100) of claim 13, wherein the image components in terms of the time schedule parameters are clustered (142a), and the clustering evaluation obtained therein is assigned (142b) to belong to different object instances.

15. The method (100) according to any one of claims 1 to 14, wherein the evaluation of motion (8) comprises: Interpolate (143) the intermediate state of the scene shown in the sequence (1) of the images between the two separate images (2a-2f).

16. The method (100) according to any one of claims 1 to 15, wherein • A control signal (150a) is generated (150) from the assessed motion (8), and • Using the control signal (150a) to control (160) the vehicle (50), the driver assistance system (51), the robot (60), the system for quality control (70), the system for monitoring the area (80), and / or the system for medical imaging (90).

17. A computer program comprising machine-readable instructions that, when executed on one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the method (100) according to any one of claims 1 to 16.

18. A machine-readable data carrier and / or downloadable product having the computer program according to claim 17.

19. One or more computers and / or computing instances having a computer program as claimed in claim 17, and / or having a machine-readable data carrier and / or downloadable product as claimed in claim 18.

Images