A cloud-edge collaborative visual SLAM method and system based on feature point prediction

Abstract

This invention discloses a cloud-edge collaborative visual SLAM method and system based on feature point prediction. The method includes: adding several consecutive frames of images into a sliding window; extracting feature points from all pixels; establishing a quadtree as prior information for the corresponding frame images and merging them to establish global prior information; using the global prior information to predict the feature point distribution of the next frame image using a multi-step Markov probability model; adjusting the branches in the new quadtree; adding the next frame image into the sliding window, removing the earliest frame image and its corresponding prior information, and using the images within the sliding window and their corresponding prior information to predict the feature point distribution of subsequent frames. This invention considers the continuity of camera motion in space and the spatial relationship between feature points in consecutive frames. By constructing a multi-step Markov probability model, it reduces computational load while ensuring effective localization and mapping, significantly saving CPU computing resources.

Description

Technical Field

[0001] This invention belongs to the technical field of robot vision, specifically relating to a cloud-edge collaborative vision SLAM method and system based on feature point prediction. Background Technology

[0002] With the widespread adoption of smart terminal devices and the rapid development of technologies such as mobile robots, drones, and augmented reality, Simultaneous Localization and Mapping (SLAM) technology faces new challenges and opportunities. Traditional SLAM systems typically rely on a single device to complete all computational tasks, which places extremely high demands on resource-constrained mobile devices. To achieve stable and efficient SLAM functionality under limited computing power, storage space, and battery life, a cloud-edge collaborative SLAM architecture has emerged and become one of the current hot research directions.

[0003] Cloud-edge collaborative SLAM achieves a balance between system performance and resource consumption by rationally allocating computing tasks, offloading computationally intensive work to the cloud while retaining necessary real-time processing capabilities on edge devices. This architecture is particularly suitable for miniaturized devices equipped with low-power processors, such as smartphones, drones, and service robots. Although these devices have limited computing power, cloud-edge collaboration can overcome the limitations of local hardware and achieve a near-high-performance computing platform SLAM experience.

[0004] In practical applications, low-computing-power devices have raised many requirements for cloud-edge collaborative SLAM architectures, including minimizing computational load and CPU burden. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a cloud-edge collaborative visual SLAM method and system based on feature point prediction. It extracts feature points based on quadtrees and Markov properties, reducing the amount of computation while ensuring the localization and mapping effects.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] One aspect of the present invention provides a cloud-edge collaborative visual SLAM method based on feature point prediction, comprising the following steps:

[0008] Add the initial consecutive frames of images to a sliding window;

[0009] Feature points are extracted from all pixels of an initial series of consecutive frames of images.

[0010] A quadtree is built for the feature points extracted from each frame;

[0011] Save the quadtree as prior information for the corresponding frame image;

[0012] Merge the prior information corresponding to several consecutive initial frames of images and establish a new quadtree as global prior information;

[0013] The feature point distribution of the next frame image is predicted using a multi-step Markov probability model with global prior information.

[0014] The branches in the new quadtree are adjusted based on the actual feature point distribution of the next frame image.

[0015] The next frame image is added to the sliding window, and the earliest frame image and its corresponding prior information are removed. The feature point distribution of subsequent frame images is predicted using the images in the sliding window and their corresponding prior information.

[0016] As a preferred technical solution, the step of building a quadtree from the feature points extracted in each frame includes the following steps:

[0017] Create a root node representing the entire image region;

[0018] Starting from the root node, check if the current node has been partitioned. If it has not been partitioned and the spatial density of feature points within the node is ( N / Area )> ρ 1. Then, the current region is divided into four sub-regions, and the feature points are assigned to the corresponding sub-regions; where, N This represents the number of feature points in the current node. ρ 1 represents the set spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These represent the maximum and minimum values ​​of the x and y coordinates of the current region.

[0019] With node center ( x c , y c Divide the area using ) as the boundary:

[0020] ;

[0021] The area has reached the minimum size limit s =W / (2 n’ )× H / (2 n’ ), n’ The partitioning stops when ∈[6,8]. W and H For the width and height of the image frame, Q 1- Q 4 represents the four sub-regions obtained from the division;

[0022] The process is repeated recursively until the feature point is assigned to the appropriate leaf node.

[0023] As a preferred technical solution, the establishment of a new quadtree as global prior information, wherein the number of node feature points... ,in w 1, w 2,…, w n The weights of images at different time frames are indicated by the subscripts, and the weight values ​​are determined by the time order of the image frames. Newer image frames have greater weights, and vice versa. V 1, V 2,…, V n The number of feature points in images at different time frames is indicated by the subscripts;

[0024] Spatial density of feature points removed N / Area )< ρ The leaf nodes of 2, where, N This represents the number of feature points in the current node. ρ 2 represents the preset spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These are the maximum and minimum values ​​of the x and y coordinates of the current region.

[0025] As a preferred technical solution, the method of using global prior information and a multi-step Markov probability model to predict the feature point distribution of the next frame image specifically involves:

[0026] Define a multistep Markov probability modelP ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-k+1 ), P ( X n+1 | X n , X n-1 ,..., X 0) indicates the current feature point distribution state of the image. X n+1 ,and X n+1 All previous distribution states X n , X n-1 ,..., X 0 is related to all; P ( X n+1 | X n , X n-1 ,..., X n-k+1 ) indicates that X n+1 Restricted to only with X n+1 Previous limited k each state X n , X n-1 ,..., X n-k+1 related;

[0027] pass X n , X n-1 ,..., X n-k+1 predict n +1 Image feature point distribution status :

[0028] ;

[0029] in E [∙] represents the maximum likelihood.

[0030] As a preferred technical solution, X n+1 Only with X n+1 Previous limited k= 3 states X n , X n-1 ,..., X n-2 Related to obtaining the third-order Markov probability model P ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-2 );

[0031] Suppose that the state transition of feature point distribution between consecutive image frames within a short time satisfies a linear relationship. X n+1 = aX n + bX n-1 + cX n-2 +...+ X 1+ ε , ε If the noise is Gaussian, then:

[0032] ;

[0033] in, a, b, c These are linear coefficients.

[0034] As a preferred technical solution, adjusting the branches in the new quadtree based on the actual feature point distribution of the next frame image specifically includes the following steps:

[0035] Feature point extraction: Extract feature points from the image corresponding to the leaf node, count the number, and update the leaf node data;

[0036] Node splitting: Spatial density of feature points within a node ( N / Area )> ρ 1. The current region is divided into four sub-regions, and feature points are assigned to the corresponding sub-regions. This process is repeated recursively until a feature point is assigned to a suitable leaf node. N This represents the number of feature points in the current node. ρ 1 represents the set spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These are the maximum and minimum values ​​of the x and y coordinates of the current region; when the region reaches the minimum size limit... s = W / (2 n’ )× H / (2 n’ ), n’ The partitioning stops when ∈[6,8]. W and H The width and height of the image frame;

[0037] Local reconstruction: If a parent node P Only leaf nodes, and the spatial density of feature points in each leaf node ( N’ / Area )< ρ 3. Then complete the parent node. P The four leaf nodes; among them N’ The number of feature points in the leaf nodes. Area Let be the area of ​​the leaf node. ρ 3 represents the set threshold for the spatial density of leaf nodes;

[0038] Extract feature points from the completed leaf node region. If the parent node is completed... P If the number of feature points in each leaf node does not reach the set threshold T2, then let the parent node... P Become a leaf node and perform a partial reconstruction upwards; if the parent node is completed... P If there exists a leaf node whose number of feature points reaches a set threshold T2, then... P parent node P 1 is the root node, and the step of building a quadtree from the feature points extracted from each frame is executed;

[0039] The upward local reconstruction specifically involves: determining... P parent node P 1. Are there only leaf nodes? If so, then... P 1. Perform the steps of the local reconstruction.

[0040] As a preferred technical solution, at the beginning of the step of adjusting the branches in the new quadtree, the number of quadtree layers is set, and the corresponding leaf nodes with a number of feature points less than the set threshold T2 are assigned a value of -1, indicating the abandoned area. The abandoned nodes do not play a role during merging.

[0041] Another aspect of the present invention provides a cloud-edge collaborative visual SLAM system based on feature point prediction, including an edge terminal, an edge server, and a cloud server that communicate with each other through a communication module.

[0042] The edge end applies the aforementioned cloud-edge collaborative visual SLAM method based on feature point prediction to obtain the distribution of image feature points and construct and store a sub-map;

[0043] The edge server is used to reconstruct the data uploaded from the edge terminal based on the SLAM visual odometry, obtain the real-time path and roadside points, and then optimize the edge map through the SLAM backend; the edge map is then stored.

[0044] The cloud server is used to merge edge maps uploaded by edge servers into a global map, and to optimize and store the global map.

[0045] In another aspect, the present invention provides a cloud-edge collaborative visual SLAM device based on feature point prediction, comprising: a memory and at least one processor, wherein the memory stores instructions, and the memory and at least one processor are interconnected via a line; at least one processor invokes the instructions in the memory to cause the cloud-edge collaborative visual SLAM device based on feature point prediction to perform the cloud-edge collaborative visual SLAM method based on feature point prediction as described above.

[0046] In another aspect, the present invention also provides a storage medium storing a program that, when executed by a processor, implements the above-described cloud-edge collaborative visual SLAM method based on feature point prediction.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] (1) This invention takes into account the continuity of camera movement in space and the spatial connection between feature points between frames. By constructing a multi-step Markov probability model and then constructing an algorithm, the CPU's computing resources are greatly saved.

[0049] (2) This invention adapts to the needs of low computing power devices through cloud-edge collaboration, and performs visual odometry feature point extraction and management while ensuring system robustness. This overcomes the problem that the extraction of feature points (taking ORB feature points as an example) of the entire image in the prior art is very resource-intensive and time-consuming. Attached Figure Description

[0050] Figure 1 This is a flowchart of a cloud-edge collaborative visual SLAM method based on feature point prediction according to an embodiment of the present invention;

[0051] Figure 2 This is a flowchart illustrating the creation of a quadtree according to an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram showing the correspondence between a quadtree and an image according to an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the weighted merging of quadtrees according to an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram of the dynamic adjustment of the prior quadtree according to an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of window sliding according to an embodiment of the present invention;

[0056] Figure 7 This is a schematic diagram of the structure of a cloud-edge collaborative visual SLAM system based on feature point prediction according to an embodiment of the present invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0058] Example 1:

[0059] like Figure 1 As shown, this embodiment provides a cloud-edge collaborative visual SLAM method based on feature point prediction, including the following steps:

[0060] S1. Input Image: Add several consecutive frames of images to the sliding window.

[0061] S2. Global Feature Point Extraction: Extract feature points from all pixels of an initial series of consecutive frames of images.

[0062] S3. Construct a quadtree: For each frame, construct a quadtree to manage the extracted feature points. (Flowchart shown below) Figure 2 As shown, the correspondence between the image and the quadtree is as follows: Figure 3 As shown.

[0063] A quadtree is a tree-like data structure used for efficiently managing two-dimensional space. It recursively divides a region into four equal-sized subregions, with each node having at most four child nodes. Quadtrees are widely used in image processing, enabling fast spatial queries and collision detection. Their hierarchical structure optimizes storage and retrieval efficiency, making them particularly suitable for handling unevenly distributed data.

[0064] Furthermore, step S3 includes the following steps:

[0065] S301. Initialize the root node: Create a root node representing the entire image region.

[0066] S302, Insert Feature Points: Starting from the root node, check if the current node has been partitioned. If not, and the spatial density (points per unit area) of feature points within the node is... N / Area )> ρ 1. Divide the current region into four sub-regions and assign feature points to the corresponding sub-regions; where, N This represents the number of feature points in the current node. ρ 1 represents the set spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These are the maximum and minimum values ​​of the x and y coordinates of the current region.

[0067] Partitioning rules (based on node center) x c , y c (Divided by using ) as the boundary).

[0068] ;

[0069] The area has reached the minimum size limit s = W / (2 n’ )× H / (2 n’ ), n’ The partitioning stops when ∈[6,8]. W and H For the width and height of the image frame, Q 1- Q 4 represents the four sub-regions obtained from the division.

[0070] S303. Recursively process until the feature point is assigned to the appropriate leaf node.

[0071] S4. Store as prior information: Save the quadtree corresponding to the image frame within the sliding window as prior information.

[0072] S5. Construct a quadtree using prior information: Merge the corresponding quadtrees within the sliding window to construct a new quadtree as global prior information, with the number of node feature points. ,in w 1, w 2,…, w n The weights of images at different time frames are indicated by the subscripts, and the weight values ​​are determined by the time order of the image frames. Newer image frames have greater weights, and vice versa. V 1, V 2,…, V n The number of feature points in the image at different time frames, with subscripts indicating the sequence number. Spatial density after removing feature points ( N / Area )< ρ Leaf node of 2, ρ 2 represents the preset spatial density threshold, which is less than or equal to... ρ 1. The specific values ​​are adjusted based on the scenario and experience. The process of merging quadtrees A and B into C is as follows: Figure 4 As shown, let the spatial density of feature points in leaf node a2 of quadtree A be ( N / Area )< ρ 2. Merge nodes at the same position in quadtrees A and B using a weighted method into the corresponding position in quadtree C. During the merging process, nodes with zero feature points have no effect. For example, quadtree B does not have a leaf node a. l Since a2, a3, and a4 are the corresponding feature points, the number of their corresponding feature points is 0, and the weights are... w•v If the value is 0, then only the leaf nodes a1, a3, a4 of quadtree A, and the leaf nodes b1 and b2 of quadtree B in the third level are retained in quadtree C.

[0073] S6. Using global prior information, a multi-step Markov probability model is employed to predict the feature point distribution of the next frame image.

[0074] The Markov property refers to the characteristic of a stochastic process, given its current state, that its future states depend only on the current state and are independent of past states. Mathematical expression:

[0075] For discrete-time stochastic processes { X n If the following conditions are met:

[0076] P ( X n+1 = x | X n = x n , X n-1 = x n-1 ,..., X 0= x 0)= P ( X n+1 = x | X n = x n );

[0077] The process is said to have Markov properties.

[0078] When the state at a certain moment depends on the previous state... k When the state at each moment (not just the current moment) is considered to be... k Markov process.

[0079] The mathematical expression is as follows:

[0080] P ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-k+1 );

[0081] That is, the conditional probability of a future state depends on the nearest... k A historical state.

[0082] Define a multistep Markov probability model P ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-k+1 ), P ( X n+1 | X n , X n-1 ,..., X 0) indicates the current feature point distribution state of the image. X n+1 ,and X n+1 All previous distribution states X n , X n-1 ,..., X 0 is related to all; P ( X n+1 | X n , X n-1 ,..., X n-k+1 ) indicates that X n+1 Restricted to only with X n+1 Previous limited k each state X n , X n-1 ,..., X n-k+1 related;

[0083] pass X n , X n-1 ,..., X n-k+1predict n +1 Image feature point distribution status :

[0084] ;

[0085] in E [∙] represents the maximum likelihood.

[0086] Furthermore, in this embodiment, let X n+1 Only with X n+1 Previous limited k= 3 states X n , X n-1 ,..., X n-2 Related to obtaining the third-order Markov probability model P ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-2 );

[0087] Suppose that the state transition of feature point distribution between consecutive image frames within a short time satisfies a linear relationship. X n+1 = aX n + bX n-1 + cX n-2 +...+ X 1+ ε , ε If the noise is Gaussian, then:

[0088] ;

[0089] in, a, b, c These are linear coefficients.

[0090] More specifically, a, b, c It can be represented as:

[0091] ;

[0092] ;

[0093] .

[0094] S7. Prior Quadtree Dynamic Adjustment: Adjusts the branches in the new quadtree in step S5 according to the actual distribution of image feature points in the next frame. Flowchart as follows: Figure 5 As shown.

[0095] Furthermore, step S7 includes the following steps:

[0096] S701. Extract feature points: Extract feature points from the image corresponding to the leaf node, count the number of points, and update the leaf node data.

[0097] S702, Node Splitting: Spatial density of feature points within a node (number of points per unit area) N / Area )> ρ If 1, the current region is divided into four sub-regions, and the feature points are assigned to the corresponding sub-regions. This process is repeated recursively until the feature points are assigned to the appropriate leaf nodes.

[0098] The area has reached the minimum size limit s = W / (2 n’ )× H / (2 n’ ), n’ The partitioning stops when ∈[6,8].

[0099] S703, Partial Reconstruction: If a parent node... P Only leaf nodes, and the spatial density of feature points in each leaf node ( N’ / Area )< ρ 3. Then complete the parent node. P The four leaf nodes; among them N’ The number of feature points in the leaf nodes. Area Let be the area of ​​the leaf node. ρ 3 represents the set spatial density threshold for leaf nodes, which is less than or equal to... ρ 1. The specific value is adjusted based on the scenario and experience.

[0100] Extract feature points from the completed leaf node region. If the parent node is completed... P If the number of feature points in each leaf node does not reach the set threshold T2, then the parent node... P Become a leaf node and perform a partial reconstruction upwards; if the parent node is completed... P If there exists a leaf node whose number of feature points reaches a set threshold T2, then... P parent nodeP 1 is the root node; execute step S3 to build a quadtree.

[0101] The upward local reconstruction specifically involves: determining... P parent node P 1. Are there only leaf nodes? If so, then... P 1. Execute step S703.

[0102] Furthermore, since adding or removing branches in a quadtree is cumbersome and complex in practice, this embodiment sets the number of quadtree layers at the beginning of the step of adjusting the branches in the new quadtree, assigns a value of -1 to the corresponding leaf node whose number of leaf node feature points is less than the set threshold T2, indicating that the abandoned area is not used during merging.

[0103] S8, Window Sliding: Removes the oldest image frame from the sliding window and adds a new image frame, such as... Figure 6 As shown.

[0104] This method processes sequential data by dynamically maintaining a fixed-size sliding window of local data. The core idea is that as the sequence progresses, the sliding window, like a sliding viewport, discards old data and incorporates new data, thus enabling real-time analysis and computation of continuous data. This approach cleverly balances time and space efficiency, avoiding the overhead of processing the entire dataset while capturing local features, making it particularly suitable for scenarios such as time series analysis, real-time signal processing, and streaming data mining.

[0105] Example 2:

[0106] like Figure 7 As shown, in this embodiment, a cloud-edge collaborative visual SLAM system based on feature point prediction is provided. The system includes an edge terminal, an edge server, and a cloud server that communicate with each other through a communication module.

[0107] The edge terminal applies a cloud-edge collaborative visual SLAM method based on feature point prediction from the above embodiment to obtain the distribution of image feature points and construct and store a sub-map;

[0108] The edge server is used to reconstruct the data uploaded from the edge terminal based on the SLAM visual odometry, obtain the real-time path and roadside points, and then optimize the edge map through the SLAM backend; the edge map is then stored.

[0109] The cloud server is used to merge edge maps uploaded by edge servers into a global map, and to optimize and store the global map.

[0110] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure can be divided into different functional modules to complete all or part of the functions described above. This system is a cloud-edge collaborative visual SLAM method based on feature point prediction applied to the above embodiments.

[0111] Example 3:

[0112] In this embodiment, a cloud-edge collaborative visual SLAM device based on feature point prediction is provided, comprising: a memory and at least one processor, wherein the memory stores instructions, and the memory and at least one processor are interconnected via a line; at least one processor invokes the instructions in the memory to cause the cloud-edge collaborative visual SLAM device based on feature point prediction to execute a cloud-edge collaborative visual SLAM method based on feature point prediction as described in the above embodiment.

[0113] Example 4:

[0114] In this embodiment, a storage medium is provided, storing a program. When the program is executed by a processor, it implements a cloud-edge collaborative visual SLAM method based on feature point prediction as described in the above embodiment, specifically:

[0115] Add images from several consecutive initial frames into a sliding window;

[0116] Feature points are extracted from all pixels of the image in the initial consecutive frames;

[0117] A quadtree is built for the feature points extracted from each frame;

[0118] Save the quadtree as prior information for the corresponding frame image;

[0119] The prior information corresponding to the images in several consecutive initial frames is merged to establish a new quadtree as global prior information;

[0120] The feature point distribution of the next frame image is predicted using a multi-step Markov probability model with global prior information.

[0121] The branches in the new quadtree are adjusted based on the actual feature point distribution of the next frame image.

[0122] The image of the next frame is added to the sliding window, and the image of the earliest frame and its corresponding prior information are removed. The feature point distribution of subsequent frames is predicted using the images in the sliding window and their corresponding prior information.

[0123] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0124] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A cloud-edge collaborative visual SLAM method based on feature point prediction, characterized in that, Includes the following steps: Add the initial consecutive frames of images into a sliding window; Feature points are extracted from all pixels of an initial series of consecutive frames of images. A quadtree is built for the feature points extracted from each frame; Save the quadtree as prior information for the corresponding frame image; Merge the prior information corresponding to several consecutive initial frames of images and establish a new quadtree as global prior information; The feature point distribution of the next frame image is predicted using a multi-step Markov probability model with global prior information. The following steps are used to adjust the branches in the new quadtree based on the actual feature point distribution of the next frame image: Feature point extraction: Extract feature points from the image corresponding to the leaf node, count the number, and update the leaf node data; Node splitting: Spatial density of feature points within a node ( N / Area )> ρ 1. The current region is divided into four sub-regions, and feature points are assigned to the corresponding sub-regions. This process is repeated recursively until a feature point is assigned to a suitable leaf node. N This represents the number of feature points in the current node. ρ 1 represents the set spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These are the maximum and minimum values ​​of the x and y coordinates of the current region; when the region reaches the minimum size limit... s = W / (2 n’ )× H / (2 n’ ), n’ The partitioning stops when ∈[6,8]. W and H The width and height of the image frame; Local reconstruction: If a parent node P Only leaf nodes, and the spatial density of feature points in each leaf node ( N’ / Area )< ρ 3. Then complete the parent node. P The four leaf nodes; among them N’ The number of feature points in the leaf nodes. Area Let be the area of ​​the leaf node. ρ 3 represents the set threshold for the spatial density of leaf nodes; Extract feature points from the completed leaf node region. If the parent node is completed... P If the number of feature points in each leaf node does not reach the set threshold T2, then let the parent node... P Become a leaf node and perform a partial reconstruction upwards; if the parent node is completed... P If there exists a leaf node whose number of feature points reaches a set threshold T2, then... P parent node P 1 is the root node, and the step of building a quadtree from the feature points extracted from each frame is executed; The upward local reconstruction specifically involves: determining... P parent node P 1. Are there only leaf nodes? If so, then... P 1. Perform the steps of the local reconstruction; Add the next frame image to the sliding window and remove the earliest frame image and its corresponding prior information. Use the images in the sliding window and their corresponding prior information to predict the feature point distribution of subsequent frames.

2. The cloud-edge collaborative visual SLAM method based on feature point prediction according to claim 1, characterized in that, The process of building a quadtree from the feature points extracted in each frame includes the following steps: Create a root node representing the entire image region; Starting from the root node, check if the current node has been partitioned. If it has not been partitioned and the spatial density of feature points within the node is ( N / Area )> ρ 1. Then, the current region is divided into four sub-regions, and the feature points are assigned to the corresponding sub-regions; where, N This represents the number of feature points in the current node. ρ 1 represents the set spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These represent the maximum and minimum values ​​of the x and y coordinates of the current region. Divide the nodes into two parts by using the node center (xc, yc) as the boundary: ； The area has reached the minimum size limit s = W / (2 n’ )× H / (2 n’ ), n’ The partitioning stops when ∈[6,8]. W and H For the width and height of the image frame, Q 1- Q 4 represents the four sub-regions obtained from the division; The process is repeated recursively until the feature point is assigned to the appropriate leaf node.

3. The cloud-edge collaborative visual SLAM method based on feature point prediction according to claim 1, characterized in that, The establishment of a new quadtree is used as global prior information, in which the number of node feature points is... ,in w 1, w 2,…, w n The weights of images at different time frames are indicated by the subscripts, and the weight values ​​are determined by the time order of the image frames. Newer image frames have greater weights, and vice versa. V 1, V 2,…, V n The number of feature points in images at different time frames is indicated by the subscripts; Spatial density of feature points removed N / Area )< ρ The leaf nodes of 2, where, N This represents the number of feature points in the current node. ρ 2 represents the preset spatial density threshold. Area =( x max - x min )×( y max - y min ) represents the area of ​​the current node. x max , x min , y max and y min These represent the maximum and minimum values ​​of the x and y coordinates of the current region.

4. The cloud-edge collaborative visual SLAM method based on feature point prediction according to claim 1, characterized in that, The specific steps of using global prior information and a multi-step Markov probability model to predict the feature point distribution of the next frame image are as follows: Define a multistep Markov probability model P ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-k+1 ), P ( X n+1 | X n , X n-1 ,..., X 0) indicates the current feature point distribution state of the image. X n+1 ,and X n+1 All previous distribution states X n , X n-1 ,..., X 0 is related to all; P ( X n+1 | X n , X n-1 ,..., X n-k+1 ) indicates that X n+1 Restricted to only with X n+1 Previous limited k each state X n , X n-1 ,..., X n-k+1 related; pass X n , X n-1 ,..., X n-k+1 predict n +1 Image feature point distribution status : ； in E [∙] represents the maximum likelihood.

5. The cloud-edge collaborative visual SLAM method based on feature point prediction according to claim 4, characterized in that, make X n+1 Only with X n+1 Previous limited k= 3 states X n , X n-1 ,..., X n-2 Related to obtaining the third-order Markov probability model P ( X n+1 | X n , X n-1 ,..., X 0)= P ( X n+1 | X n , X n-1 ,..., X n-2 ); Suppose that the state transition of feature point distribution between consecutive image frames within a short time satisfies a linear relationship. X n+1 = aX n + bX n-1 + cX n-2 +...+ X 1+ ε , ε If the noise is Gaussian, then: ； in, a, b, c These are linear coefficients.

6. The cloud-edge collaborative visual SLAM method based on feature point prediction according to claim 1, characterized in that, At the beginning of the step of adjusting the branches in the new quadtree, the number of quadtree layers is set, and the corresponding leaf nodes with a number of feature points less than the set threshold T2 are assigned a value of -1, indicating that the region is abandoned. The abandoned nodes do not play a role in the merging process.

7. A cloud-edge collaborative visual SLAM system based on feature point prediction, characterized in that, This includes edge terminals, edge servers, and cloud servers that communicate with each other via communication modules; The edge terminal applies a cloud-edge collaborative visual SLAM method based on feature point prediction as described in any one of claims 1-6 to obtain the distribution of image feature points and construct and store a sub-map; The edge server is used to reconstruct the data uploaded from the edge terminal based on the SLAM visual odometry, obtain the real-time path and roadside points, and then optimize the edge map through the SLAM backend; the edge map is then stored. The cloud server is used to merge edge maps uploaded by edge servers into a global map, and to optimize and store the global map.

8. A cloud-edge collaborative visual SLAM device based on feature point prediction, characterized in that, include: A memory and at least one processor, wherein the memory stores instructions and the memory and at least one processor are interconnected via a circuit; At least one of the processors invokes the instructions in the memory to cause the cloud-edge collaborative visual SLAM device based on feature point prediction to perform a cloud-edge collaborative visual SLAM method based on feature point prediction as described in any one of claims 1-6.

9. A storage medium storing a program, characterized in that: When the program is executed by the processor, it implements the cloud-edge collaborative visual SLAM method based on feature point prediction as described in any one of claims 1-6.

Images