An image loop detection method based on adaptive quadtree partitioning

Abstract

The application discloses a kind of image loop detection methods based on adaptive quadtree partitioning.The method is: firstly, real-time image is collected as key frame, using adaptive quadtree partitioning to divide the region of current key frame, and all historical key frames are re-divided into regions according to the mode of current key frame;Calculate the similarity of each region of the current key frame and the corresponding region of all historical key frames, determine the loop candidate key frame, and calculate the mean image similarity, Euclidean distance and rotation angle difference between the current key frame and all loop candidate key frames, as well as the pose of the current key frame and the loop candidate key frame;Compare the mean image similarity, Euclidean distance and rotation angle difference with their respective corresponding thresholds respectively, and perform loop detection.The application makes the regional feature distribution of key frame more uniform, improves the calculation accuracy of image similarity, and improves the accuracy of loop detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and particularly relates to a loop detection method based on adaptive quadtree partitioning. BACKGROUND

[0002] After years of development, the SLAM (Simultaneous Localization and Mapping) technology has been mature and successfully applied to the fields of automatic driving, unmanned driving, autonomous mobile robots, etc. The visual SLAM is to position and map through the autonomous mobile carrier such as unmanned vehicle, unmanned aerial vehicle, unmanned ship or robot, etc. In the visual SLAM system, the loop detection method is usually used to improve the global optimization accuracy and mapping accuracy.

[0003] The loop detection refers to the autonomous mobile carrier identifying the current scene as a scene that has been reached before, so that the autonomous mobile carrier establishes a map to form a closed loop during the movement. However, in the prior art, the loop detection technology is not accurate enough, and false detection and missed detection often occur, thereby affecting the positioning and mapping effect. SUMMARY

[0004] In order to solve the defects and deficiencies of the prior art, the present application provides a loop detection method based on adaptive quadtree partitioning to improve the accuracy of loop detection.

[0005] To achieve the above object, the present application adopts the following technical scheme:

[0006] The method comprises the following steps:

[0007] 1) Real-time image acquisition is performed by using a monocular camera, 1 frame of image is randomly selected as a key frame every 60 frames of images, adaptive quadtree partitioning is adopted to divide the regions of the current key frame, all the key frames before the current key frame are regarded as historical key frames, and all the historical key frames are re-divided according to the region division mode of the current key frame;

[0008] 2) The similarity between each region of the current key frame and the corresponding region of all the historical key frames is calculated, the historical key frame with the maximum similarity with each region of the current key frame is selected as a loop candidate key frame, and the mean image similarity between the current key frame and all the loop candidate key frames is calculated;

[0009] 3) The pose of the current key frame and all the loop candidate key frames is calculated, and then the Euclidean distance and the rotation angle difference between the current key frame and all the loop candidate key frames are calculated according to the pose;

[0010] 4) The obtained mean image similarity, Euclidean distance and rotation angle difference are compared with the respective threshold values, respectively, to determine whether a loop occurs, and the loop detection is completed.

[0011] In the step 1), the current key frame is regionally divided by using adaptive quadtree block division, specifically as follows:

[0012] Step S1.1: Taking the geometric center point of the current key frame as an initial node, the current key frame is divided by passing the initial node along the directions parallel to the adjacent two edges of the current key frame, and the current key frame is evenly divided into four regions, and the threshold value of the feature point number of the region is set as T;

[0013] Step S1.2: The FAST feature point detection is performed on each region divided, and the feature point number in each region is compared with the threshold value T of the feature point number;

[0014] Step S1.3: According to the feature point number in each region of the current key frame, the following judgment is made:

[0015] If the feature point number in the region is less than or equal to the threshold value T of the feature point number, the region with the feature point number less than or equal to the threshold value T of the feature point number is stopped from being divided, and then the step S1.5 is jumped to;

[0016] If the feature point number in the region is greater than the threshold value T of the feature point number, the region with the feature point number greater than the threshold value T of the feature point number is evenly divided into four regions, and then the step S1.4 is jumped to;

[0017] Step S1.4: The steps S1.2 and S1.3 are repeated;

[0018] Step S1.5: When all the regions are stopped from being divided, the total number M of the regions divided by the current key frame is calculated, and the region division is completed.

[0019] In the step S1.3, a depth threshold value D is set for the quadtree to limit the maximum depth of the region division,

[0020] and when the division depth of the region is equal to or greater than the depth threshold value D, the division is stopped.

[0021] The step 2) is specifically as follows:

[0022] Step S2.1: The bag-of-words vector of each region of the current key frame is calculated, and the bag-of-words vector is calculated according to the following formula:

[0023]

[0024] wherein, is the bag-of-words vector of the kth region of the current key frame, the value range of k is {1,...,M}, M is the total number of all the regions divided by the current key frame, is the number of the i-th type of features in the k-th region, i is in the range of {1,...,N}, N is the number of types of features, is the bag-of-words weight of the i-th type of features in the k-th region of the current key frame,

[0025] The bag-of-words weight is specifically calculated according to the following formula:

[0026]

[0027] wherein n k is the number of feature points in the k-th region of the current key frame;

[0028] Step S2.2: obtaining the bag-of-words vector of the k-th region of the current key frame according to the current key frame obtained in step S2.1 The image similarity between the k-th region of the current key frame and the k-th region of each historical key frame is calculated, and the specific calculation formula is:

[0029]

[0030] wherein, is the i-th component of the bag-of-words vector corresponding to the k-th region of the current key frame, is the bag-of-words vector corresponding to the k-th region of the j-th historical key frame, is the i-th component of the bag-of-words vector corresponding to the k-th region of the j-th historical key frame is the image similarity between the k-th region of the current key frame and the k-th region of the j-th historical key frame;

[0031] Step S2.3: selecting the historical key frame with the largest image similarity with the k-th region of the current frame as the loop candidate key frame;

[0032] Step S2.4: calculating the average image similarity between the current key frame and all loop candidate key frames, which is specifically calculated according to the following formula:

[0033]

[0034] wherein S h is the average image similarity between the current key frame and the h-th loop candidate key frame, is the bag-of-words vector corresponding to the k-th region of the h-th loop candidate key frame, h is in the range of {1,...,M}.

[0035] The pose of the current key frame and the loop candidate key frame calculated in step 3) is specifically:

[0036] ​First, the pose of the current key frame is determined, specifically: extracting the feature points of the two key frames adjacent to the current key frame respectively and the feature points of the current key frame, then determining the pose of the extracted feature points, specifically: performing feature matching according to the extracted feature points to obtain a matching pair, substituting the matching pair into the epipolar constraint to obtain a fundamental matrix, decomposing the fundamental matrix to obtain a rotation matrix R c (r c ,p c ,q c ) and a translation vector T c (x c ,y c ) describing the change of the camera, and combining the depth z c recovered by triangulation to obtain the pose P c (x c ,y c ,z c ,r c ,p c ,q c ) of the current key frame.

[0037] Then, the pose of the hth loop candidate key frame is determined in the same way as the current key frame, and the pose of the hth loop candidate key frame is P h (x h ,y h ,z h ,r h ,p h ,q h ).

[0038] The determination of the pose adopts a world coordinate system, and the coordinate system is set to have a horizontal direction as an x-axis, a vertical direction as a y-axis, and a direction perpendicular to the x-axis and the y-axis as a z-axis.

[0039] Wherein, x c is the distance of the current key frame from the origin of the coordinate system in the x-axis direction, y c is the distance of the current key frame from the origin of the coordinate system in the y-axis direction, z c is the distance of the current key frame from the origin of the coordinate system in the z-axis direction, r c is the roll angle value of the current key frame, p c is the pitch angle value of the current key frame, q c is the yaw angle value of the current key frame; x h is the distance of the loop candidate key frame from the origin of the coordinate system in the x-axis direction, y h is the distance of the loop candidate key frame from the origin of the coordinate system in the y-axis direction, and z his the distance between the candidate loop key frame and the origin of the coordinate system in the z-axis direction, r h is the roll angle value of the loop candidate key frame, p h is the pitch angle value of the loop candidate key frame, q h is the yaw angle value of the loop closure candidate keyframe.

[0040] The Euclidean distance X between the current key frame and the hth loop candidate key frame in step 3) ch It is determined according to the following formula:

[0041]

[0042] The rotation angle difference R between the current keyframe and the hth loop candidate keyframe ch It is determined according to the following formula:

[0043] R ch =|r c -r h |+|p c -p h |+|q c -q h |.

[0044] The step 4) determining whether a loop occurs is specifically as follows: setting the mean similarity threshold to S, the Euclidean distance threshold to X, and the rotation angle difference threshold to R;

[0045] If the mean similarity S between the current keyframe and the hth loop candidate keyframe h Greater than the mean similarity threshold S, and the Euclidean distance X between the current key frame and the h-th loop candidate key frame ch The rotation angle difference R between the current keyframe and the hth loop candidate keyframe is less than the Euclidean distance threshold X. ch If it is less than the rotation angle difference threshold R, it means that the hth loop candidate key frame is a loop key frame and a loop occurs; otherwise, it indicates that no loop occurs.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention first uses adaptive quadtree partitioning to divide the current key frame into regions, so that regions with obvious features are divided emphatically, the features of each region are distributed more evenly, and the accuracy of image similarity calculation is improved;

[0048] Secondly, in addition to using mean image similarity to determine whether a loop occurs, the present invention also uses the Euclidean distance and rotation angle difference between the current key frame and the loop candidate key frame for judgment, thereby improving the accuracy of overall loop detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a logical flowchart of the method of the present application;

[0050] Figure 2 is a schematic diagram of a quadtree structure of the present application;

[0051] Figure 3 is a flowchart of an adaptive quadtree partitioning method of the present application. DETAILED DESCRIPTION

[0052] The present application is further described below in conjunction with the accompanying drawings and detailed description.

[0053] A user controls a robot carrying a monocular camera to move around in a closed room, collects images in front of the robot in real time, and then uses a simultaneous localization and mapping (SLAM) technology to construct a room map. Since some cumulative errors may exist in the process of map construction by the robot, the robot needs to use a loop detection method based on adaptive quadtree partitioning to optimize the pose of the collected images and improve the overall mapping accuracy. As shown in Figure 1 The loop detection method includes the following steps:

[0054] 1) Real-time image collection in front of the robot in a room using a monocular camera, random selection of 1 frame of image as a key frame every 60 frames of image, and regional division of the current key frame using adaptive quadtree partitioning. The key frames before the current key frame are all historical key frames, and all the historical key frames are re-divided into regions according to the regional division method of the current key frame, so that all the historical key frames and the current key frame have the same number of regions and region distribution.

[0055] 2) Calculation of the similarity between each region of the current key frame and the corresponding regions of all historical key frames, selection of the historical key frame with the greatest similarity to each region of the current key frame as a loop candidate key frame, and calculation of the mean image similarity between the current key frame and all loop candidate key frames.

[0056] 3) Calculation of the pose of the current key frame and all loop candidate key frames, and then calculation of the Euclidean distance and the rotation angle difference between the current key frame and all loop candidate key frames according to the pose.

[0057] 4) Comparison of the obtained mean image similarity, Euclidean distance and rotation angle difference with their respective thresholds to determine whether a loop occurs, that is, whether it is a place that has appeared before, to complete loop detection. If a loop occurs, the positioning data difference of the repeated place is compared to correct the system error.

[0058] As shown in Figure 3As shown, in step 1), the adaptive quadtree block is used to divide the current key frame into regions, specifically:

[0059] Step S1.1: Taking the geometric center point of the current key frame as an initial node, the current key frame is divided along the directions parallel to the adjacent two edges of the current key frame, and the current key frame is evenly divided into 4 regions, and the threshold value of the feature point number of the region is set as T;

[0060] Step S1.2: The FAST feature point detection is performed on each region divided, and the feature point number in each region is compared with the feature point number threshold T;

[0061] Step S1.3: According to the feature point number in each region of the current key frame, the following judgment is made:

[0062] If the feature point number in the region is less than or equal to the feature point number threshold T, the region with the feature point number less than or equal to the feature point number threshold T is stopped from being divided, and then the step S1.5 is jumped to;

[0063] If the feature point number in the region is greater than the feature point number threshold T, the region with the feature point number greater than the feature point number threshold T is evenly divided into 4 regions, and then the step S1.4 is jumped to;

[0064] Step S1.4: Steps S1.2 and S1.3 are repeated;

[0065] Step S1.5: When all regions are stopped from being divided, the total number M of regions divided by the current key frame is calculated, and the region division is completed, as shown in Figure 2 .

[0066] In step S1.3, in order to prevent the number of region division from being too large, a depth threshold D is set for the quadtree to limit the maximum depth of region division, and when the division depth of the region is equal to or greater than the depth threshold D, the division is stopped. In the specific implementation, the value of D is set to 4.

[0067] Step 2) is specifically:

[0068] Step S2.1: The bag-of-words vector of each region of the current key frame is calculated, and the bag-of-words vector is calculated according to the following formula:

[0069]

[0070] Wherein, is the bag-of-words vector of the kth region of the current key frame, the value range of k is {1,...,M}, M is the total number of all regions divided by the current key frame, is the number of the ith feature in the kth region, the value range of i is {1,...,N}, N is the number of features, is the bag-of-words weight of the i-th category feature in the k-th region of the current keyframe;

[0071] Bag of Words Weight It is calculated according to the following formula:

[0072]

[0073] Among them, n k is the number of feature points in the kth region of the current key frame;

[0074] Step S2.2: The bag-of-words vector of the kth region of the current keyframe obtained in step S2.1 Calculate the image similarity between the kth region of the current key frame and the kth region of each historical key frame. The specific calculation formula is:

[0075]

[0076] in, is the i-th component of the bag-of-words vector corresponding to the k-th region of the current keyframe, is the bag-of-words vector corresponding to the kth region of the jth historical keyframe, is the bag-of-words vector corresponding to the kth region of the jth historical keyframe The i-th component of is the image similarity between the kth region of the current key frame and the kth region of the jth historical key frame;

[0077] Step S2.3: Select the historical key frame with the greatest similarity to the k-th region image of the current frame as the loop closure candidate key frame;

[0078] Step S2.4: Calculate the mean image similarity between the current keyframe and all loop candidate keyframes, specifically according to the following formula:

[0079]

[0080] Among them, S h is the mean image similarity between the current keyframe and the hth loop candidate keyframe, is the bag-of-words vector corresponding to the kth region of the hth loop candidate keyframe, and the value range of h is {1,...,M}.

[0081] The poses of the current keyframe and the loop candidate keyframes are calculated in step 3 as follows:

[0082] First, the pose of the current key frame is determined, specifically: the feature points of the two frames of key frames adjacent to the current key frame before and after the current key frame and the feature points of the current key frame are extracted, that is, the key frame adjacent to the current key frame before the current key frame and the key frame adjacent to the current key frame after the current key frame. Then the feature points extracted are determined, specifically: feature matching is performed according to the extracted feature points to obtain a matching pair, the matching pair is substituted into the epipolar constraint to solve the fundamental matrix, and the rotation matrix R c (r c ,p c ,q c ) and the translation vector T c (x c ,y c ) describing the change of the camera are obtained by decomposing the fundamental matrix, and the depth z c is recovered by combining triangulation. The pose of the current key frame is optimized for the minimum reprojection error as P c (x c ,y c ,z c ,r c ,p c ,q c );

[0083] The feature points of the two frames of key frames adjacent to the hth loop candidate key frame before and after the hth loop candidate key frame and the feature points of the hth loop candidate key frame are extracted, feature matching is performed according to the extracted feature points to obtain a matching pair, the matching pair is substituted into the epipolar constraint to solve the fundamental matrix, and the rotation matrix R h (r h ,p h ,q h ) and the translation vector T h (x h ,y h ) describing the change of the camera are obtained by decomposing the fundamental matrix, and the depth z h is recovered by combining triangulation. The pose of the hth loop candidate key frame is optimized for the minimum reprojection error as P h (x h ,y h ,z h ,r h ,p h ,q h )。

[0084] Then the pose of the hth loop candidate key frame is determined in the same way as the current key frame, and the pose of the hth loop candidate key frame is P h (x h ,y h ,z hr h p h q h ;

[0085] The determination of the pose adopts a world coordinate system, in which a horizontal direction is set as an x-axis, a vertical direction is set as a y-axis, and a direction perpendicular to the plane where the x-axis and the y-axis are located is set as a z-axis; the embodiment takes a point of an indoor corner as a coordinate system origin to determine the pose of the key frame.

[0086] wherein x c is a distance of the current key frame from the coordinate system origin in the x-axis direction, y c is a distance of the current key frame from the coordinate system origin in the y-axis direction, z c is a distance of the current key frame from the coordinate system origin in the z-axis direction, r c is a roll angle value of the current key frame, p c is a pitch angle value of the current key frame, and q c is a yaw angle value of the current key frame; x h is a distance of the hth loop candidate key frame from the coordinate system origin in the x-axis direction, y h is a distance of the hth loop candidate key frame from the coordinate system origin in the y-axis direction, z h is a distance of the hth loop candidate key frame from the coordinate system origin in the z-axis direction, r h is a roll angle value of the hth loop candidate key frame, p h is a pitch angle value of the hth loop candidate key frame, and q h is a yaw angle value of the hth loop candidate key frame.

[0087] The Euclidean distance X ch between the current key frame and the hth loop candidate key frame in step 3) is determined according to the following formula:

[0088]

[0089] The rotation angle difference R ch between the current key frame and the hth loop candidate key frame is determined according to the following formula:

[0090] R ch = |r c -r h | + |p c -p h | + |q c -q h |.

[0091] Step 4) is to determine whether a loop occurs, which is specifically: setting a mean similarity threshold value as S, a Euclidean distance threshold value as X, and a rotation angle difference threshold value as R;

[0092] If the average similarity S between the current key frame and the hth loop candidate key frame is greater than the average similarity threshold S, the Euclidean distance X between the current key frame and the hth loop candidate key frame is less than the Euclidean distance threshold X, the rotation angle difference R between the current key frame and the hth loop candidate key frame is less than the rotation angle difference threshold R, and the hth loop candidate key frame is the loop key frame, loop occurs; otherwise, it is indicated that loop does not occur. h ch ch

[0093] The present application uses adaptive quadtree block division to divide the current key frame into regions, so that the regions with obvious features are highlighted, the feature distribution of each region is more uniform, and the image similarity calculation accuracy is improved.

[0094] Meanwhile, the average image similarity, the Euclidean distance and the rotation angle difference of the current key frame and the loop candidate key frame are used to jointly determine whether loop occurs, so that the overall loop detection accuracy is improved.

[0095] The above is only a preferred embodiment of the present application, and any equivalent changes and modifications made within the scope of the present application should be included in the scope of the present application.​​​

Claims

1. An image loop detection method based on adaptive quadtree partitioning, characterized by: The method comprises the following steps: 1) Use a monocular camera to collect images in real time, randomly select one frame as a key frame every 60 frames, and use adaptive quadtree block partitioning to divide the current key frame into regions. The key frames before the current key frame are all regarded as historical key frames, and all historical key frames are re-divided into regions according to the region division method of the current key frame; 2) Calculate the similarity between each region of the current key frame and the corresponding regions of all historical key frames, select the historical key frame with the greatest similarity to each region of the current key frame as the loop candidate key frame, and calculate the mean image similarity between the current key frame and all loop candidate key frames; 3) Calculate the poses of the current keyframe and all loop candidate keyframes respectively, and then calculate the Euclidean distance and rotation angle difference between the current keyframe and all loop candidate keyframes based on the poses; 4) Compare the obtained mean image similarity, Euclidean distance and rotation angle difference with their respective thresholds to determine whether a loop occurs and complete loop detection.

2. The image loop detection method based on adaptive quadtree partitioning according to claim 1, characterized in that: In step 1), the adaptive quadtree partitioning is used to divide the current key frame into regions as follows: Step S1.1: Using the geometric center point of the current keyframe as the initial node, divide the current keyframe along the directions parallel to the two adjacent edges of the current keyframe through the initial node, and divide the current keyframe into 4 regions on average. Set the feature point count threshold of each region to T; Step S1.2: Perform FAST feature point detection on each divided area and compare the number of feature points in each area with the feature point threshold T; Step S1.3: Make the following judgment based on the number of feature points in each area of ​​the current key frame: If the number of feature points in the region is less than or equal to the feature point threshold T, then stop dividing the region where the number of feature points is less than or equal to the feature point threshold T, and then jump to step S1.5; If the number of feature points in the region is greater than the feature point threshold T, the region with feature points greater than the feature point threshold T is evenly divided into 4 regions, and then the process jumps to step S1.4; Step S1.4: Repeat steps S1.2 and S1.3; Step S1.5: When all regions have stopped being divided, the total number of regions M divided by the current key frame is calculated, and the region division is completed.

3. The image loop detection method based on adaptive quadtree partitioning according to claim 2, characterized in that: In step S1.3, a depth threshold D is set for the quadtree to limit the maximum depth of the region division, and when the division depth of the region is equal to or greater than the depth threshold D, the division is stopped.

4. The image loop detection method based on adaptive quadtree partitioning according to claim 1, wherein: The step 2) is specifically as follows: Step S2.1: Calculate the bag-of-words vector for each region of the current keyframe. The bag-of-words vector is calculated according to the following formula: in, is the bag-of-words vector of the kth region of the current keyframe, the value range of k is {1,...,M}, M is the total number of regions split from the current keyframe, is the number of features of type i in the kth region, the value range of i is {1,...,N}, N is the number of feature types, is the bag-of-words weight of the i-th category feature in the k-th region of the current keyframe; The bag-of-words weight It is calculated according to the following formula: Among them, n k is the number of feature points in the kth region of the current key frame; Step S2.2: The bag-of-words vector of the kth region of the current keyframe obtained in step S2.1 Calculate the image similarity between the kth region of the current key frame and the kth region of each historical key frame. The specific calculation formula is: in, is the i-th component of the bag-of-words vector corresponding to the k-th region of the current keyframe, is the bag-of-words vector corresponding to the kth region of the jth historical keyframe, The bag-of-words vector corresponding to the kth region of the jth historical keyframe The i-th component of is the image similarity between the kth region of the current key frame and the kth region of the jth historical key frame; Step S2.3: Select the historical key frame with the greatest similarity to the k-th region image of the current frame as the loop closure candidate key frame; Step S2.4: Calculate the mean image similarity between the current keyframe and all loop candidate keyframes, specifically according to the following formula: Among them, S h is the mean image similarity between the current keyframe and the hth loop candidate keyframe, is the bag-of-words vector corresponding to the kth region of the hth loop candidate keyframe, and the value range of h is {1,...,M}.

5. The image loop detection method based on adaptive quadtree partitioning according to claim 1, characterized in that: The poses of the current key frame and the loop candidate key frame are calculated in step 3) as follows: First, the pose of the current key frame is determined by extracting the feature points of the two key frames adjacent to the current key frame and the feature points of the current key frame. Then, the pose of the extracted feature points is determined by performing feature matching based on the extracted feature points to obtain matching pairs. The matching pairs are substituted into the epipolar constraint to solve the basic matrix. After decomposing the basic matrix, the rotation matrix R describing the camera change is obtained. c (r c ,p c ,q c ) and the translation vector T c (x c ,y c ), combined with triangulation to recover the depth z c The minimum reprojection error is optimized to obtain the pose of the current key frame as P c (x c ,y c ,z c ,r c ,p c ,q c ); Then, the pose of the hth loop candidate key frame is determined in the same way as the current key frame, and the pose of the hth loop candidate key frame is obtained as P h (x h ,y h ,z h ,r h ,p h ,q h ); The world coordinate system is used to determine the posture. In the coordinate system, the horizontal direction is set as the x-axis, the vertical direction is set as the y-axis, and the direction perpendicular to the plane where the x-axis and y-axis are located is set as the z-axis. Among them, x c is the distance of the current keyframe from the origin of the coordinate system in the x-axis direction, c is the distance of the current keyframe from the origin of the coordinate system in the y-axis direction, z c is the distance between the current key frame and the origin of the coordinate system in the z-axis direction, r c is the roll angle value of the current key frame, p c is the pitch angle value of the current key frame, q c is the yaw angle value of the current key frame; x h is the distance between the candidate key frame of the loop and the origin of the coordinate system in the x-axis direction, and y h is the distance of the loop candidate keyframe from the origin of the coordinate system in the y-axis direction, z h is the distance between the candidate key frame of the loop and the origin of the coordinate system in the z-axis direction, r h is the roll angle value of the loop candidate key frame, p h is the pitch angle value of the loop candidate key frame, q h is the yaw angle value of the loop closure candidate keyframe.

6. The image loop detection method based on adaptive quadtree partitioning according to claim 5, characterized in that: The Euclidean distance X between the current key frame and the hth loop candidate key frame in step 3) ch It is determined according to the following formula: The rotation angle difference R between the current keyframe and the hth loop candidate keyframe ch It is determined according to the following formula: R ch =|r c -r h |+|p c -p h |+|q c -q h |。 7. The image loop detection method based on adaptive quadtree partitioning according to claim 4 or 6, characterized in that: The step 4) determining whether a loop occurs is specifically as follows: setting the mean similarity threshold to S, the Euclidean distance threshold to X, and the rotation angle difference threshold to R; If the mean similarity S between the current keyframe and the hth loop candidate keyframe h Greater than the mean similarity threshold S, and the Euclidean distance X between the current key frame and the h-th loop candidate key frame ch The rotation angle difference R between the current keyframe and the hth loop candidate keyframe is less than the Euclidean distance threshold X. ch If it is less than the rotation angle difference threshold R, it means that the hth loop candidate key frame is a loop key frame and a loop occurs; otherwise, it indicates that no loop occurs.

Images