Edge outlier processing method and device for depth image

Abstract

The application provides an edge flying point processing method and device of a depth image, wherein the method comprises: performing edge detection on a depth image to obtain an edge region, the edge region comprising a plurality of edge points including flying points; constructing a plurality of search regions corresponding to each edge point on a time bin graph corresponding to a peak value; performing clustering analysis on all pixel points in each search region to confirm the scene category of each pixel point in the search region; obtaining the scene category of an edge point in a search region, and taking the time bin of the pixel point closest to the edge point in the same category as the correction value of the edge point; and in this way, the correction of the time bin of all edge points is completed. Through the combination of the time bin graph corresponding to the peak value, the construction of the search region of each edge point in the depth image and the clustering analysis, the correction of the time bin is completed, thereby effectively improving the processing precision of the flying point phenomenon in the direct time-of-flight depth image and the edge definition.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and apparatus for processing edge flying points in depth images. Background Technology

[0002] Compared to traditional grayscale or color images, depth images possess three-dimensional feature information of objects, and are therefore increasingly used in fields such as computer vision and computer graphics. Currently, TOF (Time of Flight) cameras are mainly used to generate depth images of objects.

[0003] However, when using TOF technology, flying spot effect occurs at image edges, which is a discontinuous depth measurement caused by large differences in depth of field. Specifically, in iTOF (indirect time of flight) cameras, due to their imaging principle, the depth value at the edge is obtained by the combined effect of objects at different distances, so flying spots inevitably appear between the foreground and background. In dTOF (direct time of flight) cameras, although the depth near the edge selects the depth value with the highest probability between the foreground and background depth values, theoretically dTOF technology will not have flying spots between the foreground and background. However, due to the error in the depth calculation process of dTOF technology, flying spots with relatively large depth errors are also prone to appear at the edges.

[0004] Currently, the flying point phenomenon in iTOF is generally filtered out using image-based methods. While dTOF can adopt this approach, it is prone to missed detections and false detections. Furthermore, the filtered image is prone to forming holes, making the flying point processing for dTOF cameras inaccurate. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method and apparatus for processing flying points at the edges of depth images, so as to improve the processing accuracy of flying point phenomena in direct time-of-flight depth images and improve the edge clarity of depth images.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] The first aspect of this invention provides a method for processing flying points on the edges of a depth image, comprising the following steps:

[0008] Edge detection is performed on the depth map to obtain edge regions, which include several edge points, including flying points;

[0009] On the time-bin plot corresponding to the peak, multiple search regions are constructed with each edge point as the center.

[0010] Cluster analysis is performed on all pixels within each search area to determine the scene category of each pixel within the search area;

[0011] Within a search area, the scene category of an edge point is obtained, and the timebox of the pixel closest to that edge point in the same category is used as the correction value for that edge point. This process is repeated to correct the timeboxes of all edge points.

[0012] In one embodiment, constructing multiple search regions centered on each edge point includes:

[0013] Construct a circle with each edge point as the center and a radius of N pixels, where N is a positive integer;

[0014] The circumscribed square region corresponding to each circle is identified as the search area.

[0015] In one embodiment, the value of N ranges from 2 to 5.

[0016] In one embodiment, the depth map is obtained by the following steps:

[0017] Obtain a coarse histogram of each pixel in the pixel array at a first resolution;

[0018] The coarse flight time of each pixel is calculated based on the coarse histogram, and the exposure window area is determined based on the coarse flight time.

[0019] A fine histogram of each pixel in the pixel array is acquired at a second resolution in the exposure window area;

[0020] A depth map is generated by calculating the depth value of each pixel based on the fine histogram.

[0021] In one embodiment, the process of performing edge detection on the depth map to obtain an edge region, the edge region including a number of edge points including flying points, further includes: performing time-box correction on the edge points of the coarse histogram;

[0022] After the timebox calibration of all edge points is completed, the following steps are also included:

[0023] The coarse flight time of each pixel is calculated based on the corrected coarse histogram, and the exposure window area is determined based on the coarse flight time.

[0024] A fine histogram of each pixel in the pixel array is acquired at a second resolution in the exposure window area;

[0025] A fine depth map is generated by calculating the depth value of each pixel based on the fine histogram.

[0026] In one embodiment, after performing edge detection on the depth map to obtain edge regions, the method further includes:

[0027] The edge region is expanded using an erosion algorithm.

[0028] In one embodiment, the step of performing cluster analysis on all pixels within each search region to determine the scene category of each pixel within the search region includes:

[0029] Based on the time bins of each pixel within the search area, cluster analysis is performed on the pixels within the search area to obtain several cluster centers.

[0030] The corresponding scenario category is determined based on the timebox of each cluster center.

[0031] In one embodiment, before performing cluster analysis on the pixels within the search area to determine the scene category of each pixel within the search area, the method further includes:

[0032] The number of cluster centers is adjusted based on the number of foreground and background objects in the current scene.

[0033] In one embodiment, the number of cluster centers ranges from 2 to 5.

[0034] A second aspect of the present invention provides an edge flying point processing apparatus for a depth image, comprising:

[0035] The edge detection module is used to perform edge detection on the depth map to obtain edge regions, which include several edge points including flying points;

[0036] The building module is used to construct multiple search regions centered on each edge point on the time box plot corresponding to the peak.

[0037] The clustering analysis module is used to perform clustering analysis on all pixels in each search area to determine the scene category of each pixel in the search area.

[0038] The edge correction module is used to obtain the scene category of an edge point within a search area, and use the timebox of the pixel closest to the edge point in the same category as the correction value of the edge point. This process is repeated to correct the timeboxes of all edge points.

[0039] The beneficial effects of the present invention are as follows: It provides a method and apparatus for processing flying points on the edges of depth images. By combining the time-box map corresponding to the peak value, a search region is constructed for each edge point in the depth image and cluster analysis is performed to complete the time-box correction, which effectively improves the processing accuracy and edge clarity of flying point phenomena in direct time-of-flight depth images. Attached Figure Description

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0041] Figure 1 This is a flowchart of the edge flying point processing method for depth images in an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of an edge region in an embodiment of the present invention;

[0043] Figure 3 This is a schematic diagram of a search area in an embodiment of the present invention;

[0044] Figure 4 This is a structural diagram of the edge flying point processing device for depth images in an embodiment of the present invention. Detailed Implementation

[0045] To make the technical problems, technical solutions, and beneficial effects of the embodiments of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0046] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly on or indirectly on that other component. When a component is referred to as "connected to" another component, it can be directly connected to or indirectly connected to that other component. Furthermore, a connection can be for both fixing and circuit connection purposes.

[0047] It should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0048] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0049] The edge flying point processing method for depth images provided in this invention is applied to a depth imaging system based on dTOF (direct time of flight) technology. This depth imaging system includes at least a controller, a transmitter, and a receiver. The controller is connected to both the transmitter and the receiver. The transmitter emits a probe beam towards a target object, and at least a portion of the probe beam is reflected by the target object to form reflected light. The receiver includes a pixel array composed of multiple pixels, used to receive the reflected light from the target object. The controller synchronously controls the emission and reception of light, performs histogram statistics on the photons received by the receiver by time bin, and then calculates the flight time of the photons using the histogram to obtain the depth value of each pixel, thus constructing a depth map of the target object.

[0050] Specifically, the transmitter includes a driver and a light source, which can be a light-emitting diode (LED), a laser diode (LD), an edge-emitting laser (EEL), a vertical-cavity surface-emitting laser (VCSEL), a picosecond laser, etc. Under the drive and control of the driver, the light source emits a detection beam, which can be visible light, infrared light, ultraviolet light, etc. At least a portion of the detection beam is emitted toward the target object, and the reflected light generated by the reflection of at least a portion of the detection beam by the target object is received by the receiver.

[0051] The receiver includes a pixel array and receiving optical elements, which can be one or more combinations of lenses, microlens arrays, mirrors, etc. The receiving optical elements receive reflected light and guide it to the pixel array. The pixel array includes multiple pixels that collect photons. In one embodiment, the pixel array consists of multiple single-photon avalanche photodiodes (SPADs). The SPADs can respond to the incident single photon and output a photon signal indicating the arrival time of the received photon at each SPAD. Of course, in other embodiments, photoelectric conversion devices such as avalanche photodiodes, photomultiplier tubes, silicon photomultiplier tubes, etc., can also be used.

[0052] Currently, in dTOF cameras, for situations where depth measurement is discontinuous due to large differences in depth of field, i.e., the flying point phenomenon at image edges, if image processing methods similar to those used in iTOF cameras are applied, it is easy to have missed detections and false detections. Furthermore, the processed image is prone to forming holes, resulting in inaccurate edge flying point processing and affecting the edge sharpness of the depth map. Therefore, this invention describes how to solve this problem by correcting the timebox of edge points through edge flying point processing methods applied to dTOF imaging systems. This method utilizes the rich histogram information in dTOF to correct the flying point phenomenon in the depth map and improve edge sharpness.

[0053] like Figure 1 As shown, Figure 1 This is a flowchart of a method for processing flying points on the edge of a depth image according to one embodiment of the present invention. The method specifically includes the following steps:

[0054] S101. Perform edge detection on the depth map to obtain edge regions, wherein the edge regions include several edge points, including flying points.

[0055] In this embodiment, edge detection is performed on the depth map acquired by the dTOF camera to find the edge set. The principle of image edge detection is to detect all edge points with large changes in depth value in the image. These edge points are connected to form the image edge lines, and the set of all edge points forms the edge region.

[0056] like Figure 2 As shown, the detected edge region 20 includes several edge points, including flying point 201. That is, the edge region can consist entirely of flying points 201 with inaccurate depth values, or it can be as follows: Figure 2 The image shows both flying points 201 and normal pixels 202 with accurate depth values, which are specifically related to the edge detection results. This embodiment does not limit this.

[0057] In practice, edge detection can be performed using existing algorithms such as Soble edge detection, Laplacian edge detection, and Canny edge detection to obtain the edge region. The Canny edge detection algorithm is preferred because it is not easily affected by noise, has good edge localization, and does not perform duplicate detections, thus enabling efficient and accurate acquisition of edge regions from depth maps.

[0058] S102. On the time box plot corresponding to the peak, construct multiple search regions centered on each edge point.

[0059] Since the dTOF camera can acquire histogram information of each pixel, the peak value in the histogram of each pixel corresponds to the coarse time, which is in line with the depth value of each pixel. Therefore, this embodiment uses the time box map as the basis for edge flying point correction processing. On the time box map, multiple search regions are constructed one-to-one with each edge point as the center, so that the characteristics of each edge point and its neighboring points can be analyzed by combining the time box value.

[0060] S103. Perform cluster analysis on all pixels within each search area to confirm the scene category of each pixel within the search area.

[0061] For each edge point, cluster analysis is performed on all pixels within the search area. This cluster analysis is based on the time bins of all pixels within the search area. If there are 25 pixels in a search area, the input to the cluster analysis is the ID position of the 25 pixels and their corresponding time bins. The output is the scene category to which each pixel belongs. Pixels with the same characteristics are clustered into the same category. The scene category to which each pixel in the search area belongs is determined according to the clustering results and the pre-set scene judgment rules, such as whether it belongs to the foreground or the background.

[0062] S104. Within a search area, obtain the scene category of an edge point, and use the timebox of the pixel closest to the edge point in the same category as the correction value of the edge point. Repeat this process to complete the correction of the timeboxes of all edge points.

[0063] Then, within a search area, the scene category of the edge points within the search area is randomly obtained based on the results of cluster analysis. For example, if the edge point belongs to the foreground category, the timebox of the pixel closest to the edge point is selected from all pixels belonging to the foreground category as the correction value of the edge point to complete the correction of the timebox of the edge point. Then, the above process is repeated for the next search area to complete the correction of the timebox of the next edge point, and so on, until the correction of the timeboxes of all edge points is completed.

[0064] Specifically, taking edge points as the foreground category as an example, if the edge point is a flying point with inaccurate depth value during correction, the timebox of the flying point will be updated to the timebox of the pixel closest to it in the foreground category after correction; if the edge point is a normal pixel with accurate depth value, its timebox will also be updated to the timebox of the pixel closest to it in the foreground category after correction. At this time, the timebox of the normal pixel does not change before and after correction. That is, this embodiment performs uniform timebox correction for the edge area to ensure the coverage of flying point processing and prevent the omission of flying point processing.

[0065] In this embodiment, a search region is constructed for each edge point by combining the time-box map corresponding to the peak value and performing cluster analysis to analyze the characteristic relationship between the edge point and its neighboring pixels. Time-box correction of edge points is performed in the same scene category with the same characteristics. The correction of the time-box corresponds to the correction of the depth value, which makes the depth value correction of edge regions, including flying points, accurate and reliable. It is less likely to miss flying point processing and there will be no holes. This improves the processing accuracy of flying point phenomenon in direct time-of-flight depth images and improves the edge clarity of depth images.

[0066] In one embodiment, step S102 includes:

[0067] Construct a circle with each edge point as the center and a radius of N pixels, where N is a positive integer;

[0068] The circumscribed square region corresponding to each circle is identified as the search area.

[0069] In this embodiment, when constructing the corresponding search region for each edge point, a circle is constructed with each edge point as the center and N pixel spacing as the radius. The outer square region corresponding to the circle is then identified as the search region. That is, the region defined by the outermost pixels on the top, bottom, left, and right sides of each edge point with N pixel spacing as the radius is used as the search region, where N is a positive integer. For example, when N is 2, the constructed search region is as follows: Figure 3 As shown, there are 5*5=25 pixels. The specific value of N can be flexibly adjusted according to the needs, and this embodiment does not limit it.

[0070] In one embodiment, the value of N ranges from 2 to 5.

[0071] In this embodiment, when constructing the search region, the radius of the region is preferably between 2 and 5 pixels. A smaller radius means a smaller search region, resulting in insufficient sample size for cluster analysis and affecting the cluster analysis results. On the other hand, a larger radius means more normal pixels in the search region, making the analysis results more accurate. However, an excessively large radius can also lead to more outliers in the search region, making it less accurate to use timeboxes of neighboring pixels for outlier correction. Selecting an appropriate radius can effectively improve the accuracy of cluster analysis and timebox correction.

[0072] In one embodiment, the depth map is obtained by the following steps:

[0073] Obtain a coarse histogram of each pixel in the pixel array at a first resolution;

[0074] The coarse flight time of each pixel is calculated based on the coarse histogram, and the exposure window area is determined based on the coarse flight time.

[0075] A fine histogram of each pixel in the pixel array is acquired at a second resolution in the exposure window area;

[0076] A depth map is generated by calculating the depth value of each pixel based on the fine histogram.

[0077] In this embodiment, a coarse and fine depth calculation method is used. First, a coarse histogram of each pixel in the pixel array is obtained at a lower first resolution. Based on the highest peak in the coarse histogram, the coarse time of flight of each pixel is calculated, which is to perform coarse positioning within a larger range and obtain the approximate distance of the target object to be imaged. Then, based on the coarse time of flight, the exposure window area is determined, which is the measurement range containing the actual time of flight value. In the exposure window area, a fine histogram of each pixel in the pixel array is obtained at a higher second resolution, which is greater than the first resolution, so that further fine positioning can be performed on the basis of coarse positioning. Based on the highest peak in the fine histogram, the fine time of flight of each pixel is calculated, and then the depth value of each pixel is obtained to generate a high-precision depth value. For the depth map obtained by dynamic coarse and fine adjustment imaging, the edge point timebox correction is performed through the edge flying point processing process in the above embodiment to further improve the edge clarity of the depth map.

[0078] In one embodiment, step S101 is followed by:

[0079] Time-box correction of edge points in the coarse histogram;

[0080] After the timebox calibration of all edge points is completed, the following steps are also included:

[0081] The coarse flight time of each pixel is calculated based on the corrected coarse histogram, and the exposure window area is determined based on the coarse flight time.

[0082] A fine histogram of each pixel in the pixel array is acquired at a second resolution in the exposure window area;

[0083] A fine depth map is generated by calculating the depth value of each pixel based on the fine histogram.

[0084] In this embodiment, similar to the previous embodiment, dynamic coarseness adjustment is also used to calculate depth and generate a depth map. The difference is that in this embodiment, the coarse histogram is also corrected for edge points in the process of generating the depth map. That is, after obtaining the coarse histogram of each pixel in the pixel array at a lower first resolution, the edge flying point processing process in the above embodiment is used to correct the edge points in the coarse histogram for time boxes. After the time boxes of all edge points in the coarse histogram are corrected, the subsequent fine positioning imaging process is performed based on the corrected coarse histogram, thus realizing edge flying point processing in the depth map generation process.

[0085] In one embodiment, after step S101, the method further includes:

[0086] The edge region is expanded using an erosion algorithm.

[0087] In this embodiment, after edge detection obtains the edge region, the edge region is further expanded to increase the number of edge points in the edge region. Specifically, through the erosion algorithm, the distance of each edge point is expanded by one pixel in all directions, and the outermost contour constitutes the expanded edge region. Since there may be missed detection of flying points after edge detection, the detection rate of flying points can be improved after expansion, ensuring the coverage of subsequent edge flying point processing.

[0088] In one embodiment, step S103 includes:

[0089] Based on the time bins of each pixel within the search area, cluster analysis is performed on the pixels within the search area to obtain several cluster centers.

[0090] The corresponding scenario category is determined based on the timebox of each cluster center.

[0091] In this embodiment, when performing cluster analysis on pixels in each search area, the time-bin map corresponding to the peak value is combined with the cluster analysis based on the time-bins of each pixel in the search area. Specifically, existing clustering algorithms such as K-means clustering and KNN clustering can be used for cluster analysis to obtain several cluster centers. Each cluster center represents a cluster, and each cluster includes several pixels. Pixels belonging to the same cluster center have similar time-bin characteristics. Therefore, after confirming the corresponding scene category based on the time-bins of each cluster center, the scene category of each pixel in the search area can be obtained. Specifically, the correspondence between time-bins and scene categories can be pre-set to achieve efficient scene category division.

[0092] Furthermore, prior to step S103, the method also includes:

[0093] The number of cluster centers is adjusted based on the number of foreground and background objects in the current scene.

[0094] In this embodiment, the number of cluster centers can be flexibly set based on the complexity of the scene, preferably 2-5 cluster centers. Taking a relatively simple scene, including only one target object and one background, as an example, 2 cluster centers can be set. That is, during cluster analysis, the pixels in the search area are divided into two categories. When performing edge flypoint correction processing, if the time box value of the pixel on the left side of the edge is 4 (classified as A), the time box value of the pixel on the right side of the edge is 2 (classified as B), and the time box value of a certain edge point in the edge region is 3.5 (classified as A), this edge point will be replaced by the time box value of the left pixel, which is also in class A, after cluster analysis, that is, it will be corrected to 4, thereby realizing edge flypoint correction processing and avoiding holes in the image filtering method. In more complex scenes, there may be more time box values, so they can be further divided into 3-5 categories based on the number of foreground and background objects to achieve more accurate edge flypoint processing.

[0095] It should be noted that there is no necessary order between the above steps. Those skilled in the art will understand from the description of the embodiments of the present invention that the above steps may have different execution orders in different embodiments, that is, they may be executed in parallel or in turn, etc.

[0096] The present invention also provides a corresponding edge flying point processing device for depth images, such as... Figure 4 As shown, Figure 4 This is a structural diagram of a depth image edge flying point processing device according to an embodiment of the present invention. It includes an edge detection module 401, a construction module 402, a clustering analysis module 403, and an edge correction module 404, which are electrically connected sequentially. The edge detection module 401 performs edge detection on the depth map to obtain edge regions, which include several edge points, including flying points. The construction module 402 constructs multiple corresponding search regions centered on each edge point on the timebox map corresponding to the peak value. The clustering analysis module 403 performs clustering analysis on all pixels within each search region to determine the scene category of each pixel within the search region. The edge correction module 404 obtains the scene category of an edge point within a search region and uses the timebox of the pixel closest to that edge point in the same category as the correction value for that edge point, and so on, to complete the correction of the timeboxes of all edge points. Since the edge flying point correction process has been described in detail in the above method embodiments, please refer to the corresponding method embodiments above for details, and it will not be repeated here.

[0097] In summary, this invention provides a method and apparatus for processing flying points on the edges of a depth image. The method includes: performing edge detection on the depth image to obtain edge regions, each edge region including several edge points, including flying points; constructing multiple search regions centered on each edge point on the time-box map corresponding to the peak value; performing cluster analysis on all pixels within each search region to determine the scene category of each pixel within the search region; obtaining the scene category of an edge point within a search region, and using the time-box of the pixel closest to that edge point in the same category as the correction value of that edge point, and so on, to complete the correction of the time-boxes for all edge points. By combining the time-box map corresponding to the peak value, constructing search regions for each edge point in the depth image, and performing cluster analysis to complete the time-box correction, the processing accuracy and edge clarity of flying point phenomena in direct time-of-flight depth images are effectively improved.

[0098] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A method for edge outlier processing of a depth image, the method comprising: The method comprises the following steps: Edge detection is performed on the depth map to obtain an edge region, the edge region comprising a plurality of edge points including the flying points, the depth map comprising depth values of all pixel points; A plurality of search regions are constructed with each edge point as the center on a time bin chart corresponding to the peak value, the time bin chart being constructed by obtaining histogram information of each pixel point and confirming a rough time bin corresponding to the peak value in the histogram of each pixel point, the rough time bin of each pixel point corresponding to the depth value thereof; Cluster analysis is performed on all pixel points in each search region to confirm a scene category of each pixel point in the search region; In a search region, the scene category of an edge point is obtained, and the time bin of the pixel point closest to the edge point in the same category is taken as the correction value of the edge point, and the correction of the time bins of all edge points is completed in this way.

2. The method of claim 1, wherein, The plurality of search regions are constructed with each edge point as the center, and the radius is N pixel intervals, wherein N is a positive integer; Each circle is circumscribed by a square region, and the square region is taken as the search region. The value range of N is 2-5.

3. The method of claim 2, wherein, The depth map is obtained by the following steps:

4. The method of claim 1, wherein, A rough histogram of each pixel point in a pixel array is obtained at a first resolution; A rough time of flight of each pixel point is calculated according to the rough histogram, and an exposure window region is confirmed according to the rough time of flight; A fine histogram of each pixel point in the pixel array is obtained at a second resolution in the exposure window region; A depth value of each pixel point is calculated according to the fine histogram, and a depth map is generated. After the edge detection on the depth map to obtain the edge region comprising a plurality of edge points including the flying points, the method further comprises:

5. The method of claim 1, wherein, Time bin correction of the edge points is performed on the rough histogram of each pixel point obtained in the pixel array at the first resolution; After the correction of the time bins of all edge points is completed, the method further comprises: A rough time of flight of each pixel point is calculated according to the corrected rough histogram, and an exposure window region is confirmed according to the rough time of flight; A fine histogram of each pixel point in the pixel array is obtained at a second resolution in the exposure window region; A depth value of each pixel point is calculated according to the fine histogram, and a fine depth map is generated.

6. The method of claim 1, wherein, After the edge detection on the depth map to obtain the edge region, the method further comprises: The edge region is expanded by an erosion algorithm.

7. The method of claim 1, wherein, The cluster analysis is performed on all pixel points in each search region to confirm the scene category of each pixel point in the search region, comprising: The pixel points in the search region are subjected to cluster analysis to obtain a plurality of cluster centers according to the time bins of the pixel points in the search region; The corresponding scene category is confirmed according to the time bins of the cluster centers.

8. The method of claim 7, wherein, Before the cluster analysis is performed on the pixel points in the search region to confirm the scene category of each pixel point in the search region, the method further comprises: The number of cluster centers is adjusted according to the number of foreground objects and background objects in the current scene.

9. The method of claim 7, wherein, The number of the cluster centers ranges from 2 to 5.

10. An apparatus for edge outlier processing of a depth image, the apparatus comprising: Comprise: An edge detection module, configured to perform edge detection on a depth map to obtain an edge region, the edge region comprising a plurality of edge points including a flying point, the depth map comprising depth values of all pixel points; A construction module, configured to construct a plurality of search regions corresponding to each edge point on a time bin map corresponding to a peak value, the time bin map being constructed by: obtaining histogram information of each pixel point, and confirming a rough time bin corresponding to the peak value in the histogram of each pixel point, the rough time bin of each pixel point corresponding to a depth value thereof; A cluster analysis module, configured to perform cluster analysis on all pixel points in each search region to confirm a scene category of each pixel point in the search region; An edge correction module, configured to obtain a scene category of an edge point in a search region, and take a time bin of a pixel point closest to the edge point in the same category as a correction value of the edge point, and in this way, correct time bins of all edge points.

Images