Motion segmentation and optical flow field joint constraint motion estimation method based on multivariable kernel density function

Abstract

The invention discloses a motion segmentation and optical flow field joint constraint motion estimation method based on a multivariable kernel density function. The method is suitable for complex dynamic scenes such as satellite videos. The method comprises the following steps: firstly, respectively carrying out background separation and target motion estimation, spatial domain motion estimation and light watershed motion estimation on an input video sequence to obtain preliminary results of a segmentation domain, a spatial domain and a light watershed; then, areas which are all marked as motion pixels in the three domains are reserved through coarse screening, and false motion areas are filtered according to a set pixel threshold value; on this basis, a multivariable kernel density motion fine estimation model is constructed by combining information of a spatial domain, a light basin and a segmentation domain, clustering optimization is performed in a five-dimensional feature space by using a mean shift clustering algorithm, and finally a high-precision and high-consistency motion estimation vector field is output. According to the method, multi-domain information is effectively fused, the limitation of a single method is overcome, and the robustness and accuracy of motion estimation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing and computer vision, and particularly relates to a motion estimation method combining motion segmentation and optical flow field joint constraint based on multivariate kernel density function, which is used to improve the motion estimation accuracy in video images. BACKGROUND

[0002] Image motion estimation is one of the core tasks in the field of computer vision and video processing, which aims to estimate the motion vector of each pixel or region in the scene from a sequence of consecutive images. Accurate motion estimation plays a crucial role in many fields, such as super-resolution reconstruction and moving target monitoring in satellite video analysis, behavior recognition and anomaly detection in video surveillance, and environment perception and obstacle avoidance in autonomous driving.

[0003] However, in complex scenes such as satellite video imaging, images often contain a large range of static background and scattered dynamic targets with small pixel proportion (such as moving vehicles, ships). The coexistence of static background and dynamic targets poses a severe challenge to motion estimation algorithms. Traditional methods often struggle to balance global consistency and local accuracy when dealing with such complex scenes, and have obvious limitations.

[0004] Traditional motion estimation methods can be mainly divided into the following categories:

[0005] Motion estimation based on optical flow: classic optical flow methods such as Lucas-Kanade and Horn-Schunck estimate pixel-level motion vectors by assuming that pixel intensity remains constant over a short period of time. This method performs well in areas with rich texture and smooth motion. (Reference: Brox T, Malik J. Large displacement optical flow: descriptor matching in variational motion estimation[J]. IEEE transactions on pattern analysis and machine intelligence, 2011, 33(3): 500-513.) However, in complex dynamic scenes such as satellite video, it is difficult for optical flow methods to provide accurate estimates for small-sized, texture-sparse moving targets. At the same time, due to the influence of imaging quality, the edges of moving targets and background are often blurred, which can lead to errors in optical flow calculation at the boundary, and may estimate meaningless small motion noise in the vast static background area.

[0006] Background modeling and motion segmentation based methods: These methods (e.g., Gaussian Mixture Model (GMM), frame differencing) separate foreground (moving objects) from background by establishing a background model or differencing consecutive frames. They have an advantage in separating static background (Reference: Barnich O, Van Droogenbroeck M. ViBe: a universal background subtraction algorithm for video sequences[J]. IEEE Transactions on Image processing, 2011, 20(6): 1709-1724.), but are easily disturbed by changes in lighting, target shadows, and noise, resulting in broken "salt and pepper noise" or misjudging target shadows as foreground. In particular, in satellite videos, moving targets may exhibit "thin" characteristics, leading to incomplete segmentation and loss of important motion information.

[0007] Feature information and multi-target tracking based methods: These methods focus on extracting and tracking significant feature points or regions in images. Although they can effectively capture the trajectories of specific moving targets, their estimation results often lack global consistency, and they have weak tracking capabilities for feature points that are sparse or have dramatic appearance changes, making it difficult to generate dense and continuous motion vector fields.

[0008] The comprehensive limitations of existing methods are that they usually rely on a single type of information and fail to fully utilize the complementarity of multi-dimensional information such as spatial structure, motion continuity, and foreground-background semantics in complex scenes. For example, optical flow methods can provide dense motion fields but are easily disturbed by local interference; motion segmentation methods can effectively separate dynamic / static regions but cannot provide accurate motion vectors. This "each for himself" strategy makes it difficult for existing methods to balance estimation accuracy in complex scenes with dynamic backgrounds and target interlacing.

[0009] Multivariate kernel density estimation, as a non-parametric statistical method, has the potential to fuse multi-source heterogeneous data and reveal its underlying distribution structure. In motion estimation, if a kernel density function that combines spatial domain, optical flow domain, and segmentation domain can be constructed and optimized through clustering algorithms such as mean shift, it can theoretically effectively overcome the errors of a single information source and achieve more accurate and consistent motion estimation. However, existing technologies have not fully exploited and utilized the inherent correlation and joint constraint potential between the three.

[0010] Therefore, aiming at the deficiencies of the prior art, the application provides a motion segmentation and optical flow field joint constraint motion estimation method based on a multivariate kernel density function. The core innovation of the application is that a unified multivariate kernel density estimation framework is constructed to deeply fuse and jointly optimize motion segmentation, optical flow estimation and spatial position information. The method can not only accurately segment static backgrounds and dynamic targets, but also effectively corrects the estimation error of the optical flow field at the edges of the moving targets and the micro targets and suppresses the motion noise of the background area in the clustering process by using the mutual constraint and supplement of the multi-domain information, so that a high-precision and high-consistency motion vector field is finally output, thereby providing reliable motion information basis for subsequent video super-resolution reconstruction, target tracking and other advanced tasks. SUMMARY

[0011] Aiming at the above technical problems, the application aims to provide a motion segmentation and optical flow field joint constraint motion estimation method based on a multivariate kernel density function. The method constructs a unified multivariate kernel density estimation model by jointly using the motion estimation in the spatial domain, the vector information of the optical flow field and the foreground / background information in the segmentation domain, and uses a mean shift clustering algorithm for joint optimization, so that a high-precision and high-consistency motion estimation vector field is finally obtained .

[0012] To achieve the above-mentioned purpose, the application provides a motion segmentation and optical flow field joint constraint motion estimation method based on a multivariate kernel density function, which comprises the following steps:

[0013] S1: inputting an original video image sequence , wherein is a reference frame, represents the number of sequence frames, i=1, 2, 3,..., t;

[0014] S2: segmenting the background and the target in the segmentation domain by using an average background modeling method, taking as the reference frame, performing motion estimation on other frames and , and outputting a segmentation domain motion estimation result sequence , wherein each represents the motion estimation result of the i+1th frame and the reference frame ;

[0015] S3: performing motion estimation in the spatial domain by using a diamond search method, outputting a spatial domain result sequence , wherein each represents the motion estimation result of the i+1th frame and the reference frame ;

[0016] S4: performing motion estimation in the optical flow domain by using The sparse optical flow method is used for motion estimation, and the output is a sequence of optical flow domain results. , each of Represents the (i+1)th frame and the reference frame. The motion estimation results;

[0017] S5: Motion estimation results combining spatial domain, optical flow domain, and segmentation domain , , Coarse screening;

[0018] S6: Determine whether the sum of non-zero pixels in the coarse screening result area is greater than a given threshold;

[0019] S7: Adjust parameters if the value is less than the threshold. ;

[0020] S8: If it is greater than the threshold, a joint multivariate kernel density motion estimation model of spatial domain, optical flow domain and segmentation domain will be constructed;

[0021] S9: Solve the motion estimation model, combining mean-shift clustering and feature space clustering methods to solve for the output motion estimation vector field. ;

[0022] In step S5, for each pixel location in the image, the pixel is retained in the coarse screening result area only if it is marked as a moving pixel in the sparse spatial domain result, the dense optical flow domain result, and the binarized segmentation domain result.

[0023] In step S6, the "threshold" is an empirical value set according to the image resolution. It needs to be large enough to filter out the pseudo-motion regions formed by image noise, optical flow calculation jitter, and small false detections during the segmentation process. At the same time, it also needs to be small enough to ensure that typical regions containing real moving targets (such as vehicles and ships in satellite video) can be preserved. Based on the above principles, and considering that moving targets in satellite video are usually scattered entities with a small pixel ratio, the preferred value range of the validity threshold is 50 to 300 pixels.

[0024] In step S7, the parameters are key control parameters for the segmentation domain, spatial domain, and optical flow domain, respectively. Specifically: the parameter is the foreground discrimination threshold of the segmentation domain, with a preferred value range of 10 to 20; the parameter is the motion vector threshold of the spatial domain, with a preferred value range of 0.1 to 1.0 pixels; and the parameter is the motion threshold of the optical flow domain, with a preferred value range of 0.1 to 1.0 pixels.

[0025] The expression for the multivariate kernel density motion precision estimation model jointly constructed from the spatial domain, optical flow domain, and segmentation domain in step S8 is as follows:

[0026] in: Represents the total number of sample points. Represents the motion estimation results, These represent the coordinates of the target point in the spatial domain, optical flow domain, and segmentation domain, respectively. Representing the first The coordinates of each sample point in the three domains Representing spatial domains Light Basin and partitioning domain The weight function, represents the variable bandwidth matrices of the spatial domain, optical flow domain, and segmentation domain, respectively, and k() is the Gaussian kernel function;

[0027] In step S9, the "mean-shift clustering and solution" is performed in the aforementioned five-dimensional feature space. This five-dimensional feature space is composed of two pixel coordinate dimensions in the spatial domain, two motion vector dimensions in the optical flow domain, and one foreground / background discriminant value dimension in the segmentation domain. This process achieves motion pattern clustering by finding density extrema points in the feature space. After clustering, the motion vectors within each cluster are optimized to the median of all motion vectors within that cluster, ultimately generating a high-precision, highly consistent motion estimation vector field. .

[0028] To implement the method described in this invention and ensure the motion estimation vector field The accuracy requires setting the following key parameters. These will be explained from four aspects: spatial domain, optical flow domain, segmentation domain, and multivariate kernel density estimation. Each parameter is closely related to the specific characteristics of the input video sequence (such as resolution, frame rate, noise level, and moving target size), and those skilled in the art can flexibly adjust them within the preferred ranges described below.

[0029] 1. Spatial Domain Motion Estimation Parameters

[0030] In spatial domain motion estimation, the analysis scale, as a fundamental parameter, refers to the size of the local window used for block matching or feature extraction. Its preferred range is 8×8 to 32×32 pixels. The setting of this parameter needs to strike a balance between sensitivity to capturing the motion of small targets and robustness in suppressing noise: a size that is too small (e.g., 8×8) is susceptible to noise interference, leading to increased estimation errors; a size that is too large makes it difficult to accurately represent the motion vector of moving targets with a small pixel footprint.

[0031] 2. Parameters for motion estimation in optical watersheds

[0032] Optical flow methods rely on grayscale changes in image sequences to estimate motion, and their parameter settings are particularly crucial for preserving motion boundaries and suppressing noise. Specifically, these include:

[0033] Smoothing Constraint Weight: This parameter controls the smoothness of the optical flow field at motion boundaries, and its value is negatively correlated with the intensity of the target's motion in the scene. When the motion speed changes greatly, a smaller smoothing weight should be used to preserve motion details.

[0034] Maximum number of feature points: To effectively capture motion in the scene while taking into account computational efficiency, the maximum number of feature points used for optical flow calculation is preferably set between 300 and 800.

[0035] Feature quality threshold: Used to filter feature points with high saliency, preferably in the range of 0.001 to 0.01. This threshold can effectively preserve highly stable feature points while reducing the interference of noise points on optical flow calculation.

[0036] 3. Motion estimation parameters of the segmented domain

[0037] Within the segmentation domain, parameter settings aim to distinguish moving foreground targets from static backgrounds, primarily including:

[0038] Foreground discrimination threshold: For 8-bit grayscale images, this threshold is used to distinguish the grayscale difference between moving targets in the foreground and static background. Its typical value range is 10 to 20. A threshold that is too low will introduce noise and misclassify it as foreground, while a threshold that is too high may cause small moving targets to be missed.

[0039] Target area range: This range is filtered based on the pixel area of ​​connected regions after morphological processing to eliminate small noise and large non-target areas. This range should be set according to the actual pixel proportion of moving targets (such as vehicles or ships) in the video; the minimum area is typically a few pixels, and the maximum area can be hundreds of pixels.

[0040] 4. Multivariate kernel density estimation parameters

[0041] The variable bandwidth matrix in step S8 These are the core parameters of the model; they are related to the aforementioned motion estimation parameters and together determine the clustering effect in the feature space.

[0042] Spatial domain bandwidth matrix Its scale is positively correlated with the spatial domain analysis scale. The larger the analysis scale, the larger the corresponding bandwidth should be, so as to perform clustering over a larger range in space.

[0043] Optical flow domain bandwidth matrix Its scale is negatively correlated with the optical flow smoothing constraint weight. The smaller the smoothing weight, the greater the allowable change in optical flow, and the smaller the corresponding bandwidth should be to distinguish different motion modes.

[0044] Segmentation domain bandwidth matrix Its scale is negatively correlated with the foreground discrimination threshold. The higher the discrimination threshold, the more stringent the foreground discrimination, and the smaller the corresponding bandwidth should be, in order to strengthen the separation of the foreground and background regions in the feature space.

[0045] Compared with the prior art, the features and beneficial effects of the present invention are as follows:

[0046] Multi-domain information fusion overcomes the limitations of single methods: This invention, for the first time, places spatial domain motion estimation, optical flow field information, and motion segmentation information within a unified multivariate kernel density estimation framework, enabling the three to mutually verify and complement each other. The global consistency of the spatial domain, the local accuracy of the optical flow field, and the target-level semantic information of the segmentation domain are effectively combined, significantly improving the robustness of motion estimation in complex scenes.

[0047] Joint optimization for high accuracy: By using the mean shift algorithm to jointly cluster the features of the three domains, the estimation error of optical flow at the edge of the moving target can be effectively corrected, motion information in incompletely segmented areas can be filled in, and false motion noise in the background area can be suppressed, resulting in a motion field with higher accuracy and better consistency.

[0048] High applicability: This method does not rely on a specific motion model and is a data-driven nonparametric method that can adapt well to motion estimation tasks under various complex imaging conditions such as satellite video and surveillance video, providing a more reliable motion information foundation for subsequent advanced vision tasks such as super-resolution reconstruction and target tracking. Attached Figure Description

[0049] Figure 1 This is a flowchart of the motion estimation method based on motion segmentation and joint constraints of optical flow field using multivariable kernel density function according to the present invention; Detailed Implementation

[0050] See below. Figure 1 The motion estimation method based on motion segmentation and joint constraints of optical flow field based on multivariable kernel density function described in this invention is explained in detail.

[0051] like Figure 1 As shown, in order to obtain better motion estimation vector field Mt results, a motion estimation method based on motion segmentation and joint constraint of optical flow field using multivariable kernel density function is provided, including the following steps:

[0052] S1: Input the original video image sequence ,in As a reference frame, Represents the sequence frame number, i = 1, 2, 3, ..., t;

[0053] S2: In the segmentation domain, the background and target are segmented using the average background modeling method, so as to... For reference frame, other frames and Motion estimation is performed, and a sequence of partition domain motion estimation results is output , wherein each represents the motion estimation result of the i+1th frame and the reference frame ;

[0054] S3: Motion estimation is performed in the spatial domain using a diamond search method, and a sequence of spatial domain results is output , wherein each represents the motion estimation result of the i+1th frame and the reference frame ;

[0055] S4: Motion estimation is performed in the optical flow domain using a sparse optical flow method, and a sequence of optical flow domain results is output , wherein each represents the motion estimation result of the i+1th frame and the reference frame ;

[0056] S5: The motion estimation results of the spatial domain, the optical flow domain, and the partition domain are combined , , coarse screening;

[0057] S6: Determine whether the total sum of non-zero pixels in the coarse screening result area is greater than a given threshold value;

[0058] S7: If less than the threshold value, adjust the parameter ;

[0059] S8: If greater than the threshold value, construct a joint multivariate kernel density motion fine estimation model of the spatial domain, the optical flow domain, and the partition domain;

[0060] S9: Solve the motion estimation model, combine the mean shift clustering and feature space clustering methods, and output the motion estimation vector field ;

[0061] In the step S5, for each pixel position in the image, only when it is marked as a motion pixel in the sparse spatial domain result, the dense optical flow domain result, and the binary partition domain result, is the pixel retained in the coarse screening result area;

[0062] In step S6, the "threshold" is an empirical value set according to the image resolution. It needs to be large enough to filter out the pseudo-motion regions formed by image noise, optical flow calculation jitter, and small false detections during the segmentation process. At the same time, it also needs to be small enough to ensure that typical regions containing real moving targets (such as vehicles and ships in satellite video) can be preserved. Based on the above principles, and considering that moving targets in satellite video are usually scattered entities with a small pixel ratio, the preferred value range of the validity threshold is 50 to 300 pixels.

[0063] In step S7, the parameters are key control parameters for the segmentation domain, spatial domain, and optical flow domain, respectively. Specifically: the parameter is the foreground discrimination threshold of the segmentation domain, with a preferred value range of 10 to 20; the parameter is the motion vector threshold of the spatial domain, with a preferred value range of 0.1 to 1.0 pixels; and the parameter is the motion threshold of the optical flow domain, with a preferred value range of 0.1 to 1.0 pixels.

[0064] The expression for the multivariate kernel density motion precision estimation model jointly constructed from the spatial domain, optical flow domain, and segmentation domain in step S8 is as follows:

[0065]

[0066] in:

[0067] Represents the total number of sample points. Represents the motion estimation results, These represent the coordinates of the target point in the spatial domain, optical flow domain, and segmentation domain, respectively. Representing the first The coordinates of each sample point in the three domains Representing spatial domains Light Basin and partitioning domain The weight function, represents the variable bandwidth matrices of the spatial domain, optical flow domain, and segmentation domain, respectively, and k() is the Gaussian kernel function;

[0068] In step S9, the "mean-shift clustering and solution" is performed in the aforementioned five-dimensional feature space. This five-dimensional feature space is composed of two pixel coordinate dimensions in the spatial domain, two motion vector dimensions in the optical flow domain, and one foreground / background discriminant value dimension in the segmentation domain. This process achieves motion pattern clustering by finding density extrema points in the feature space. After clustering, the motion vectors within each cluster are optimized to the median of all motion vectors within that cluster, ultimately generating a high-precision, highly consistent motion estimation vector field. .

[0069] To implement the method and ensure the precision of the motion estimation vector field the following key parameters need to be set. Those skilled in the art can adjust them within the following preferred ranges according to the specific characteristics of the input video sequence, such as resolution, frame rate, noise level and motion target size.

[0070] 1. Spatial domain motion estimation parameters

[0071] In spatial domain motion estimation, the analysis scale is used as a basic parameter, which refers to the size of the local window used for block matching or feature extraction. Its preferred range is 8x8 to 32x32 pixels. The setting of this parameter needs to balance the sensitivity of capturing small target motion and the robustness of suppressing noise: too small a size (such as 8x8) is easily disturbed by noise, leading to increased estimation error; too large a size makes it difficult to accurately represent the motion vector of a motion target with a small pixel occupancy.

[0072] 2. Optical flow domain motion estimation parameters

[0073] Optical flow method relies on the gray scale change in the image sequence to estimate motion, and the parameter setting is particularly important for the preservation of motion boundaries and noise suppression. Specifically, it includes:

[0074] Smoothness constraint weight: This parameter controls the smoothness of the optical flow field at the motion boundary, and its value is negatively related to the motion intensity of the target in the scene. When the motion speed changes greatly, a smaller smoothness weight should be used to preserve the motion details.

[0075] Maximum number of feature points: To effectively capture motion in the scene while considering computational efficiency, the maximum number of feature points used for optical flow calculation is preferably set between 300 and 800.

[0076] Feature quality threshold: used to filter feature points with high saliency, the preferred range is 0.001 to 0.01. This threshold can effectively retain feature points with high stability while reducing the interference of noise points on optical flow calculation.

[0077] 3. Segmentation domain motion estimation parameters

[0078] In the segmentation domain, parameter setting aims to distinguish between foreground moving targets and static backgrounds, mainly including:

[0079] Foreground discrimination threshold: for 8-bit grayscale images, the grayscale difference threshold is used to distinguish between foreground moving targets and static backgrounds, and its typical value range is 10 to 20. Too low a threshold will introduce noise errors as foreground, and too high a threshold may cause small motion targets to be missed.

[0080] Target area range: filter the connected region's pixel area after morphological processing to eliminate small noise and large non-target area. The range should be set according to the actual pixel proportion of the moving target (such as vehicles, ships) in the video, and the minimum area is usually a few pixels, and the maximum area can be hundreds of pixels.

[0081] 4. Multivariate kernel density estimation parameters

[0082] Variable bandwidth matrix in step S8 is the core parameter of the model, which is associated with the aforementioned motion estimation parameters and jointly determines the clustering effect of the feature space:

[0083] Spatial domain bandwidth matrix : its scale is positively correlated with the spatial domain analysis scale, and the larger the analysis scale, the larger the corresponding bandwidth should be to cluster a larger range in space.

[0084] Optical flow domain bandwidth matrix : its scale is negatively correlated with the optical flow smoothing constraint weight. The smaller the smoothing weight, the greater the allowed optical flow change, and the smaller the corresponding bandwidth should be to distinguish different motion patterns.

[0085] Segmentation domain bandwidth matrix : its scale is negatively correlated with the foreground discrimination threshold. The higher the discrimination threshold, the stricter the judgment of the foreground, and the smaller the corresponding bandwidth should be to strengthen the separation of the foreground and background areas in the feature space.

Claims

1. A method of motion estimation with joint constraints of motion segmentation and optical flow field based on multivariate kernel density function, characterized in that, The method comprises the following steps: S1: input original video image sequence wherein is a reference frame, represents the sequence frame number, i = 1, 2, 3,... t; S2: segmenting the background from the target in the segmentation domain using an average background modeling method to obtain a segmentation result S3: taking a frame as a reference frame, performing motion estimation on other frames to obtain a sequence of motion estimation results of the segmentation domain , where each represents a motion estimation result of the i+1th frame and the reference frame​​ S3: motion estimation in spatial domain using diamond search method, output spatial domain result sequence , wherein each represents the motion estimation result of the i+1th frame and the reference frame ; S4: using the optical flow field in sparse optical flow method to perform motion estimation, outputting an optical flow field result sequence , wherein each represents the motion estimation result of the i+1th frame and the reference frame ; S5: motion estimation result combining spatial domain, optical flow domain and segmentation domain 、 、 coarse screening S6: judging whether the total sum of non-zero pixels in the coarse screening result area is greater than a given threshold value; S7: adjust parameter if less than threshold ; S8: if greater than the threshold value, then constructing a multi-variable kernel density motion fine estimation model in combination with the space domain, the optical flow domain and the segmentation domain; S9: solving the motion estimation model, combining the mean shift clustering, the feature space clustering method, solving the output motion estimation vector field ; In the step S5, for each pixel position in the image, only when it is marked as a motion pixel in the sparse space domain result, the dense optical flow domain result and the binary segmentation domain result, the pixel is retained in the coarse screening result area; In the step S6, the "threshold value" is an empirical value set according to the image resolution, which needs to be large enough to filter out the pseudo motion area formed by image noise, optical flow calculation jitter and small false detection in the segmentation process; at the same time, it also needs to be small enough to ensure that the typical area containing the real motion target (such as vehicles and ships in satellite video) can be retained, based on the above principles and considering that the motion target in satellite video is usually a small pixel ratio dispersed entity, the preferred value range of the effectiveness threshold value is 50 to 300 pixels; In the step S7, the parameters are key control parameters of the segmentation domain, the space domain and the optical flow domain, specifically: the parameter is the foreground discrimination threshold value of the segmentation domain, and the preferred value range is 10 to 20; the parameter is the motion vector threshold value of the space domain, and the preferred value range is 0.1 to 1.0 pixel; the parameter is the motion threshold value of the optical flow domain, and the preferred value range is 0.1 to 1.0 pixel; In the step S8, the expression of the multi-variable kernel density motion fine estimation model constructed in combination with the space domain, the optical flow domain and the segmentation domain is: Wherein: represent the total number of sample points, represent the motion estimation results, respectively represent the target point coordinates in the spatial domain, the optical flow domain and the segmentation domain, respectively represent the coordinates of the first sample point in the three domains, respectively represent the weight functions of the spatial domain , the optical flow domain and the segmentation domain , respectively represent the variable bandwidth matrices of the spatial domain, the optical flow domain and the segmentation domain, and k() is a Gaussian kernel function; In the step S9, the "mean shift clustering and solving" is carried out in the above-mentioned five-dimensional feature space composed of two pixel coordinate dimensions of the space domain, two motion vector dimensions of the optical flow domain and one foreground / background discrimination value dimension of the segmentation domain, and the process realizes the clustering of motion patterns by finding the density extreme points in the feature space; after clustering, the motion vector in each cluster is optimized as the median of all motion vectors in the cluster, and finally a high-precision and high-consistency motion estimation vector field Mt is generated.

Images