A visual inertial positioning method based on sub-pixel feature merging and optimization

Abstract

The present invention discloses a visual inertial positioning method based on sub-pixel feature merging optimization. For the two frames of images, a dual-thread extraction method of point and line features is adopted: the sub-pixel point feature extraction algorithm is used to perform feature detection on the image to obtain sub-pixel corner points; sub-image-level line features are obtained by line feature extraction, so that the point and line feature extraction accuracy reaches the sub-pixel level. The obtained point and line features are matched to obtain the inter-frame pose of two adjacent frames of images, and are fused with the carrier pose information obtained by inertial device measurement to obtain the position information of the carrier. Due to the large amount of line feature data, the system backend optimization time is long, so the repeated line features are merged in the sliding window to reduce the dimension of the state quantity to be estimated and increase the proportion of effective line features. The present invention reduces the hardware facility cost required for the algorithm, reduces the amount of calculation required for optimization, improves the running speed, and makes the positioning accuracy more accurate.

Description

Technical Field

[0001] The present invention relates to the field of visual inertial positioning and navigation technology, and in particular to a visual inertial positioning method based on sub-pixel point and line feature fusion. Background Art

[0002] With the rapid development of information science, automatic control and other disciplines, the autonomous unmanned equipment industry has flourished, which can better meet the growing industrial needs and gradually achieve commercial use. Visual sensors have the advantages of rich image information, low price, lightness and simplicity, and the ability to integrate with a variety of sensors, which has enabled the visual positioning and orientation navigation system based on visual sensors to develop rapidly, and a large number of theories on visual SLAM positioning and mapping systems have emerged.

[0003] At present, autonomous unmanned equipment usually uses two or more sensors at the same time, and realizes information fusion through the complementary advantages of each sensor, so as to ensure the strong robustness of the system. Among them, the fusion algorithm of vision and inertial measurement unit data has become a popular research direction because of its complementary advantages. The combination of vision and inertial system can compensate for the zero bias and drift of low-level inertial measurement units through image information, and correct the offset of inertial devices in time. The inertial measurement unit can ensure that the carrier's posture information is not interfered by external factors through accelerometers and gyroscopes. However, in environments with drastic changes in lighting or sparse features, it is difficult for a single point feature visual positioning system to obtain accurate image features. During a certain period of time when the algorithm is running, it is impossible to obtain high-precision inter-frame posture information by relying solely on inertial devices, which makes the algorithm positioning inaccurate. Summary of the invention

[0004] In response to the above-mentioned problems, the present invention aims to provide a visual inertial positioning method based on sub-pixel point and line feature fusion, which can improve the system positioning accuracy while ensuring the operation efficiency of the algorithm to address the problems of insufficient number of pixel-level point features and incomplete information in low-texture environments.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A visual inertial positioning method based on sub-pixel feature merging optimization, characterized in that it comprises the following steps:

[0007] S1: Preprocess the camera image to obtain a grayscale image as the input image of the visual system;

[0008] S2: Extract point features of the input image using a multi-level pixel feature detection algorithm;

[0009] S3: while performing point feature extraction in step S2, perform line feature extraction on the input image;

[0010] S4: Fuse the point features extracted in step S2 and the line features extracted in step S3 to achieve visual inertial positioning.

[0011] Furthermore, the specific operation of step S2 includes the following steps:

[0012] S201: Performing a first round of corner point detection on the input image using a pixel-level feature detection method;

[0013] S202: performing secondary screening on the detected corner points using a sub-pixel corner point detection method;

[0014] S203: performing image edge constraint on the sub-pixel corner point detection result, and removing sub-pixel corner points that are not at the edge of the image;

[0015] S204: Use the LK optical flow method to perform feature matching on the sub-pixel corner points finally obtained in step S203 to obtain the inter-frame pose of the input image based on the point features.

[0016] Furthermore, the specific operation of step S201 includes the following steps:

[0017] S2011: Select a pixel point p in the input image, take p as the center and 3 pixels as the radius, and determine a detection neighborhood with a radius of 3 and an edge of 16 pixels;

[0018] S2012: Determine the gray value of pixel p, and make a difference between the gray value of 16 pixels on the circumference and the gray value of p. If the difference of n consecutive pixels is greater than the set threshold, then point p is set as the corner point of the detection neighborhood, and the sum of the difference of n consecutive pixels is recorded, which is recorded as

[0019] S2013: using other pixel points in the detection neighborhood determined in step S2011 as new pixel points p, repeating the above steps S2011-S2012 to determine whether there are other feature points that meet the characteristics in the detection neighborhood, until all pixel points in the input image are traversed;

[0020] S2014: Divide the input image into regions, select the difference and The largest corner point is taken as the corner point of the region, and the remaining corner points in the region are discarded.

[0021] Furthermore, the specific operation of step S202 includes the following steps:

[0022] S221: taking the corner point selected in step S214 as the base point q0, and taking a window centered at q0 as the sub-pixel neighborhood;

[0023] S2022: Use the Sobel operator to calculate the p in the window i Find pixel gradient G at point i , where p i is the i-th pixel in the sub-pixel neighborhood centered on q0;

[0024] S2023: Using the least squares method, p i Point and corresponding pixel gradient G i Substitute into the formula In the above formula, a systematic method is established to solve the new sub-pixel point q1; where ω i represents Gaussian weight.

[0025] S2024: With q1 as the center, repeat steps S2021-2023 for iterative calculation to obtain a new sub-pixel point q i , until the optimal solution is obtained.

[0026] Furthermore, the specific operation of step S3 includes the following steps:

[0027] S301: Analyze the image frames of the input image in sequence, calculate the gradient value and line area angle of each pixel in the image frame, set the gradient threshold, merge the pixels greater than the gradient threshold, and obtain line features;

[0028] S302: Filter the line features extracted in step S2021 according to the distance, length and number of the line features;

[0029] S303: Use the LBD descriptor to match the line features to obtain the inter-frame pose of the input image based on the line features.

[0030] Furthermore, the gradient value and line area angle of the pixel points in the image frame of the input image are calculated, and the gradient threshold ρ is set to suppress the pixel points with a gradient value less than ρ, and the coordinates of the remaining pixel points that meet the conditions are placed in the same space;

[0031] S3012: Setting direction threshold Select the pixel with the largest gradient value among all the pixels that meet the gradient threshold, and take the coordinates of the pixel from the space as the seed pixel seed;

[0032] S3013: Based on the seed, select the pixel gradient direction within the direction threshold If the pixel meets the conditions, it will be taken out of the space so that the pixel will not be judged repeatedly;

[0033] S3014: Every time a pixel point j that meets the threshold is found, the angle of the area composed of the pixel points that meet the threshold in the image frame is updated to Where, the subscript j is used to traverse all pixels in the region, level_line_angel j is the line area angle of the pixel points in the image frame calculated in step S3011;

[0034] S3015: All pixels in the area are judged to obtain a line segment support domain, scattered points in the line segment support domain are fitted to obtain a minimum circumscribed rectangle of the line segment support domain, and the rectangle is corrected to obtain a line feature.

[0035] Furthermore, the specific operation of step S4 includes the following steps:

[0036] S401: Select key frames;

[0037] S402: setting two conditions, 3D space and 2D plane line feature distance thresholds, in the sliding window to merge repeated line features;

[0038] S403: In the sliding window, repeat the above steps S401 to S402 for each image frame, select a key frame sequence, merge repeated line features, calculate the error state function of each feature, and obtain the inter-frame pose matrix, thereby achieving visual inertial positioning of the target.

[0039] Furthermore, the specific operation of step S401 includes the following steps: filtering the image frame sequence input by the camera, there are 15 consecutive image frames between the latest key frame and the most recently inserted key frame; the number of point and line tracking in the current frame must reach 20; when the inter-frame pose between the latest image frame and the current key frame exceeds a preset threshold, the key frame is removed and a new key frame is introduced.

[0040] Furthermore, the specific operation of step S402 includes the following steps:

[0041] S4021: Number the line features in the key frame and the current frame, and assign the line features of the current frame to i Switch to the Manhattan coordinate system and set the 3D space line feature distance threshold to threshold1. If the line feature line in the keyframe j The line feature line in the current frame i The distance is less than threshold1, and the two-line features are transferred to a two-dimensional plane;

[0042] S4022: Set the 2D plane line feature distance threshold to threshold2, the line feature line in the current frame i With the line feature line in the keyframe j When the distance between them is less than the threshold threshold2, the two line features are considered to overlap;

[0043] S4023: Merge two overlapping line features.

[0044] The beneficial effects of the present invention are:

[0045] 1. The visual inertial positioning method based on sub-pixel feature merging optimization in this invention aims at the problems of insufficient number of pixel-level point features and incomplete information in low-texture environments. Considering the rich feature information of super-resolution images, a feature detection algorithm based on pixel / sub-pixel fusion is proposed to extract and select point features of multi-level pixels, obtain deep image information, and then analyze the scene image frame. The experimental results show that after multi-layer iterative selection of image point features, the accuracy of corner points is improved and invalid feature points are reduced. Compared with traditional experimental methods, the positioning accuracy of the system can be improved by up to 25%.

[0046] 2. In order to solve the problem of affecting positioning accuracy by using only a small number of point features in complex environments, the present invention introduces two-dimensional line features to achieve point and line feature fusion. At the same time, the line features are constrained according to the angle and size of the feature vector to refine the effective features, thereby improving the accuracy and ensuring the efficiency of the algorithm.

[0047] 3. In order to further improve the algorithm running speed and optimize the algorithm, the present invention merges repeated line features, uses threshold judgment to select key frames, matches and merges repeated line features according to three-dimensional landmark point information and two-dimensional line feature information, reduces the dimension of the sliding window state quantity to be estimated, increases the proportion of effective line features, reduces the algorithm complexity, and improves the algorithm positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the visual inertial positioning method in the present invention.

[0049] Figure 2 This is a schematic diagram of extracting point features of an input image using a multi-level pixel feature detection algorithm in the present invention.

[0050] Figure 3 It is a schematic diagram of extracting line features from an input image in the present invention.

[0051] Figure 4 This is a schematic diagram of point-line feature fusion in the present invention.

[0052] Figure 5 This is the trajectory absolute error diagram of the visual inertial positioning method with sub-pixel point and line feature fusion in the present invention. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0054] A visual inertial positioning method based on sub-pixel feature merging optimization is shown in the attached Figure 1 As shown, the following steps are included:

[0055] S1: Preprocess the camera image to obtain a grayscale image as an input image for the visual system; the grayscale processing can be performed using the grayscale processing method in the prior art, which will not be described in detail in the present invention.

[0056] Further, S2: use the multi-level pixel feature detection algorithm to extract the point features of the input image; take the EuRoC public dataset to verify the method, taking a frame point feature extraction image during the algorithm operation as an example, as shown in the attached figure. Figure 2 shown.

[0057] Specifically, S201: using a pixel-level feature detection method, performing a first round of corner point detection on the input image;

[0058] More specifically, S2011: Figure 2 Select a pixel point p in the image, take p as the center and 3 pixels as the radius, and determine a detection neighborhood with a radius of 3 and an edge of 16 pixels;

[0059] S2012: Determine the gray value of pixel p, and make a difference between the gray value of 16 pixels on the circumference and the gray value of p. If the difference of n consecutive pixels is greater than the set threshold, then point p is set as the corner point of the detection neighborhood, and the sum of the difference of n consecutive pixels is recorded, which is recorded as

[0060] S2013: using other pixel points in the detection neighborhood determined in step S2011 as new pixel points p, repeating the above steps S2011-S2012 to determine whether there are other feature points that meet the characteristics in the detection neighborhood, until all pixel points in the input image are traversed;

[0061] S2014: Divide the input image into regions, select the difference and The largest corner point is taken as the corner point of the region, and the remaining corner points in the region are discarded.

[0062] Further, S202: performing secondary screening on the detected corner points using a sub-pixel corner point detection method;

[0063] More specifically, S2021: taking the corner point selected in step S2014 as the base point q0, and taking a window centered on q0 as the sub-pixel neighborhood;

[0064] S2022: Use the Sobel operator to calculate the p in the window i Find pixel gradient G at pointi ; Among them, p i is the i-th pixel in the sub-pixel neighborhood centered on q0;

[0065] S2023: Using the least squares method, p i Point and pixel gradient G i Substitute into the formula In the above formula, a systematic method is established to solve the new sub-pixel point q1; where ω i represents Gaussian weight.

[0066] S2024: With q1 as the center, repeat steps S2021-2023 for iterative calculation to obtain a new sub-pixel point q i , specify the number of iterations as 10, and the iteration stops automatically after reaching this number of iterations, then it is considered that the optimal solution is obtained.

[0067] S203: Perform image edge constraint on the sub-pixel corner point detection result to remove sub-pixel corner points that are not at the edge of the image; specifically, determine the gray value of the pixel point and set the gray value threshold α. If the gray value is less than the threshold α, the corner point is considered to be at the edge of the image; if the gray value is greater than the threshold α, the corner point is not at the edge of the image and is discarded.

[0068] S204: Use the LK optical flow method to perform feature matching on the sub-pixel corner points finally obtained in step S203 to obtain the inter-frame pose of the input image based on point features; the LK optical flow method is a prior art and will not be described in detail in this application.

[0069] Further, S3: while performing point feature extraction in step S2, line feature extraction is performed on the input image; to ensure that the extraction accuracy of line features is at the sub-pixel level, LSD line feature extraction (a Line Segment Detector) is used. To ensure the real-time performance of the algorithm, the number and direction of the extracted line features are screened to obtain high-quality line features. The EuRoC public dataset is used to verify the method, and two consecutive frames of line feature extraction images are selected during the algorithm operation, as shown in the attached figure. Figure 3 It should be noted that the point feature extraction in step S2 and the line feature extraction in step S3 in this application are performed simultaneously in two threads.

[0070] Specifically, S301: sequentially analyzing the image frames of the input image, calculating the gradient value and line area angle of each pixel in the image frame, setting a gradient threshold, merging the pixels greater than the gradient threshold, and obtaining line features;

[0071] More specifically, S3011: calculating the gradient value and line area angle of the pixel points in the image frame of the input image, setting the gradient threshold ρ, suppressing the pixel points with a gradient value less than ρ, and placing the coordinates of the remaining pixel points that meet the conditions into the same space;

[0072] Establish a coordinate system in the image frame, let the coordinates of the pixel point be (x, y), then the gradient value of the pixel point in the image frame is defined as in,

[0073] The line area angle is defined as

[0074] S3012: Setting direction threshold Select the pixel with the largest gradient value among all the pixels that meet the gradient threshold, and take the coordinates of the pixel from the space as the seed pixel seed;

[0075] S3013: Based on the seed, select the pixel gradient direction within the direction threshold If the pixel meets the conditions, it will be taken out of the space so that the pixel will not be judged repeatedly;

[0076] S3014: Every time a pixel point j that meets the threshold is found, the angle of the area composed of the pixel points that meet the threshold in the image frame is updated to Where, the subscript j is used to traverse all pixels in the region, level_line_angel j is the line area angle of the pixel points in the image frame calculated in step S3011;

[0077] S3015: All pixels in the area are judged to obtain a line segment support domain, scattered points in the line segment support domain are fitted to obtain a minimum circumscribed rectangle of the line segment support domain, and the rectangle is corrected to obtain a line feature.

[0078] S302: Filter the line features extracted in step S2021 according to the distance, length and number of the line features to reduce the number of line features that need to be matched;

[0079] S303: Use the LBD descriptor to match the line features to obtain the inter-frame pose of the input image based on the line features.

[0080] Further, S4: the point features extracted in step S2 and the line features extracted in step S3 are fused to achieve visual inertial positioning.

[0081] The inter-frame pose based on point features finally obtained in step S2 and the inter-frame pose based on line features finally obtained in step S3 are optimized, and the optimized pose information is jointly processed with the pose information obtained by the inertial device to obtain more accurate positioning information. The method for processing inertial device information is a prior art. IMU data obtains the carrier specific force and angular velocity through a gyroscope and an accelerometer. The pre-integration of the inertial device is to integrate the specific force and angular velocity to obtain the pose at the current moment and the relative motion of the next frame in the IMU coordinate system. The output frequency of the IMU data is different from that of the camera image, so a timestamp is used to align the IMU data with the camera frame number. The timestamp is used to initialize the visual inertial joint, and the relative pose of the camera and the inertial device is solved to improve the positioning accuracy of the algorithm. Specifically,

[0082] S401: Select key frames; filter the image frame sequence input by the camera, there are 15 consecutive image frames between the latest key frame and the most recently inserted key frame; the number of point and line tracking in the current frame must reach 20; when the inter-frame pose between the latest image frame and the current key frame exceeds a preset threshold, the key frame is removed and a new key frame is introduced.

[0083] S402: To avoid the same line feature being detected repeatedly in the previous and next frames, merge the repeated line features; set the three-dimensional space and two-dimensional plane line feature distance thresholds in the sliding window, merge the repeated line features, reduce the back-end error state function, and improve the positioning accuracy of the algorithm. The EuRoC public data set is used to verify the method. The repeated line features detected during the algorithm operation are marked in red, as shown in the attached figure. Figure 4 shown.

[0084] The image frame in the sliding window is used as the key frame, and the real-time image is used as the current frame. The following processing is adopted: the line features in the key frame and the current frame are numbered, and the line features of the current frame are numbered. i Switch to the Manhattan coordinate system and set the 3D space line feature distance threshold to threshold1. If the line feature line in the keyframe j The line feature line in the current frame i The distance is less than threshold1, and the two-line features are transferred to a two-dimensional plane;

[0085] Set the 2D plane line feature distance threshold to threshold2, the line feature line in the current frame i With the line feature line in the keyframe j When the distance between them is less than the threshold threshold2, the two line features are considered to overlap; the two overlapping line features are merged.

[0086] S403: In the sliding window, repeat the above steps S401 to S402 for each image frame, select a key frame sequence, merge repeated line features, calculate the error state function of each feature, and obtain the inter-frame pose matrix, thereby achieving visual inertial positioning of the target.

[0087] The predicted trajectory generated by the algorithm of the present invention is compared with the actual trajectory in the data set to obtain an error trajectory diagram, as shown in the attached figure. Figure 5 Attached Figure 5 In the figure, the error trajectory diagrams of MH01, MH02, MH03, MH04, and MH05 correspond to five different scenes MH01, MH02, MH03, MH04, and MH05 of the EuRoC public dataset. Compared with the traditional point-line feature algorithm, the present invention improves the point-line feature extraction accuracy to the sub-pixel level, selects key frames, merges repeated line features, reduces the interference of useless line features on back-end optimization, improves the algorithm positioning accuracy, and shortens the algorithm running time.

[0088] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A visual inertial positioning method based on sub-pixel feature merging optimization, characterized in that: The following steps are included: S1: Preprocess the camera image to obtain a grayscale image as the input image of the visual system; S2: Extract point features of the input image using a multi-level pixel feature detection algorithm; S3: while performing point feature extraction in step S2, perform line feature extraction on the input image; S4: Fusing the point features extracted in step S2 and the line features extracted in step S3 to achieve visual inertial positioning; The specific operation of step S2 includes the following steps: S201: Performing a first round of corner point detection on the input image using a pixel-level feature detection method; S202: performing secondary screening on the detected corner points using a sub-pixel corner point detection method; S203: performing image edge constraint on the sub-pixel corner point detection result, and removing sub-pixel corner points that are not at the edge of the image; S204: Use the LK optical flow method to perform feature matching on the sub-pixel corner points finally obtained in step S203 to obtain the inter-frame pose of the input image based on the point features.

2. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 1 is characterized in that: The specific operation of step S201 includes the following steps: S2011: Select a pixel point p in the input image, take p as the center and 3 pixels as the radius, and determine a detection neighborhood with a radius of 3 and an edge of 16 pixels; S2012: Determine the gray value of pixel p, and make a difference between the gray value of 16 pixels on the circumference and the gray value of p. If the difference of n consecutive pixels is greater than the set threshold, then point p is set as the corner point of the detection neighborhood, and the sum of the difference of n consecutive pixels is recorded, which is recorded as S2013: using other pixel points in the detection neighborhood determined in step S2011 as new pixel points p, repeating the above steps S2011-S2012 to determine whether there are other feature points that meet the characteristics in the detection neighborhood, until all pixel points in the input image are traversed; S2014: Divide the input image into regions, select the difference and The largest corner point is taken as the corner point of the region, and the remaining corner points in the region are discarded.

3. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 2 is characterized in that: The specific operation of step S202 includes the following steps: S221: taking the corner point selected in step S214 as the base point q0, and taking a window centered at q0 as the sub-pixel neighborhood; S2022: Use the Sobel operator to calculate the p in the window i Find pixel gradient G at point i , where p i is the i-th pixel in the sub-pixel neighborhood centered on q0; S2023: Using the least squares method, p i Point and corresponding pixel gradient G i Substitute into the formula In the above formula, a systematic method is established to solve the new sub-pixel point q1; where ω i represents Gaussian weight; S2024: With q1 as the center, repeat steps S2021-2023 for iterative calculation to obtain a new sub-pixel point q i , until the optimal solution is obtained.

4. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 3 is characterized in that: The specific operation of step S3 includes the following steps: S301: Analyze the image frames of the input image in sequence, calculate the gradient value and line area angle of each pixel in the image frame, set the gradient threshold, merge the pixels greater than the gradient threshold, and obtain line features; S302: Filter the line features extracted in step S2021 according to the distance, length and number of the line features; S303: Use the LBD descriptor to match the line features to obtain the inter-frame pose of the input image based on the line features.

5. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 4 is characterized in that: The specific operation of step S301 includes the following steps: S3011: Calculate the gradient value and line area angle of the pixel points in the image frame of the input image, set the gradient threshold ρ, suppress the pixel points with gradient values ​​less than ρ, and place the coordinates of the remaining pixel points that meet the conditions into the same space; S3012: Setting direction threshold Select the pixel with the largest gradient value among all the pixels that meet the gradient threshold, and take the coordinates of the pixel from the space as the seed pixel seed; S3013: Based on the seed, select the pixel gradient direction within the direction threshold If the pixel meets the conditions, it will be taken out of the space so that the pixel will not be judged repeatedly; S3014: Every time a pixel point j that meets the threshold is found, the angle of the area composed of the pixel points that meet the threshold in the image frame is updated to Where, the subscript j is used to traverse all pixels in the region, level_line_angel j is the line area angle of the pixel points in the image frame calculated in step S3011; S3015: All pixels in the area are judged to obtain a line segment support domain, scattered points in the line segment support domain are fitted to obtain a minimum circumscribed rectangle of the line segment support domain, and the rectangle is corrected to obtain a line feature.

6. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 5, characterized in that: The specific operation of step S4 includes the following steps: S401: Select key frames; S402: setting two conditions, 3D space and 2D plane line feature distance thresholds, in the sliding window to merge repeated line features; S403: In the sliding window, repeat the above steps S401 to S402 for each image frame, select a key frame sequence, merge repeated line features, calculate the error state function of each feature, and obtain the inter-frame pose matrix, thereby achieving visual inertial positioning of the target.

7. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 6 is characterized in that: The specific operation of step S401 includes the following steps: filtering the image frame sequence input by the camera, there are 15 consecutive image frames between the latest key frame and the most recently inserted key frame; the number of point and line tracking in the current frame must reach 20; when the inter-frame pose between the latest image frame and the current key frame exceeds a preset threshold, the key frame is removed and a new key frame is introduced.

8. The visual inertial positioning method based on sub-pixel feature merging optimization according to claim 7, characterized in that: The specific operation of step S402 includes the following steps: S4021: Number the line features in the key frame and the current frame, and assign the line features of the current frame to i Switch to the Manhattan coordinate system and set the 3D space line feature distance threshold to threshold1. If the line feature line in the keyframe j The line feature line in the current frame i The distance is less than threshold1, and the two-line features are transferred to a two-dimensional plane; S4022: Set the 2D plane line feature distance threshold to threshold2, the line feature line in the current frame i With the line feature line in the keyframe j When the distance between them is less than the threshold threshold2, the two line features are considered to overlap; S4023: Merge two overlapping line features.

Images