Indoor SLAM (Simultaneous Localization and Mapping) method based on Asia world and semantic information

Abstract

The invention discloses an indoor SLAM (Simultaneous Localization and Mapping) method based on an Asia world and semantic information, and belongs to the technical field of simultaneous localization and mapping. According to the method, an improved semantic segmentation network is combined, a plane instance mask with semantic information is generated, and plane parameters are extracted through a PCA (principal component analysis) method; plane matching is completed by the constant-speed tracking model and the plane parameters, and planes which do not conform to the Asia large structure are removed by using the gravity direction; and finally, detecting the Atritrate large coordinate system, completing matching of the Atritrate large coordinate system through a plane matching result, performing pose estimation by using a matching result, and correcting the Atritrate large coordinate system. The method can be applied to an actual physical environment, the influence of accumulative errors on visual odometer pose estimation in a weak texture environment is reduced, and robust and high-precision positioning information is provided.

Description

Technical Field

[0001] The present invention belongs to the technical field of simultaneous positioning and mapping, and specifically relates to a visual SLAM method applied to indoor environments. Background Art

[0002] With the rapid progress of artificial intelligence and computer technology, intelligent autonomous mobile robots have gradually become a key direction in the field of robotics research. Intelligent autonomous mobile robots combine a variety of intelligent technologies, among which simultaneous localization and mapping (SLAM) technology is the core of its value realization. SLAM technology provides robots with the ability to self-locate and map the environment in unknown environments, which is the basis for the smooth operation of intelligent robots.

[0003] Simultaneous Localization and Mapping (SLAM) refers to a technology that uses a subject equipped with specific sensors to estimate its own position in real time during movement and simultaneously build a map of the surrounding environment in the absence of prior information about the environment. According to the type of sensor used, SLAM algorithms are mainly divided into two categories: visual SLAM algorithms based on monocular cameras, RGB-D cameras, and binocular cameras, and laser SLAM algorithms based on lidar. In recent years, visual SLAM algorithms have been widely used in fields such as intelligent autonomous mobile robots and autonomous driving due to their low cost, rich environmental perception information, and high indoor positioning accuracy.

[0004] Most traditional visual SLAM algorithms are based on traditional point and line features. However, in indoor environments, weak textures will greatly reduce the number of feature points, resulting in irreversible cumulative errors in positioning, affecting the accuracy of visual odometer pose estimation. In recent years, in order to address the problem of poor positioning accuracy of visual SLAM algorithms in indoor weak texture environments, domestic and foreign scholars have used the method of detecting indoor Manhattan structures to extract planes that conform to the Manhattan structures in images, and converted inter-frame tracking into Manhattan coordinate tracking to avoid the impact of cumulative errors in indoor weak texture environments. However, this method has some limitations: (1) Traditional plane detection uses a hierarchical clustering method, which is significantly affected by sensor noise and may lead to incorrect plane detection. In addition, this method only uses information between pixels and is prone to misjudging curved surfaces as planes. (2) The increase in the number of Manhattan coordinate systems will lead to Manhattan coordinate system matching errors, resulting in reduced recognition accuracy. (3) The use of point features for mixed positioning will lead to the impact of cumulative errors.

[0005] Therefore, providing a visual SLAM method with strong robustness that can be applied to most indoor weak-texture scenes is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the invention

[0006] The main purpose of the present invention is to propose a visual SLAM method based on the Atlanta coordinate system for indoor weak-texture environments. The method can be applied to actual physical environments, reduce the impact of cumulative errors on visual odometer pose estimation in weak-texture environments, and provide robust and high-precision positioning information. The present invention aims to solve the following problems:

[0007] 1. Traditional visual SLAM algorithms have irreversible cumulative errors in positioning in indoor weak-texture environments;

[0008] 2. When there are too many Manhattan coordinate systems in the environment, the SLAM algorithm using Manhattan structure positioning will have the problem of reduced recognition accuracy.

[0009] 3. Using point features for mixed positioning will lead to cumulative errors.

[0010] The present invention proposes an RGB-D visual SLAM method based on the Atlanta structure for indoor weak-texture environments. The method combines an improved semantic segmentation network to generate a plane instance mask with semantic information, and extracts plane parameters through a PCA principal component analysis method; a uniform speed tracking model and plane parameters are used to complete plane matching, and the plane that does not conform to the Atlanta structure is removed using the gravity direction; finally, the Atlanta coordinate system is detected and the Atlanta coordinate system is matched through the result of plane matching, the matching result is used to perform pose estimation, and the Atlanta coordinate system is corrected.

[0011] The specific working process of the present invention is as follows:

[0012] Step 1: Use an RGB-D depth camera to capture standard indoor images and obtain continuous RGB images and depth image sequences;

[0013] Step 2: For the RGB image collected in step 1, a planar semantic image with a Manhattan structure is segmented using an improved Yolov8 planar instance segmentation network;

[0014] Step 3, combining the plane image obtained in step 2 with the depth image obtained in step 1, and using the PCA principal component analysis method to obtain the plane parameters;

[0015] Step 4, completing plane matching based on the uniform speed tracking model and plane parameters;

[0016] Step 5, obtain the gravity direction according to the semantic information and the parameters of the matching plane, and use the gravity direction to remove the Manhattan plane that does not conform to the Atlanta structure;

[0017] Step 6, perform orthogonal detection on the plane, detect all Atlanta coordinate systems, use the Atlanta coordinate system of the current frame to match the Atlanta coordinate system in the map, match to the key frame, and mark the Atlanta coordinate system status.

[0018] Step 7: Based on the key frame matched to the current frame, pose estimation and mapping are performed, and the pose of the current frame is used to correct the inaccurate Atlanta coordinate system.

[0019] The semantic segmentation network described in step 2 is an improved Yolov8-seg plane instance segmentation model, which can perform pixel-level instance segmentation on indoor planes. The specific steps are as follows:

[0020] To preprocess the RGB image, first adjust the image to a size of 640*640*3. Next, use OpenCV functions to normalize pixels and convert R and B color channels to ensure the uniformity and standardization of image data. The preprocessed image is passed to the backbone network of the Yolov8-seg instance segmentation network for feature extraction. The extracted features are passed to the multi-scale feature fusion module. The fused features are enhanced for attention. Through the channel attention module and the spatial attention module, the network pays more attention to the edge features of the plane. Finally, it is passed to the multi-scale instance segmentation module to generate two outputs output0 and output1. The first output output0 is 1*116*8400 in size. The first 84 items of 116 include cx, cy, w, h and the recognition accuracy scores of 80 plane semantic categories. cx and cy represent the center point coordinates of the target detection box, and w and h represent the width and height of the target detection box; the last 32 columns are used to calculate the plane mask data. The second output output1 is 1*32*160*160 in size. The mask data of the corresponding target is obtained by matrix multiplication of the last 32 fields of output0 and the data of output1, thereby obtaining the plane instance segmentation mask and plane semantic information. After filtering the planes using the plane semantic information, the planes with Manhattan structure are extracted, and the semantic information is saved as the planes of objects that conform to the Manhattan structure, such as ceilings, floors, walls, and desktops, and uncommon indoor planes such as road planes are deleted.

[0021] Step 3 is as follows:

[0022] The three-dimensional coordinates of the point cloud are obtained through the depth image and the camera intrinsic parameters. For each pixel point (u, v), its depth value d is obtained, and the three-dimensional coordinates (x, y, z) of the pixel point are calculated according to the camera intrinsic parameters (fx, fy, u0, v0), where fx is the vertical focal length, fy is the horizontal focal length, u0 is the horizontal coordinate of the image principal point, and v0 is the vertical coordinate of the image principal point, as shown in formula (1):

[0023]

[0024] z=d (1)

[0025] According to the filtered plane obtained in step 2, the three-dimensional point cloud coordinates in the depth image are obtained through the coordinates of the plane in the RGB image. The center point coordinates and the plane normal vector of the plane in the current frame are extracted using the PCA principal component analysis method. The three-dimensional coordinates of each point cloud are recorded as (x i ,y i ,z i ), each point is subtracted from the mean of the dimension to obtain the P matrix, as shown in formula (2):

[0026]

[0027] Calculate the covariance matrix: Use the centered coordinate matrix P to construct the covariance matrix C, as shown in formula (3):

[0028]

[0029] The covariance matrix C is decomposed to obtain the eigenvalues ​​and eigenvectors of the point cloud, where the eigenvector corresponding to the minimum eigenvalue is the normal vector n of the plane. c .

[0030] Step 4 is as follows:

[0031] The pose of the current frame is roughly estimated based on the uniform speed tracking model. The pose of the current frame estimated by the uniform speed tracking model is recorded as The displacement of the current frame is D. The coordinates of the center point of the plane in the current frame obtained in step 2 are M. c , the normal vector of the plane in the current frame is n c , the parameter representation of the calculation plane in the initial coordinate system is shown in formula (4):

[0032]

[0033] M w =M c +D (4)

[0034] Calculate the angle difference and plane distance difference between the current plane and the map plane, where n map is the normal vector of the map plane, d map is the vertical distance from the world origin to the map plane, as shown in formula (5):

[0035] △angle=n w ·n map

[0036] △d=d map +n w ·Mw (5)

[0037] When the angle difference is lower than the threshold of 0.05 and the plane distance difference is lower than 0.03 meters, the plane matching is considered successful.

[0038] Step 5 is as follows:

[0039] Match the plane marked as perpendicular to the gravity direction, or there is a plane with semantics as the ground in the current image, use this plane to extract the gravity direction, and multiply this gravity direction with all the planes in the current frame. If the result is less than 0.05, it means that the plane is perpendicular to the gravity direction, or the result is greater than 0.95, it means that the plane is parallel to the gravity direction. Therefore, this plane is a plane that conforms to the Atlanta structure, and the remaining planes are eliminated.

[0040] Step 6 is as follows:

[0041] All plane normal vectors are vertically detected by multiplying the normal vectors. If three mutually perpendicular normal vectors are detected, it means that these three normal vectors constitute the Atlanta coordinate system. It is queried whether the map planes matching these three planes belong to a certain key frame at the same time. If they do, it means that the match is successful. If there is a successfully matched Atlanta coordinate system, the Atlanta pose estimation is used, otherwise the point features are used for pose estimation.

[0042] There are three rules for marking the status of the Atlanta coordinate system: 1. In the period when SLAM positioning has not yet used point feature tracking, the status of the Atlanta coordinate system detected in all frames is set to accurate; 2. If point feature pose estimation has been performed, the Atlanta coordinate system encountered again is temporarily set to inaccurate. 3. If an accurate Atlanta coordinate system is matched, the status is also set to accurate.

[0043] Step 7 is as follows:

[0044] Use the current frame and the accurate Atlanta coordinate system in the map to estimate the rotation. Arrange the plane normal vectors in the xyz direction in n order. x 、n y 、n z , get the relative rotation matrix R of the current frame relative to the reference Atlanta coordinate system frame cy , and use singular value decomposition (SVD) to ensure its orthogonality. is the rotation matrix between the Atlanta coordinate system and the reference frame in which it appears, is the rotation matrix between the reference frame and the world coordinate system, and the current frame is rotated by R cw As shown in formula (6).

[0045] R cy =[n x ,n y,n z ]

[0046]

[0047] After completing the rotation estimation, the ORB feature points in the current frame are detected and matched using the ORB feature descriptor. The reprojection error measured by the camera is constructed through the matched ORB feature points. After fixing the rotation matrix, the translation estimation is performed by minimizing the reprojection error, as shown in formula (7).

[0048]

[0049] Where M is the number of point features, t cw is the displacement of the current frame, K is the camera intrinsic parameter matrix, P i is the three-dimensional coordinate of the feature point, Z is the z-axis value of the three-dimensional coordinate of the feature point, u i is the pixel coordinate of the feature point in the current image frame.

[0050] After completing the displacement and rotation estimation, use the inaccurate Atlanta coordinate system in this frame to correct it, and arrange the plane normal vectors in the order of xyz directions as n' x 、n' y 、n' z , get the relative rotation matrix R of the current frame relative to the inaccurate Atlanta coordinate system frame cy' , calculate the rotation matrix R of the inaccurate Atlanta coordinate system relative to the world coordinate system y'w , as shown in formula (8),

[0051] R cy' =[n' x ,n' y ,n' z ]

[0052]

[0053] And set the Atlanta coordinate system state to accurate. Get the world coordinates of all feature points through the current frame rotation matrix to complete the mapping task. Use the current rotation matrix R cw With the translation matrix t cw As the current pose, the positioning task is completed.

[0054] The present invention has the following advantages:

[0055] (1) The present invention improves the Yolov8-seg semantic segmentation network and significantly improves the detection accuracy of planes. (2) The present invention can be applied to most actual indoor scenes, and can screen planes based on semantic information, thus overcoming the recognition errors caused by the increase of Manhattan structures, and can effectively improve the positioning accuracy of SLAM algorithms in indoor weak-texture environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart of the RGB-D visual SLAM method applied to indoor weak texture environment of the present invention;

[0057] Figure 2 Planar instance mask image generated by semantic segmentation network inference results

[0058] a. RGB image

[0059] b. Planar instance mask image generated by the original yolov5 semantic segmentation network;

[0060] c. Plane instance mask image generated by the improved semantic segmentation network;

[0061] Figure 3 Average plane segmentation accuracy mAP50-95 before and after network improvement (95%>IoU>50%)

[0062] Figure 4 The present invention is applied to the positioning trajectory of the RGB-D visual SLAM method in indoor weak texture environment

[0063] a. True trajectory

[0064] b. Trajectory of the method proposed by the present invention DETAILED DESCRIPTION

[0065] In order to describe the technical solution of the invention in more detail, the present invention is further described below with reference to the accompanying drawings and examples.

[0066] like Figure 1 As shown, the specific steps of the present invention are as follows:

[0067] 1. Collect RGB images and depth images

[0068] The image acquisition device uses an RGB-D camera, which is calibrated and the parameters are written into the SLAM framework sensor parameter yaml file. The collected RGB image frame stream and depth image frame stream are transmitted to the computer in real time, and the tracking thread of the SLAM framework receives the image data released by the sensor.

[0069] 2. Segmentation and elimination of indoor plane instances

[0070] To preprocess the RGB image, first adjust the image to a size of 640*640*3. Next, use OpenCV functions to normalize pixels and convert R and B color channels to ensure the uniformity and standardization of image data. The preprocessed image is passed to the backbone network of the Yolov8-seg instance segmentation network for feature extraction. The extracted features are passed to the multi-scale feature fusion module. The fused features are enhanced for attention. Through the channel attention module and the spatial attention module, the network pays more attention to the edge features of the plane. Finally, it is passed to the multi-scale instance segmentation module to generate two outputs output0 and output1. The first output output0 is 1*116*8400 in size. The first 84 items of 116 include cx, cy, w, h and the recognition accuracy scores of 80 plane semantic categories. cx and cy represent the center point coordinates of the target detection box, and w and h represent the width and height of the target detection box; the last 32 columns are used to calculate the plane mask data. The second output output1 is 1*32*160*160 in size. The mask data of the corresponding target is obtained by matrix multiplication of the last 32 fields of output0 and the data of output1, thereby obtaining the plane instance segmentation mask and plane semantic information. After filtering the planes using the plane semantic information, the planes with Manhattan structure are extracted, and the semantic information is saved as the planes of objects that conform to the Manhattan structure, such as ceilings, floors, walls, and desktops, and uncommon indoor planes such as road planes are deleted.

[0071] 3. Calculate plane parameters

[0072] The three-dimensional coordinates of the point cloud are obtained through the depth image and the camera internal parameters, and the depth value d is obtained for each pixel point (i, j) i,j , according to the camera internal parameters (f x ,f y ,u0,v0) calculate the three-dimensional coordinates (x,y,z) of the pixel point, as shown in formula (1):

[0073]

[0074] z=d (1)

[0075] According to the filtered plane obtained in step 2, the three-dimensional point cloud coordinates in the depth image are obtained through the coordinates of the plane in the RGB image. The center point coordinates and the plane normal vector of the plane in the current frame are extracted using the PCA principal component analysis method. The three-dimensional coordinates of each point cloud are recorded as (x i ,y i ,z i ), each point is subtracted from the mean of the dimension to obtain the P matrix, as shown in formula (2):

[0076]

[0077] Calculate the covariance matrix: Use the centered coordinate matrix P to construct the covariance matrix C, as shown in formula (3):

[0078]

[0079] The covariance matrix C is decomposed to obtain the eigenvalues ​​and eigenvectors of the point cloud, where the eigenvector corresponding to the minimum eigenvalue is the normal vector n of the plane. c .

[0080] 4. Tracking the model matching plane at a constant speed

[0081] The pose of the current frame is roughly estimated based on the uniform speed tracking model. The pose of the current frame estimated by the uniform speed tracking model is recorded as The displacement of the current frame is D. The coordinates of the center point of the plane in the current frame obtained in step 2 are M. c , the normal vector of the plane in the current frame is n c , the parameter representation of the calculation plane in the initial coordinate system is n w , as shown in formula (4):

[0082]

[0083] M w =M c +D (4)

[0084] Calculate the angle difference and plane distance difference between the current plane and the map plane, where n map is the normal vector of the map plane, d map is the vertical distance from the world origin to the map plane, as shown in formula (5):

[0085] △angle=n w ·n map △d=d map +n w ·M w (5) When the angle difference is lower than the threshold of 0.05 and the plane distance difference is lower than 0.03 meters, the plane matching is considered successful.

[0086] 5. Eliminate planes in the direction of gravity

[0087] Match the plane marked as perpendicular to the gravity direction, or the plane with semantics as the ground in the current image, use this plane to extract the gravity direction, and multiply this gravity direction with all the planes in the current frame. If the size is close to 0, or close to 1, it means that the plane is perpendicular to the gravity direction, or the plane is parallel to the gravity direction. Therefore, this plane is a plane that conforms to the Atlanta structure, and the rest of the planes are eliminated.

[0088] 6. Atlanta coordinate system tracking

[0089] All plane normal vectors are vertically detected by multiplying the normal vectors. If three mutually perpendicular normal vectors are detected, it means that these three normal vectors constitute the Atlanta coordinate system. It is queried whether the map planes matching these three planes belong to a certain key frame at the same time. If they do, it means that the match is successful. If there is a successfully matched Atlanta coordinate system, the Atlanta pose estimation is used, otherwise the point features are used for pose estimation.

[0090] There are three rules for marking the status of the Atlanta coordinate system: 1. In the period when SLAM positioning has not yet used point feature tracking, the status of the Atlanta coordinate system detected in all frames is set to accurate; 2. If point feature pose estimation has been performed, the Atlanta coordinate system encountered again is temporarily set to inaccurate. 3. If an accurate Atlanta coordinate system is matched, the status is also set to accurate.

[0091] 7. Rotation and displacement estimation

[0092] Use the current frame and the accurate Atlanta coordinate system in the map to estimate the rotation. Arrange the plane normal vectors in the xyz direction in n order. x 、n y 、n z , get the relative rotation matrix R of the current frame relative to the reference Atlanta coordinate system frame cy , and use singular value decomposition (SVD) to ensure its orthogonality. is the rotation matrix between the Atlanta coordinate system and the reference frame in which it appears, is the rotation matrix between the reference frame and the world coordinate system, and the current frame is rotated by R cw As shown in formula (6).

[0093] R cy =[n x ,n y ,n z ]

[0094]

[0095] After completing the rotation estimation, the ORB feature points in the current frame are detected and matched using the ORB feature descriptor. The reprojection error measured by the camera is constructed through the matched ORB feature points. After fixing the rotation matrix, the translation estimation is performed by minimizing the reprojection error, as shown in formula (7).

[0096]

[0097] Where M is the number of point features, t cw is the displacement of the current frame, K is the camera intrinsic parameter matrix, P iis the three-dimensional coordinate of the feature point, and Z is the z-axis value of the three-dimensional coordinate of the feature point.

[0098] After completing the displacement and rotation estimation, the Atlanta coordinate system, which is inaccurate in this frame, is corrected using R cy' , get the rotation matrix R of the inaccurate Atlanta coordinate system relative to the world coordinate system cw , as shown in formula (8),

[0099] R cy' =[n' x ,n' y ,n' z ]

[0100]

[0101] And set the Atlanta coordinate system state to accurate. Get the world coordinates of all feature points through the current frame rotation matrix to complete the mapping task. Use the current rotation matrix R cw With the translation matrix t cw As the current pose, the positioning task is completed.

[0102] The above description of the specific implementation method of the present invention is only used for those skilled in the art to understand the present invention and is not used to limit the scope of the present invention.

Claims

1. An indoor SLAM method based on Atlanta world and semantic information, characterized in that, The steps include: Step 1: Use an RGB-D depth camera to capture standard indoor images and obtain continuous RGB images and depth image sequences; Step 2: For the RGB image collected in step 1, a planar semantic image with a Manhattan structure is segmented using an improved Yolov8 planar instance segmentation network; Step 3, combining the plane image obtained in step 2 with the depth image obtained in step 1, and using the PCA principal component analysis method to obtain the plane parameters; Step 4, completing plane matching based on the uniform speed tracking model and plane parameters; Step 5, obtain the gravity direction according to the semantic information and the parameters of the matching plane, and use the gravity direction to remove the Manhattan plane that does not conform to the Atlanta structure; Step 6, perform orthogonal detection on the plane, detect all Atlanta coordinate systems, use the Atlanta coordinate system of the current frame to match the Atlanta coordinate system in the map, and match it to the key frame; Step 7: Based on the key frame matched to the current frame, perform pose estimation and local mapping.

2. An indoor SLAM method based on Atlanta world and semantic information as claimed in claim 1, characterized in that, Step 2 The specific steps are as follows: The RGB image is passed into the backbone network of the Yolov8-seg instance segmentation network for feature extraction, and the extracted features are passed into the multi-scale feature fusion module. The fused features are enhanced for attention, and the planar target detection box is obtained through the channel attention module and the spatial attention module. Finally, it is passed into the multi-scale instance segmentation module to generate a plane instance segmentation mask and plane semantic information. The plane semantic information is used to filter the plane and then extract the plane with Manhattan structure.

3. An indoor SLAM method based on Atlanta world and semantic information as claimed in claim 1, characterized in that, The PCA principal component analysis method is used to extract the center point coordinates and plane normal vector of the plane in the current frame. The three-dimensional coordinates of all point clouds are recorded as (x i ,y i ,z i ), each point is subtracted from the mean of the dimension to obtain the P matrix, as shown in formula (1): Calculate the covariance matrix: Use the centered coordinate matrix P to construct the covariance matrix C, as shown in formula (2): Perform eigendecomposition on the covariance matrix C to obtain the eigenvalues ​​and eigenvectors of the point cloud, where the eigenvector corresponding to the minimum eigenvalue is the plane normal vector n c .

4. An indoor SLAM method based on Atlanta world and semantic information as claimed in claim 1, characterized in that, Step 4 is as follows: The pose of the current frame is roughly estimated based on the uniform speed tracking model. The pose of the current frame estimated by the uniform speed tracking model is recorded as The displacement of the current frame is D, and the coordinates of the center point of the plane in the current frame obtained in step 2 are M c , the normal vector of the plane in the current frame is n c , the parameter representation of the calculation plane in the initial coordinate system is shown in formula (4): M w =M c +D (4) Calculate the angle difference and plane distance difference between the current plane and the map plane, where n map is the normal vector of the map plane, d map is the vertical distance from the world origin to the map plane, as shown in formula (5): △angle=n w ·n map △d=d map +n w ·M w (5) When the angle difference is lower than the threshold of 0.05 and the plane distance difference is lower than 0.03 meters, the plane matching is considered successful.

5. An indoor SLAM method based on Atlanta world and semantic information as claimed in claim 1, characterized in that, Step 5 is as follows: Match the plane marked as perpendicular to the gravity direction, or there is a plane with semantics as the ground in the current image, use this plane to extract the gravity direction, and multiply this gravity direction with all the planes in the current frame. If the result is less than 0.05, it means that the plane is perpendicular to the gravity direction, or the result is greater than 0.95, it means that the plane is parallel to the gravity direction. Therefore, this plane is a plane that conforms to the Atlanta structure, and the remaining planes are eliminated.

6. An indoor SLAM method based on Atlanta world and semantic information as claimed in claim 1, characterized in that, Step 6 is as follows: All plane normal vectors are vertically detected by multiplying the normal vectors. If three mutually perpendicular normal vectors are detected, it means that these three normal vectors constitute the Atlanta coordinate system. It is queried whether the map planes matching these three planes belong to a certain key frame at the same time. If there are multiple key frames, the earliest key frame in the time series is taken out.

7. An indoor SLAM method based on Atlanta world and semantic information as claimed in claim 1, characterized in that, Step 7 is as follows: The relative rotation relationship of the current frame with respect to the key frame is obtained by making a rotation estimate between the Atlanta coordinate system in the key frame and the Atlanta coordinate system of the current frame, and then the pose of the current frame with respect to the world coordinate system is obtained through the pose of the key frame with respect to the world coordinate system.

Images