A dynamic object removal method based on average blur

Abstract

The present invention proposes a method for removing dynamic objects based on average blur. First, through image preprocessing, including correction and denoising, an image pyramid is constructed and ORB feature points are extracted. Secondly, the regional merging result of each layer is obtained by using the regional merging operation based on the image pyramid and average blur, and multiple layers are fused to generate the regional merging result of the current frame. Thirdly, the RANSAC algorithm result is optimized by the improved RANSAC algorithm, and dynamic and static region identification is performed to generate a dynamic region mask. Post-processing operations are performed to remove dynamic objects in the current frame using the dynamic region mask, and camera pose recovery, local mapping and closed-loop detection operations are performed on the remaining static regions. The present invention effectively uses the local texture information of the image to process the image. In the process of removing dynamic objects, the limitation that the network can only target known objects is avoided, and the applicable scenarios of the dynamic SLAM system when facing dynamic targets are expanded.

Description

Technical Field

[0001] The invention belongs to the technical field of image processing, and in particular relates to a dynamic object removal method based on average blur. Background Art

[0002] At present, the mainstream solution of dynamic SLAM is to use target detection or semantic segmentation network to obtain the area of ​​moving objects and remove the feature points on dynamic objects, so as to remove the adverse effects of the feature points on dynamic objects on the positioning system in the subsequent system. This improves the positioning accuracy, but also obviously sacrifices the real-time performance of the system. DS-SLAM, SLAM++ and DynaSLAM are the current representative solutions. They use semantic segmentation and RANSAC algorithms to remove feature points on dynamic areas, but they are generally time-consuming.

[0003] SLAM++ uses RGB-D input, detects objects using a 3D object detector, and uses the mapping results generated by SLAM. DynaSLAM is a dynamic stereoscopic vision SLAM system that can simultaneously and densely reconstruct static backgrounds, moving objects, and objects that may move but are currently stationary, but it is very time-consuming. Other solutions, such as RDS-SLAM, improve real-time performance through independent semantic threads and mobile probability transfer, but the problem of long time consumption for frames that need semantic processing still exists.

[0004] Comparative document CN115588133A is a visual SLAM method suitable for dynamic environments. The application discloses a visual SLAM method suitable for dynamic environments, including: extracting feature points from an input image, and dividing the input image into an area inside a frame and an area outside a frame; filtering out moving points outside the frame through geometric constraints to obtain static points outside the frame for the feature points in the area outside the frame; calculating the speed of the feature points in the area inside the frame to divide them into moving points inside the frame and static points inside the frame; and incorporating the static points inside the frame and the static points outside the frame into pose solution to obtain the optimal pose.

[0005] The segmentation network and target recognition network used in the above patents can usually only recognize objects of a specified category. However, in actual scenarios, this is obviously difficult to cover all situations. For example, the network can remove the impact of ground vehicle occlusion, but it is difficult to remove the impact of shadows moving with it or animals that are not in the specified category. The method proposed in this case can uniformly and effectively eliminate this impact. Summary of the invention

[0006] In order to solve the problems existing in the background technology, the purpose of the present invention is to provide a dynamic object removal method based on average blur, which effectively utilizes the local texture information of the image to process the image. In the process of dynamic object removal, the use of deep learning networks is avoided, and the limitation that the network can only target known objects is also avoided, thereby improving the real-time and universality of the dynamic SLAM system when facing dynamic targets.

[0007] A dynamic object removal method based on average blur includes the following steps:

[0008] S1: performing image preprocessing operations, correcting the image, including but not limited to lens distortion correction, and denoising the image; performing ORB feature point extraction operations, scaling the image N-1 times, and forming an N-layer image pyramid, where N is an integer greater than 1; and then segmenting each layer of the image pyramid through quadtree segmentation to obtain a segmented image region;

[0009] S2: Perform image region merging operation, use the image pyramid to perform region division and region merging operation based on average blur on each layer of the image, obtain the region merging result of each layer, fuse all layers in the image pyramid, obtain the region merging result of the current frame, and thus obtain the ORB feature point list contained in the region merging result;

[0010] S3: Dynamic region mask generation operation, performing improved RANSAC matching on the current frame and the previous frame, using the RANSAC matching results and region segmentation results, performing dynamic and static region recognition and dynamic mask generation operations, judging the dynamic and static nature of the image region based on the number of ORB feature points in the region, and generating a dynamic region mask based on this;

[0011] S4: Perform post-processing operations, use the dynamic area mask to remove dynamic objects in the current frame, and perform camera pose recovery, local mapping and closed-loop detection operations on the remaining static area.

[0012] Further, S2 comprises the following steps:

[0013] S21: using the image pyramid obtained in S1, performing an image region merging operation based on average blur on the image of each layer therein, so as to obtain a region merging result corresponding to the N layer images in the image pyramid;

[0014] S22: restore the size of the region merging result of each level in the image pyramid to the original image size with the center of the image pyramid as the origin;

[0015] S23: using the different information contained in each level of the image pyramid, a fusion operation is performed on the boundary of the image region, and after the merging is completed, a region merging result of the current frame image is obtained;

[0016] S24: Associate all ORB feature points to corresponding regions according to their coordinate positions, and obtain a list of ORB feature points included in the current frame region merging result.

[0017] Further, the image region merging operation based on average blur in S21 includes the following steps:

[0018] For any level image of the image pyramid, a total of k regions are processed as the result of quadtree segmentation, where k is a positive integer;

[0019] SS1: Perform average blur on k regions in the image of the current layer to obtain the average pixel value averageValue of each region num , num=1,2,...,k;

[0020] SS2: Take the average pixel value averageValue of k regions num The difference between the maximum and minimum values ​​of is the average pixel value range R, and one tenth of the average pixel value range R is taken as the threshold Y;

[0021] SS3: Sort the sizes of the image segmented regions and obtain a region size list List(A);

[0022] SS4: Take the largest area A in the current list List(A) that has not been merged. i ;

[0023] SS5: Count the number of its adjacent areas as m, and take its adjacent area A in clockwise direction ij ;

[0024] SS6: For Area A i and A ij The absolute value of the average pixel value difference is calculated and recorded as D Ai,Aij ;

[0025] SS7: Judgement D Ai,Aij Is it less than the threshold Y? If it is less than the threshold, the current area A ij Mark as area to be merged;

[0026] SS8: Area A i Merge with all the regions to be merged obtained in the previous step to obtain a new region, which is updated to A i , and recalculate area A i Average pixel value after average blur averageValue(A i);

[0027] SS9: Update area size list List(A), update area A i The range and parameters of the object and delete the surrounding area merged by it;

[0028] SS10: Repeat the operations from SS4 to SS9 for the expanded area until area A i There are no surrounding areas that meet the merging conditions, so area A i Mark as merged region and update region size list List(A);

[0029] SS11: Repeat steps SS4 to SS10 for the largest unmerged area in the current area list List(A) until all areas cannot be merged with other areas, and obtain the range map AM of all areas in the current layer. n , n=1,2,...,N.

[0030] Furthermore, the improved RANSAC matching in S3 includes the following steps:

[0031] S31: Perform RANSAC consistency detection on the current frame and the previous frame using all extracted ORB feature points to obtain the main direction of inter-frame motion and a list of internal and external points;

[0032] S32: Calculate the ratio of the number of inner points to the number of all ORB feature points, innerRate, and the ratio of the number of inner points to the number of outer points, ioRate;

[0033] S33: Count the number of inner points and outer points in each area, and calculate the ratio of the number of inner points to the number of all ORB feature points in the area, innerRate i , and the ratio of the number of internal points to the number of external points ioRate i ;

[0034] S34: For each region, if the innerRate of the region i ≥ innerRate, and ioRate i ≥ioRate, the area is considered to be a high-quality area;

[0035] S35: Sort the number of inliers in all high-quality areas from most to least, select 80% of the ORB feature points of the total number of front inliers as the range of the second RANSAC calculation, perform the second RANSAC calculation, and obtain the main direction of the current frame.

[0036] Further, the dynamic and static area identification in S3 includes the following steps:

[0037] N1: Perform dynamic and static region identification on all regions in the current frame, and identify the parts of the identified high-quality regions mentioned in S34 as static regions;

[0038] N2: In the RANSAC matching process, the area without any successfully matched ORB feature points is considered to be a dynamic area;

[0039] N3: For the regions with successfully matched ORB feature points but not selected for calculation in N2, recalculate the ORB feature points and the RANSAC main direction of these regions; if the main direction of the region is inconsistent with the main direction determined in S35, it is considered to be a dynamic region, otherwise it is considered to be a static region;

[0040] N4: All dynamic areas in N3 are merged into one area according to the area boundary of the current frame, and this area is the dynamic area mask of the current frame.

[0041] The beneficial effects of the present invention are as follows:

[0042] The present invention utilizes information extracted from ORB features and the image itself in a visual SLAM task, and performs region segmentation on the image based on ORB feature points, thereby avoiding the use of a complex deep learning network when dynamic SLAM performs tasks of dynamic boundary determination, and effectively improving the real-time performance of the dynamic SLAM system; at the same time, the present invention fully utilizes the process of the visual SLAM system, effectively combines the features required for positioning with the features required for object segmentation, and optimizes the RANSAC process, thereby improving the accuracy of motion recovery and mapping.

[0043] The present invention effectively utilizes the local texture information of the image to process the image. In the process of removing dynamic objects, it avoids the limitation that the network can only be used for known objects, and expands the applicable scenarios of the dynamic SLAM system when facing dynamic targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a flow chart of the method for removing dynamic objects based on average blur of the present invention;

[0045] Figure 2 It is a flow chart of the region merging operation based on image pyramid and average blur in the present invention;

[0046] Figure 3 This is a flowchart of the image region merging operation based on average blur in the present invention;

[0047] Figure 4 It is a flow chart of the region-based RANSAC algorithm optimization operation in the present invention;

[0048] Figure 5This is a flowchart of the dynamic and static area recognition and dynamic mask generation operation in the present invention;

[0049] Figure 6 This is a schematic diagram of the current frame after preprocessing;

[0050] Figure 7 A schematic diagram for extracting an image pyramid for the current frame;

[0051] Figure 8 This is a schematic diagram of the result after ORB feature point matching between the current frame and the previous frame;

[0052] Fig. 9 It is a schematic diagram of the dynamic area of ​​the current frame;

[0053] Fig.10 A frame of image input from the KITTI dataset;

[0054] Fig.11 It is the quadtree segmentation result of one level in the image pyramid of the KITTI dataset image;

[0055] Fig.12 It is the result of the average blur operation at one level in the image pyramid of the KITTI dataset image;

[0056] Fig.13 This is the dynamic area mask result finally obtained for the KITTI dataset image. DETAILED DESCRIPTION

[0057] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0058] A dynamic object removal method based on average blur, such as Figure 1 As shown, the following steps are included:

[0059] S1: Perform image preprocessing operations to obtain the captured image from the camera for processing, including correcting the image using the camera parameters in the YAML format configuration file, where the image correction includes but is not limited to lens distortion correction; then perform Gaussian blur denoising on the image to prevent noise from interfering with the extraction of image features. The image obtained after preprocessing is as follows Figure 6 As shown, the objects with orange edges are dynamic objects; Figures 10 to 13 The image processing effect of the input KITTI dataset is shown. The acquired image is scaled N-1 times with a certain scaling factor to form an N-layer image pyramid with image sizes from small to large, such as Figure 7As shown, wherein N is an integer greater than 1;

[0060] In order to extract an appropriate number of ORB feature points from each layer of the image, the number of ORB feature points that need to be extracted from each layer of the image is calculated in descending order, and then quadtree segmentation is used in each layer to extract ORB feature points and segment the image area. The quadtree segmentation result of one level in the image pyramid is as follows: Fig.11 This process ensures that each area of ​​each layer has at most one ORB feature point, as shown in Figure 8 shown.

[0061] Different from the common visual SLAM module process, the present invention needs to perform a merging operation based on the image pyramid and the average blurred image area on the image pyramid from which the ORB feature points are obtained. The specific operation is as follows:

[0062] S2: Perform image region merging operation, use the image pyramid to perform region division and region merging operation based on average blur on each layer of the image, obtain the region merging result of each layer, fuse all layers in the image pyramid, obtain the region merging result of the current frame, and thus obtain the ORB feature point list contained in the region merging result;

[0063] S3: Dynamic region mask generation operation, performing improved RANSAC matching on the current frame and the previous frame, using the RANSAC matching results and region segmentation results, performing dynamic and static region recognition and dynamic mask generation operations, judging the dynamic and static nature of the image region based on the number of ORB feature points in the region, and generating a dynamic region mask based on this;

[0064] S4: Perform post-processing operations, use the dynamic area mask to remove dynamic objects in the current frame, and perform camera pose recovery, local mapping and closed-loop detection operations on the remaining static areas. The specific steps are as follows:

[0065] Removing dynamic objects: All ORB feature points in the current frame are judged using the dynamic area mask obtained by S3. ORB feature points that fall within the dynamic area mask are removed, and ORB feature points that do not fall within the dynamic area mask are retained.

[0066] Camera pose recovery: Use the retained ORB feature points to restore the camera pose of the current frame relative to the previous frame. If tracking is successful, update the map and maintain the keyframe database; if tracking fails, the system attempts to relocate to resume tracking. Determine whether the current frame is created as a keyframe based on the degree of change in the current environment and the length of the time interval.

[0067] Local Mapping: Manage existing keyframes and update map points, add new map points brought in by new keyframes, and remove low-quality map points. Optimize the map points of keyframes and local maps to ensure the accuracy of map points.

[0068] Closed loop detection: Merge the local map with the global map and update the map structure. Monitor the similarity between the current scene and the existing scenes in the map to identify potential closed loop opportunities. When a closed loop is detected, confirm the closed loop by matching the features of the current keyframe with the historical keyframes. Use all the information currently forming the closed loop for global optimization to correct the accumulated errors in the map. Merge the local map with the global map and update the map structure to maintain the high accuracy and robustness of the visual SLAM system.

[0069] Wherein S2 includes the following steps, such as Figure 2 shown.

[0070] S21: using the image pyramid obtained in S1, performing an image region merging operation based on average blur on the image of each layer therein, so as to obtain a region merging result corresponding to the N layer images in the image pyramid;

[0071] S22: restore the size of the region merging result of each level in the image pyramid, and restore it to the original image size with the center of the image pyramid as the origin; the specific method is: restore the size of the region range map AM1~AM n According to the corresponding zoom scale, restore with the center as the origin to obtain the area range map AM1~AM of the original image size n .

[0072] S23: using the different information contained in each level of the image pyramid, a fusion operation is performed on the boundary of the image region, and after the fusion is completed, a region fusion result of the current frame image is obtained; wherein the fusion operation is performed as follows:

[0073] First, take the AM1 layer and AM n (n=2,3,4,...,N) layers, compare and merge the area range maps of the two layers to AM n The layer is used as the basis for optimizing the region and updating the region boundary. Secondly, for a smaller region A whose area is less than or equal to 1 / 64 of the current frame area i , when this region exists only in the AM1 layer, and in the AM n If the layer is completely covered by the adjacent area, the small area is considered to be part of the covered area and its area is merged with the covered area. i , the current area division will not be changed, but will be divided according to the AM1 area boundary. i(i=1,2,3,...,k) to traverse and update the processed regional results to AM1. Finally, loop until the AM1 layer and AM N layer, and after the update is completed, the area range map AM1 is stored in the final area range map AM.

[0074] S24: All ORB feature points are counted by region according to the final area range map AM to obtain the current frame area merging result, and obtain the ORB feature point list List (A, List (ORBpoints)) corresponding to each region.

[0075] Further, the image region merging operation based on average blur in S21 is as follows: Figure 3 As shown, the following steps are included:

[0076] For any level image of the image pyramid, a total of k regions are processed as the result of quadtree segmentation, where k is a positive integer;

[0077] SS1: For the image of the current layer n (n=1,2,3,...,N) of the image pyramid, the quadtree segmentation result obtained during the ORB feature point extraction process for the layer is used as the processing object, and a total of k image regions are processed. Average blur is performed on each quadtree segmented image region to obtain the average pixel value averageValue of each region num , num=1,2,...,k. The specific method is: in each area obtained by quadtree segmentation, the average pixel values ​​of all pixels are added and averaged, and then this average value is applied to all pixels in the block. The processing result of the input image is as follows Fig.12 shown.

[0078] SS2: Take the average pixel value averageValue of k regions num The difference between the maximum and minimum values ​​of is the average pixel value range R, and one tenth of the average pixel value range R is taken as the threshold Y;

[0079] SS3: Considering that the average pixel value of the area with more pixels is more stable and more representative, the sizes of the image segmentation areas are sorted and the area size list List(A) is obtained;

[0080] SS4: Merge the surrounding areas of the unmerged areas in List(A) in descending order. Take the largest area A in the current list List(A) that has not been merged. i .

[0081] SS5: For Area A i , count the number of its adjacent regions as m, and take its adjacent region A in clockwise directionij ;

[0082] SS6: For Area A i and A ij Region, the average pixel value calculated in the two regions is subtracted and the absolute value is obtained to obtain D Ai,Aij D Ai,Aij =|averageValue(A i )- averageValue(A ij )|,

[0083] Judgement D Ai,Aij Is it less than the threshold Y? If it is less than the threshold, the current area A ij Mark as the area to be merged. If it is greater than the threshold, no operation is performed.

[0084] SS7: Judgement D Ai,Aij Is it less than the threshold Y? If it is less than the threshold, the current area A ij Mark as area to be merged;

[0085] SS8: Area A i Merge with all the areas to be merged obtained by SS7 to obtain the new area A i ' . And let A i =A i ' . Recalculate area A i Average pixel value after average blur averageValue(A i ) to replace the previous value.

[0086] SS9: Update area size list List(A), update area A i range and parameters and remove the image area merged by it from the list;

[0087] SS10: Repeat the operations from step SS4 to step SS9 for the enlarged image area until area A i If there is no area that meets the merging conditions, update the area size list List(A) and merge area A i Mark the region as merged to exclude it from the next region merge.

[0088] SS11: Repeat steps SS4 to SS10 for the largest unmerged region in the current region list List(A) until all regions cannot be merged with other regions, and obtain the range map AM of all regions. n .

[0089] Different from the usual visual SLAM module process, this system does not directly use the results of the RANSAC algorithm, but performs region-based RANSAC algorithm optimization operations based on the current results, such as Figure 4 As shown in Figure 3, the improved RANSAC matching in S3 includes the following steps:

[0090] S31: Use all extracted ORB feature points to perform RANSAC consistency detection on the current frame and the previous frame to obtain the main direction of inter-frame motion and the list of internal and external points, such as Figure 8 As shown in the figure, green indicates the points that are successfully matched, and red indicates the points that fail to match. The points that are successfully matched on dynamic objects will cause interference and errors to the visual SLAM system. Figure 8 The green ORB feature points on the vehicles and pedestrians are the ORB feature points that need to be removed, but the RANSAC algorithm cannot remove them. These ORB feature points will cause errors to the visual SLAM system.

[0091] S32: Using the number of internal and external points and the main direction counted by RANSAC, calculate the ratio of the number of internal points to all ORB feature points in the entire frame, innerRate, and the ratio of the number of internal points to the number of external points, ioRate.

[0092] S33: Count the number of inliers and outliers in each image region. i The ratio of the number of inner points to the number of all ORB feature points in the area innerRate i , and the ratio of the number of internal points to the number of external points ioRate i .

[0093] S34: For each region, if the innerRate of the region i ≥ innerRate, and ioRate i ≥ioRate, the area is considered to be a high-quality area.

[0094] S35: Sort the number of inliers in all high-quality areas from most to least, select 80% of the ORB feature points of the total number of front inliers as the range of the second RANSAC calculation, perform the second RANSAC calculation, and obtain the main direction of the current frame.

[0095] So far, the main direction of the current frame has been obtained, and the dynamic and static area recognition in S3 is as follows: Figure 5 As shown, the specific steps are as follows:

[0096] N1: Perform dynamic and static region identification on all regions in the current frame, and identify the parts of the identified high-quality regions mentioned in S34 as static regions;

[0097] N2: In the RANSAC matching process, the area without any successfully matched ORB feature points is considered to be a dynamic area;

[0098] N3: For the regions with successfully matched ORB feature points but not selected in N2, recalculate the ORB feature points and the RANSAC main direction of these regions, where the main direction is the RANSAC calculation result; if the main direction of the region is inconsistent with the main direction determined in S35, it is considered to be a dynamic region, otherwise it is considered to be a static region;

[0099] N4: All dynamic regions in N3 are merged into one region according to the region boundary of the current frame, and this region is the dynamic region mask of the current frame. The schematic diagram of the dynamic region mask of the current frame is as follows: Fig. 9 As shown in FIG. 1 , the orange shaded area is the dynamic area mask. The method designed by the present invention removes the points on the dynamic objects in the image that are not removed by the RANSAC algorithm, thereby removing the interference caused by the dynamic objects. Fig.13 As shown, the dynamic area mask result obtained for the KITTI dataset image.

[0100] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A dynamic object removal method based on average blur, characterized in that: The following steps are involved: S1: performing image preprocessing operations, correcting the image, including but not limited to lens distortion correction, and denoising the image; performing ORB feature point extraction operations, scaling the image N-1 times, and forming an N-layer image pyramid, where N is an integer greater than 1; and then segmenting each layer of the image pyramid through quadtree segmentation to obtain a segmented image region; S2: Perform image region merging operation, use the image pyramid to perform region division and region merging operation based on average blur on each layer of the image, obtain the region merging result of each layer, fuse all layers in the image pyramid, obtain the region merging result of the current frame, and thus obtain the ORB feature point list contained in the region merging result; S3: Performing a dynamic region mask generation operation, performing an improved RANSAC match on the ORB feature points in the current frame and the previous frame image, using the RANSAC matching results and the region segmentation results, performing dynamic and static region recognition and dynamic mask generation operations, and making dynamic and static judgments on the image region based on the number of ORB feature points in the region, and thereby generating a dynamic region mask; S4: Perform post-processing operations, use the dynamic area mask to remove dynamic objects in the current frame, and perform camera pose recovery, local mapping and closed-loop detection operations on the remaining static areas; The improved RANSAC matching in S3 includes the following steps: S31: Perform RANSAC consistency detection on the current frame and the previous frame using all extracted ORB feature points to obtain the main direction of inter-frame motion and a list of internal and external points; S32: Calculate the ratio of the number of inner points to the number of all ORB feature points, innerRate, and the ratio of the number of inner points to the number of outer points, ioRate; S33: Count the number of inner points and outer points in each area, and calculate the ratio of the number of inner points to the number of all ORB feature points in the area, innerRate i , and the ratio of the number of internal points to the number of external points ioRate i , i=1,2,...,k; S34: For each region, if the innerRate of the region i ≥ innerRate, and ioRate i ≥ioRate, the area is considered to be a high-quality area; S35: Sort the number of inliers in all high-quality areas from large to small, select 80% of the ORB feature points of the total number of inliers in turn, perform the second RANSAC calculation and obtain the main direction of the current frame.

2. The method for removing dynamic objects based on average blur according to claim 1, characterized in that: The S2 comprises the following steps: S21: using the image pyramid obtained in S1, performing an image region merging operation based on average blur on the image of each layer therein, so as to obtain a region merging result corresponding to the N layer images in the image pyramid; S22: restore the size of the region merging result of each level in the image pyramid to the original image size with the center of the image pyramid as the origin; S23: using the different information contained in each level of the image pyramid, a fusion operation is performed on the boundary of the image region, and after the merging is completed, a region merging result of the current frame image is obtained; S24: Associate all ORB feature points to corresponding regions according to their coordinate positions, and obtain a list of ORB feature points included in the current frame region merging result.

3. The method for removing dynamic objects based on average blur according to claim 2, characterized in that: The image region merging operation based on average blur in S21 includes the following steps: For any level image of the image pyramid, a total of k regions are processed as the result of quadtree segmentation, where k is a positive integer; SS1: Perform average blur on k regions in the image of the current layer to obtain the average pixel value averageValue of each region num , num=1,2,...,k; SS2: Take the average pixel value averageValue of k regions num The difference between the maximum and minimum values ​​of is the average pixel value range R, and one tenth of the average pixel value range R is taken as the threshold Y; SS3: Sort the sizes of the image segmented regions and obtain a region size list List(A); SS4: Take the largest area A in the current list List(A) that has not been merged. i , i=1,2,...,k; SS5: Count the number of its adjacent areas as m, and take its adjacent area A in clockwise direction ij , j = 1, 2, ..., m; SS6: For Area A i and A ij The absolute value of the average pixel value difference is calculated and recorded as D Ai,Aij ; SS7: Judgement D Ai,Aij Is it less than the threshold Y? If it is less than the threshold, the current area A ij Mark as area to be merged; SS8: Area A i Merge with all the regions to be merged obtained in the previous step to obtain a new region, which is updated to A i , and recalculate area A i Average pixel value after average blur averageValue(A i ); SS9: Update area size list List(A), update area A i The range and parameters of the object and delete the surrounding area merged by it; SS10: Repeat the operations from SS4 to SS9 for the enlarged image area until area A i If there is no area around that meets the merging conditions, update the area size list List(A) and merge area A i Mark as merged area to exclude it from the next area merge; SS11: Repeat the operations from SS4 to SS10 for the largest unmerged area in the current area list List(A) until all areas cannot be merged with other areas, and obtain the range map AM of all areas in the current layer. n , n=1,2,...,N.

4. The method for removing dynamic objects based on average blur according to claim 1, characterized in that: The dynamic and static area identification in S3 includes the following steps: N1: Perform dynamic and static region recognition on all regions in the current frame, and identify the parts that are identified as high-quality regions in S34 as static regions; N2: In the RANSAC matching process, the area without any successfully matched ORB feature points is considered to be a dynamic area; N3: For the regions with successfully matched ORB feature points but not selected for calculation in N2, recalculate the ORB feature points and the RANSAC main direction of these regions; if the main direction of the region is inconsistent with the main direction determined in S35, it is considered to be a dynamic region, otherwise it is considered to be a static region; N4: All dynamic areas in N3 are merged into one area according to the area boundary of the current frame, and this area is the dynamic area mask of the current frame.

5. The method for removing dynamic objects based on average blur according to claim 1, characterized in that: When performing image pyramid segmentation in S1, each region of each layer of the image has at most one ORB feature point.

Images