CN118038239A - An indoor visual loop detection method based on point-line feature global descriptor - Google Patents

Abstract

The invention provides an indoor vision loop detection method based on a dotted line feature global descriptor, which belongs to the technical field of indoor vision positioning and comprises the following steps: step 1, extracting image dotted line characteristics and descriptors; step 2, establishing a key frame; step 3, generating a global descriptor; and 4, matching the historical key frames. The invention is based on the image appearance structured dot line characteristic information, is oriented to the positioning requirement of indoor scenes, and can meet the indoor scene loop detection with consistent visual long-term positioning track.

Description

An indoor visual loop detection method based on point-line feature global descriptor

Technical Field

The present invention belongs to the technical field of visual indoor positioning, and specifically relates to an indoor visual loop detection method based on a point-line feature global descriptor, which can be used to achieve long-term positioning of low-cost portable visual sensors in indoor scenes.

Background technique

With the rapid development of society, indoor positioning has great demand in various fields of social life. In the commercial field, navigation services are needed in large buildings such as shopping malls, airports, hospitals or large office buildings to improve user experience; in the entertainment field, indoor positioning technology can be used for virtual and augmented reality applications to provide a more realistic and interactive experience; in the industrial field, various IoT mobile machines need to know their exact location in order to better perform tasks. The Global Navigation Satellite System (GNSS) is the most traditional and widely used positioning technology at present, but in indoor scenes, GNSS signals are blocked and blocked by buildings, and the accuracy of positioning results is greatly affected. At the same time, its positioning accuracy at a large scale level is difficult to meet the movement in a small scale environment. In addition, some traditional indoor positioning methods such as UWB positioning technology and Wi-Fi fingerprint technology, which calculate the current position through triangulation and other technologies based on the signal strength received by different base stations, but they must rely on pre-deployed base stations, do not have the ability of autonomous positioning, are applicable to few scenarios, and have many restrictions, so it is necessary to introduce autonomous positioning technology.

The vision-based SLAM technology is currently the most mainstream autonomous positioning technology. The core goal of SLAM (Simultaneous Localization and Mapping) is to enable machines or pedestrians to navigate autonomously in unknown environments and build maps at the same time. The sensors it relies on can be roughly divided into three categories: visual sensors, lidar, and inertial measurement units (IMUs). Compared with inertial sensors, visual sensors have the ability to perceive the surrounding environment, and compared with lidar, they have the advantages of low price and portability. Therefore, they are the most commonly used sensors in autonomous positioning technology. The vision-based SLAM is usually divided into three links: front-end tracking, back-end optimization, and loop detection. Since the pose calculation of visual SLAM essentially depends on the matching recursion between images, errors inevitably accumulate and eventually lead to pose divergence. Loop detection is to achieve global constraints by identifying the same position and detecting the arrival at the same place, thereby avoiding pose divergence.

The most common loop detection method is based on the bag-of-words model, which relies on a large number of scenes for training to obtain a large number of bag-of-words libraries. The image builds its own word vector based on the bag-of-words library and matches it accordingly. This method relies on feature points and is simple to implement, but the huge bag-of-words library is slow to load and is not suitable for all scenes. At the same time, indoor scenes often have the characteristics of structure and low texture, that is, the feature point information in the image is reduced and the line feature information is increased. Combining the above characteristics, for indoor positioning scenes, it is urgent to build a global description matching method based on structured point and line features, which does not rely on a priori word library construction and only matches based on information between images, so as to realize loop detection in the visual SLAM process and meet the requirements of global trajectory consistency of visual autonomous positioning in indoor scenes.

Summary of the invention

In order to solve the problem of divergence of visual positioning in structured indoor scenes with sufficient line features but fewer point features in the prior art, the present invention provides an indoor visual loop detection method based on a global descriptor of point and line features. This method is based on the structured point and line feature information of image appearance, is oriented to the positioning needs of indoor scenes, and can meet the indoor scene loop detection with consistent visual long-term positioning trajectory.

In order to achieve the above object, the present invention adopts the following technical scheme:

An indoor visual loop detection method based on a point-line feature global descriptor comprises the following steps:

Step 1: Use the SuperPoint model of the convolutional neural network to complete the extraction of the feature points and their descriptors of the image, use the line segment detection method to complete the line feature detection, and use the method of representing line features with point features to complete the extraction of the descriptor of the line feature, and finally complete the extraction of the point features, line features and descriptors of the image;

Step 2: According to the three criteria of frame interval, feature quantity and co-viewing relationship, each frame of the image is tested to complete the establishment of the key frame;

Step 3: For the key frame, cluster the point features and group the line features according to the intersection relationship, and calculate the global feature residual according to the clustering result and the grouping result and the descriptor of the local feature point to obtain the global descriptor of the key frame;

Step 4: Calculate the similarity between the global descriptor generated by the current key frame and the global descriptor of the historical key frame to obtain a candidate matching key frame; perform geometric verification and continuity verification on the candidate matching key frame to obtain a matching key frame confirmed to be at the same position, and complete loop detection.

Beneficial effects:

The method in the present invention is based on image matching technology. According to the structured characteristics of indoor environments, it utilizes the appearance information of images themselves and realizes a global image description method from local to overall through point and line feature clustering and grouping. It can realize detection and recognition of the same position without the need for a priori constructed visual word bags. It is widely applicable to indoor structured environments, thereby achieving long-term trajectory consistency of visual positioning.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of an indoor visual loop detection method based on a point-line feature global descriptor of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

As shown in FIG1 , an indoor visual loop detection method based on a point-line feature global descriptor of the present invention comprises the following steps:

Step 1: Extract image point and line features and descriptors;

Step 2: Establish key frames;

Step 3: Generate a global descriptor;

Step 4: Match historical key frames.

Specifically, the embodiment of the present invention takes an indoor scene as an example to describe the specific steps of a visual loop detection method based on a point-line feature global descriptor.

Step 1: Extract image point and line features and descriptors:

In SLAM positioning and image matching, directly operating on pixels is too computationally intensive, so feature extraction is often performed first to achieve dimensionality reduction.

Step 1.1 Feature point extraction and description:

In order to improve the effect of feature point extraction, the present invention uses the SuperPoint model based on convolutional neural network to perform automatic feature point extraction, which uses an encoder-decoder architecture: the encoder Encoder realizes image dimensionality reduction and feature extraction through multiple convolution layers and maximum pooling layers. The decoder is responsible for decoding the feature information compressed by the encoder to obtain the required information type. There are two decoders in the SuperPoint model: one is responsible for obtaining the feature point position, and the other is responsible for obtaining the feature point descriptor. Therefore, after the SuperPoint model, a series of feature point sets {x _i } in the image can be obtained, and each feature point x _i is composed of its position (u _i , _vi ) in the image pixel coordinate system and a 256*1 dimensional descriptor. constitute.

Step 1.2 Line feature extraction:

Considering the structured characteristics of the indoor environment, that is, more appearance information exists in the form of line segments, the image is further subjected to line feature detection. LSD (Line Segment Detector) is a line segment detection algorithm based on image gradients, which can achieve sub-pixel accuracy. The principle is that the grayscale values of pixels at both ends of the line segment as the detection target are different, which will inevitably produce gradient values. Therefore, the pixels are merged and distinguished according to the size and direction of the gradient, and finally the line area is generated, thereby obtaining a series of line feature sets {l _i }, where each line feature l _i consists of its two endpoints e _i1 , e _i2 and the slope k _i .

Considering that the line features extracted by LSD are often too fragmented, the extracted line features need to be merged. The merging standard is based on the minimum distance between the line segment endpoints and the line segment slope: when the minimum endpoint distance min{||{e _i1 ,e _i2 }-{{e _j1 ,e _j2 }||} between two line features (line segments) l _i and l _j is less than the threshold T _d and the slope difference |k _i -k _j | is less than the threshold T _k , it can be considered that the slopes of the two line features are consistent and the endpoints are continuous, and they can be merged into a line feature l _n . The endpoints e _n1 and e _n2 of the new line feature l _n are composed of the two endpoints with the farthest distance between the two line segments l _i and l _j , and the slope k _n is recalculated based on the new endpoints. Among them, {e _j1 ,e _j2 } represents the two endpoints of the line feature l _j , and k _j represents the slope of the line feature l _j .

To avoid line feature redundancy, when the number of line features is greater than the threshold n _l , the line features are screened by quadtree uniformization to obtain a line feature set {l _i } that removes redundant line features in overcrowded areas.

Step 1.3 Line feature description:

In order to save computing resources, the present invention uses a method of using feature point description to represent line feature description, and determines whether a feature point belongs to a line feature according to the distance from the feature point to the line feature.

First, according to the endpoints and slope, the straight line where the line feature _lj is located is expressed as a parameterized straight line: _Aj ·x+ _Bj ·y+ _Cj , where x and y are the image pixel coordinates, _Aj , _Bj , and _Cj are the parameters required to describe the straight line where the line feature _lj is located, and the distance _dij from the feature point _xi to the line feature _lj is:

Among them, (u_i, vi_i) is the position of the feature point x_i in the image pixel coordinate system. When the distance d _ij is less than the threshold, it is considered that the feature point belongs to the line feature. After calculating the distance between points and lines, the set of feature points belonging to the line feature l _j can be obtained And the feature point set {x _p } that does not belong to any line feature. The description of the line feature l _j is described by all the feature point descriptors belonging to the line feature.

Step 1.4 Point feature homogenization:

For a set of feature points that do not belong to any line feature, in order to prevent redundancy, when the number of points is greater than the threshold _np , the feature points are screened by quadtree uniformization to obtain a set of feature points {x _pi } that removes redundant feature points in overcrowded areas.

Step 2: Establish key frames:

In order to improve the efficiency of loop detection, the present invention establishes key frames through image frames, and then the image matching required for loop detection is only performed in the key frames, reducing the amount of calculation required for matching, including:

Step 2.1 Determine the frame interval:

Taking into account the normal positioning movement speed, the visually perceived scene will not change much in a short period of time. Therefore, the current key frame must be separated from the previous key frame by a certain number of frames. When the interval number of frames is greater than the threshold N _GAP , the current frame may be established as a key frame. The threshold N _GAP changes according to the frame rate of the visual image.

Step 2.2 Determine the number of features:

After point and line feature extraction, in order to ensure that the subsequent matching process has sufficient information, the number of features needs to be guaranteed. When the number of point features is greater than the threshold P _kf and the number of line features is greater than L _kf , the current frame may be established as a key frame.

Step 2.3 Determine the common view relationship:

The newly established key frame needs to maintain a certain interval with the previous key frame to reduce information redundancy, and also needs to maintain information continuity with the previous key frame, so it is tested through the common view relationship. Common view means that the same feature appears in both images, so the common view ratio of the feature in the current frame and the feature in the previous key frame needs to be maintained within the threshold range [Cov _L ,Cov _U ].

If the current frame passes all three test criteria of steps 2.1 to 2.3 above, it will be established as a new key frame KF _n . After judging by the above three test criteria, a representative key frame can be obtained, and then the loop detection only extracts and matches the subsequent global descriptors for the key frame.

Step 3: Extract global descriptor:

In order to obtain the descriptor vector of the global image, the present invention adopts the idea of VLAD (Vector of Local Aggregated Descriptors), but is not limited to the cluster description of feature points, but also realizes the cluster description of line features, thereby realizing the extraction of global descriptors including point and line features.

Step 3.1 Point feature clustering:

For point features, the present invention uses Kmeans++ clustering, which is simple and fast. It selects K cluster centers one by one during initialization, and the sample points farther away from other cluster centers are more likely to be selected as the next cluster center. Then, the K clusters and cluster centers are continuously updated according to the distance between the feature points and the cluster centers. After clustering is completed, K feature point clusters and corresponding cluster center feature points are finally obtained.

Step 3.2 Line feature grouping:

Taking into account real indoor scenes, line features often appear at the edges of objects, and usually multiple continuous (with the same endpoint) line segments constitute a structure (such as a wall corner, door frame). It can be simply summarized as line segments with a continuous relationship approximately belong to the same structure, and the present invention groups line features based on this feature.

During initialization, line segments are selected step by step to ensure that the distance between the minimum endpoint of the newly selected line segment and the selected line segment is greater than the threshold, and finally L initial line segments are obtained, which belong to L line segment groups. When L such line segments cannot be found in the image, the insufficient line segments are randomly selected.

Then, the distance calculation between the remaining line segments and the line segments in the established line segment group is started. When the minimum endpoint distance min{||{e _l1 ,e _l2 }-{e _a1 ,e _a2 }|| between a remaining line segment l _l and any line segment l _a in a line segment group is less than the threshold T _c , the two are considered continuous and added to the line segment group. When a line segment satisfies the above relationship with two line segment groups at the same time, the two line segment groups are merged. The two endpoints of the remaining line segment l _l are e _l1 ,e _l2 , and the two endpoints of any line segment l _a are e _a1 ,e _a2 . All line features are traversed in this way. After the traversal is completed, two situations usually occur:

At this time, there are still line segments that do not belong to any line segment cluster. At this time, add them to the line segment group with the smallest number of line segments;

At this time, the number of line segment groups is less than L, so the line segment group with the largest number is split and evenly divided into two line segment groups until L line segment groups are met.

After obtaining L line segment groups, its group center point is calculated. By calculating the centroid of the feature points belonging to the group segment, the feature point closest to the centroid is used as the group center.

Step 3.3 Global feature residual calculation:

After obtaining L line feature groups and K point feature clusters and the corresponding (L+K) central feature points, by subtracting all feature points from their central feature points, we can get the feature distribution that can express the L+K clustering ranges, and these L+K local set features constitute the global descriptor. The specific calculation formula is as follows:

Where V is the global descriptor matrix, V(m,j) is the element in the mth row and jth column of the matrix, and j represents the jth dimension of the feature point descriptor, with a total of 256 dimensions. m represents the mth local set, of which the first L are L line feature groups and the last K are K point feature clusters, with a total of L+K dimensions. First, calculate the L line feature groups. At this time, _xi represents the points in the feature point set { _xl } in the line feature. represents the descriptor of feature point x _i , and/> Represents the descriptor of the central feature point c _m of the mth line feature group. a _k is the belonging function. When the feature point _xi belongs to the mth cluster, a _k is 1, otherwise a _k is 0. Then calculate K feature point clusters. At this time, _xi represents a point in the feature point set {x _p } in the line feature./> represents the descriptor of feature point x _i , and/> Represents the descriptor of the feature point c _mL of the cluster center of the (mL)th point feature cluster. The final V is a (L+K)*256 matrix, in which each row vector is a description vector of a line feature group or a point feature cluster. This formula accumulates all feature residuals of each cluster and finally obtains (L+K) global features. These (L+K) global features express a certain distribution of local features within the cluster range. This distribution eliminates the feature distribution differences of the image itself by subtracting each feature point from the feature center, and only retains the distribution differences between the local features and the cluster center.

Step 3.4 L2 norm normalization:

For the convenience of calculation, the (K+L)*256 dimensional matrix is expanded in one dimension, that is, the row vectors are concatenated to obtain a (L+K)*256 one-dimensional vector V. Then the description vector is normalized by the L2 norm: Therefore, when calculating the similarity later, it can be ensured that the expression of each vector is numerically uniform, making it easier to compare the similarity between images.

Step 4: Match historical keyframes:

After extracting the global descriptor for each key frame, the descriptor of the current key frame must be matched with the historical key frames.

Step 4.1 Similarity calculation:

The present invention uses cosine similarity to perform vector matching, traverses the global descriptors _Vj of all historical key frames _KFj and compares them with the global descriptor _Vn of the current key frame _KFn , and calculates their similarity Sim(n,j) as:

Since the descriptor vector has been L2-norm normalized, || _Vn ||＝|| _Vj ||＝1, then Sim(n,j}＝ _Vn * _Vj . After completing the similarity calculation for all historical key frames, if there is no key frame with a similarity greater than the threshold _Simth , it means that the position in the image has never appeared in the historical information. Otherwise, the top three candidate key frames { _KFc } with the largest similarity greater than the threshold _Simth are selected (if there are less than three, all key frames greater than the threshold _Simth are selected), and geometric tests are performed in turn according to the similarity.

Step 4.2: Perform geometry verification:

The current keyframe and the candidate keyframes are matched with feature points and line features, and the pose solution is completed after RANSAC (RAndomSAmple Consensus) outliers are eliminated. Bidirectional reprojection is performed based on the pose to obtain optimized matching inliers. When the number of inliers is greater than the threshold _Pth , it can be considered that the two conform to the actual geometric relationship, that is, they are at the same location in the real world. When all candidate keyframes fail to pass the geometric check, it is considered that they are only similar in appearance, but not at the same location.

Step 4.3: Continuity check:

When a candidate key frame KF _c passes the geometric check, to ensure that it is at the same position, it is also necessary to perform geometric checks on the three consecutive historical frames F _c+1 , F _c+2 , and F _c+3 after the key frame and the three consecutive frames F _n+1 , F _n+2 , and F _n+3 after the current key frame V _n . When all consecutive frames pass the geometric check, it can be confirmed that the current position is the same, and then the matching relationship between the current key frame and the candidate key frame is output to the positioning system to complete the loop detection.

It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims (5)

1. An indoor vision loop detection method based on a dotted line feature global descriptor is characterized by comprising the following steps:

Step 1, extracting feature points and descriptors of an image by utilizing a SuperPoint model of a convolutional neural network, detecting line features by utilizing a line segment detection method, extracting descriptors of the line features by utilizing a method for representing the line features by utilizing the point features, and finally extracting the point features, the line features and the descriptors of the image;

Step2, checking each frame of image according to three standards of frame number interval, feature quantity and common view relation, so as to complete establishment of key frames;

step 3, clustering the point features, grouping the line features according to the intersecting relation, and completing calculation of global feature residual errors according to clustering results, grouping results and the local feature point descriptors to which the clustering results and the grouping results belong, so as to obtain global descriptors of the key frames;

step 4, performing similarity calculation on the global descriptor generated according to the current key frame and the global descriptor of the historical key frame to obtain candidate matching key frames; and performing geometric check and continuity check on the candidate matching key frames to obtain matching key frames confirmed to be in the same position, and completing loop detection.

2. The method for detecting indoor visual loop back based on global descriptor with dotted line features according to claim 1, wherein the step 1 comprises:

Step 1.1, performing automatic feature point extraction by using SuperPoint model based on convolutional neural network, which uses encoder-decoder architecture, wherein the encoder realizes the dimension reduction and feature extraction of the image through a plurality of convolutional layers and a maximum pooling layer, and the decoder is responsible for decoding the feature information compressed by the encoder to obtain the required information type; in SuperPoint model, there are two decoders, one is in charge of obtaining feature point position, and the other is in charge of obtaining feature point descriptor; after SuperPoint model, a series of feature point sets { x _i } in the image are obtained, each feature point x _i is composed of its position (u _i,v_i) in the image pixel coordinate system and a 256 x 1-dimensional descriptor Constructing;

Step 1.2, combining the characteristic of indoor environment structuring, namely that more appearance information exists in the form of line segments, continuing to perform line feature detection on the image, and adopting a line segment detection algorithm based on image gradients to realize sub-pixel level precision; the difference of the pixel gray values of the two end regions of the line segment serving as a detection target inevitably generates gradient values, so that pixels are combined and distinguished according to the size and the direction of the gradient, and a line region is finally generated, so that a series of line feature sets { l _i } are obtained, and each line feature l _i is composed of two end points e _i1,e_i2 and a slope k _i;

Combining the extracted line features, and judging the combination standard according to the minimum distance between line segment endpoints and the slope of the straight line: when the minimum endpoint distance min { ||{ e _i1,e_i2}-{e_j1,e_j2 } || } between the two line features l _i、l_j is smaller than the threshold T _d and the difference of the slopes |k _i-k_j | is smaller than the threshold T _k, the slopes of the two line features are considered to be consistent, the endpoints are continuous, namely the two line features are combined into one line feature l _n, the endpoint e _n1,e_n2 of the new line feature l _n is formed by two endpoints with the farthest two line segments of l _i、l_j, and the slope k _n is recalculated according to the new endpoint; where { e _j1,e_j2 } represents the two endpoints of line feature l _j, and k _j represents the slope of line feature l _j;

When the number of the line features is larger than a threshold value n _l, performing quadtree homogenization screening on the line features to obtain a line feature set { l _i } with redundant line features in an over-dense area removed;

step 1.3, judging whether the characteristic points are assigned to the line characteristics according to the distances from the characteristic points to the line characteristics by using a method for representing the line characteristics by using the characteristic point descriptions:

First, the line feature l _j is expressed as a parameterized straight line according to the endpoints and slopes: a _j·x+B_j·y+C_j, where x, y are image pixel coordinates, a _j、B_j、C_j is a parameter required for expressing a straight line where the line feature l _j is located, and a distance d _ij from the feature point x _i to the line feature l _j is:

Wherein, (u _i,v_i) is the position of the feature point x _i in the image pixel coordinate system; when the distance d _ij is smaller than the threshold value, the feature point is considered to belong to the line feature; after the calculation of the distance between the dotted lines, a characteristic point set belonging to the line characteristic l _j is obtained And a feature point set { x _p } that does not belong to any line feature; a description of a line feature l _j, i.e., by a global feature point descriptor attributed to the line feature;

And 1.4, for the feature point set which does not belong to any line feature, in order to prevent redundancy, performing quadtree homogenization screening on the feature points when the number of points is larger than a threshold value n _p, and obtaining a feature point set { x _pi } with redundant feature points in an over-dense region removed.

3. The method for detecting indoor visual loop back based on global descriptor with dotted line features according to claim 2, wherein the step 2 comprises:

Step 2.1 determining a frame number interval: an image in which a current key frame is spaced from a previous key frame by a certain number of frames, the current frame being established as a key frame when the number of frames spaced is greater than a threshold N _GAP, the threshold N _GAP being varied according to a frame rate of the visual image;

Step 2.2, determining the feature quantity: the current frame may be established as a key frame only when the number of point features is greater than the threshold P _kf and the number of line features is greater than L _kf;

Step 2.3, determining a common view relation: common view means that the same feature appears in both images, so that the common view ratio of the feature in the current frame and the feature in the previous key frame is maintained at a threshold interval [ Cov _L,Cov_U ];

If the current frame passes all three of the above inspection criteria of steps 2.1-2.3, it is established as a new key frame KF _n; and judging through the three inspection standards to obtain a representative key frame, and then performing loop detection to extract and match the subsequent global descriptor only for the key frame.

4. The method for detecting indoor visual loop back based on global descriptor with dotted line features according to claim 3, wherein the step 3 comprises:

Step 3.1, clustering the point characteristics: for the point characteristics, K cluster centers are selected one by one in the initialization process by utilizing Kmeans++ clustering, and sample points which are farther away from other cluster centers are more likely to be selected as the next cluster center, and then the K clusters and the cluster centers are continuously updated according to the distance between the characteristic points and the cluster centers; finally obtaining K characteristic point clusters and corresponding clustering center characteristic points after completing clustering

Step 3.2 line feature grouping:

Gradually selecting line segments during initialization, ensuring that the distance between the newly selected line segments and the minimum endpoint of the selected line segments is greater than a threshold value, and finally obtaining L initial line segments which belong to L line segment groups; when L line segments cannot be found in the picture, randomly selecting the insufficient line segments;

Then, starting to calculate the distance between the rest line segment and the line segment in the established line segment group, when the minimum endpoint distance min of a certain rest line segment l _l and any line segment l _a in a certain line segment group is smaller than a threshold T _c, considering that the rest line segment l _l and the line segment group are continuous, adding the rest line segment into the line segment group, and when a certain line segment simultaneously meets the above relation with the two line segment groups, merging the two line segment groups; two endpoints of the remaining line segment l _l are e _l1,e_l2, and two endpoints of the arbitrary line segment l _a are e _a1,e_a2;

Thus, all line features are traversed, and after the traversing is completed, two cases occur:

(1) If the line segments do not belong to any line segment cluster, adding the line segments into the line segment group with the minimum line segment number;

(2) If the number of the line segment groups is less than L, splitting from the line segment group with the largest number, and uniformly dividing the line segment group into two line segment groups until the line segment groups meet the L line segment groups;

after obtaining L line segment groups, calculating the group center point, calculating the mass center of the characteristic points belonging to the group of line segments, and taking the characteristic point closest to the mass center in the characteristic points as the group center; thus, L line characteristic groups and corresponding group center characteristic points are obtained

Step 3.3 global feature residual calculation:

After obtaining L line feature groups, K point feature clusters and corresponding (L+K) central feature points, subtracting all the feature points from the central feature points to obtain feature distribution expressing L+K clustering ranges, wherein the L+K local set features form a global descriptor, and the specific calculation formula is as follows:

wherein V is a global descriptor matrix, V (m, j) is an element of an mth row and an mth column of the matrix, j represents an mth dimension of a feature point descriptor, and m represents an mth local set, wherein the former L are L line feature groups, the latter K are K point feature clusters, and L+K dimensions are formed; l line feature sets are first computed, where x _i represents a point in the feature point set x _l that belongs to the line feature, Descriptor representing feature point x _i/(A descriptor representing the m-th line feature group center feature point c _m; a _k is a attribution function, when the feature point x _i belongs to the m-th cluster, a _k is 1, otherwise a _k is 0; k feature point clusters are then computed, where x _i represents a point in the feature point set { x _p }, which belongs to the line feature >Descriptor representing feature point x _i/(A descriptor representing the (m-L) th point feature cluster class center feature point c _m-L; the final V is an (L+K) 256 matrix, wherein each row of vectors is a description vector of a line feature group or a point feature cluster; the method adds up all the feature residuals of each cluster to finally obtain (L+K) global features;

Step 3.4L2, norm normalization: one-dimensional expansion is carried out on the obtained (K+L) 256-dimensional matrix, namely, row vectors are spliced, and a (L+K) 256-dimensional vector V is obtained; the description vector is then L2 norm normalized: so that the expression form of each vector is ensured to be numerically uniform when the similarity is calculated later, and the similarity between images is more easily compared.

5. The method for detecting indoor visual loop back based on global descriptor with dotted line features according to claim 4, wherein the step 4 comprises:

Step 4.1 similarity calculation: vector matching is carried out by utilizing cosine similarity, the global descriptor V _j of all the historical key frames KF _j is traversed, the global descriptor V _n of the current key frame KF _n is traversed, and the similarity Sim (n, j) is calculated as follows:

Since the descriptor vector has been L2 norm normalized, ||v _n||＝||V_j |=1, then Sim (n, j) =v _n*V_j; after the similarity calculation of all the historical key frames is completed, if no key frame with the similarity being greater than a threshold value Sim _th exists, the position in the image is represented to not appear in the historical information, otherwise, the first 3 candidate key frames { KF _c } with the maximum similarity being greater than the threshold value Sim _th are selected, and if the number of the candidate key frames is less than 3, all the key frames with the similarity being greater than the threshold value Sim _th are selected, and geometric inspection is sequentially carried out according to the similarity;

And 4.2, performing geometric verification: matching the characteristic points and the line characteristics of the current key frame and the candidate key frame, removing the RANSAC outliers, completing pose calculation, and carrying out two-way reprojection according to the pose to obtain matched interior points after optimization, wherein when the number of the interior points is greater than a threshold value P _th, the two are considered to be in accordance with the actual geometric relationship, namely the same position in the real world; when all candidate key frames cannot pass the geometric verification, the candidate key frames are considered to be similar in appearance at the moment, and are not in the same place;

Step 4.3, performing continuity check: when candidate key frames KF _c pass the geometric check, in order to ensure the same position, the continuous three-frame historical frame F _c+1、F_c+2、F_c+3 after the key frames and the continuous three-frame F _n+1、F_n+2、F_n+3 after the current key frame V _n are subjected to the geometric check in turn; when the continuous frames pass the geometric verification, the current arrival at the same position can be confirmed, and then the matching relation between the current key frame and the candidate key frame is output to the positioning system, so that loop detection is completed.