Empirical map reusing method
Through the combination of incremental binary search tree and dynamic island mechanism, the problem of low visual perception accuracy and efficiency in the RatSLAM system is solved, and rapid closed-loop detection and map fusion are realized. It is suitable for robots with limited endurance and improves environmental construction. The energy efficiency and map reuse capability of the graph.
Patent Information
- Application Number
- CN202310588138.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-24
- Publication Date
- 2024-11-26
AI Technical Summary
When the existing RatSLAM system is navigating and positioning in an unknown environment, the visual perception accuracy is poor and the efficiency is low. The robot needs to continuously traverse the entire environment to build a map. The lack of map reuse makes it impossible for robots with limited battery life to complete the construction and differences of large-environment maps. Map reuse between robots.
Using incremental binary search tree and dynamic island mechanism, rapid closed-loop detection is achieved through ORB feature extraction and binary search tree retrieval, allowing local environmental maps to fusion to form a global map, and reuse the experience maps between different robots.
It improves the efficiency and accuracy of position recognition, reduces the time complexity of loop detection, realizes efficient construction and reuse of maps, is suitable for robots with limited battery life, and enhances the energy efficiency of environmental map construction.
Smart Images

Figure CN119027598A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of mobile robot navigation and positioning, and specifically is a map reuse method established by using a brain-like navigation method Rat SLAM, and is a system for robot navigation and positioning in an unknown environment. Background Art
[0002] With the development of robotics technology, mobile robots have been widely used due to their wide range of applications and high mobility. When a robot moves in an unknown environment to perform tasks, it is crucial to estimate its position in the environment. Simultaneous localization and mapping (SLAM) is one of the most promising and practical strategies to solve this problem. SLAM is a process in which a robot uses sensor information to model the environment, draw an environmental map, and simultaneously estimate its position in the environmental map while moving in an unknown environment.
[0003] The RatSLAM method, which is based on the neural mechanism of navigation in the rodent brain, can solve the positioning and mapping problems using only a monocular camera, and has good application value. However, the original RatSLAM system uses the sum of the absolute differences of pixel values after average pooling to measure the similarity between images, and needs to be compared with the saved visual templates one by one to achieve position recognition or closed-loop detection. The visual perception accuracy is poor and the efficiency is low. Moreover, the robot needs to traverse the entire environment continuously at one time to build an environmental map. For robots with limited endurance, it is impossible to complete the construction of environmental maps of large environments. At the same time, for the same environment, different robots need to traverse the environment again to build an environmental map if they want to navigate in the environment. There is a lack of map reuse among different robots, and the energy efficiency of environmental mapping is low. Summary of the invention
[0004] In view of the above problems, the present invention proposes a method for reusing empirical maps. The method uses an incrementally constructed binary search tree to save environmental image features, adopts a dynamic island mechanism and uses the idea of sequence matching to perform closed-loop verification, and can efficiently and accurately achieve closed-loop detection; the method allows robots with limited endurance to explore the large environment in partitions, establish local environmental maps, and form a global map through map fusion; the method allows empirical maps to be reused in different individuals of the same type of robots, reducing repeated exploration and greatly improving the energy efficiency of environmental mapping.
[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:
[0006] An experience map reuse method comprises the following steps:
[0007] 1) Using incremental binary search tree and dynamic island mechanism to improve local view cells and obtain environmental perception information;
[0008] 2) Construct a pose cell network to convert the environmental perception information and the collected odometer information into the expression of the robot's position in the environment, and output the network state of the pose cells;
[0009] 3) Update the experience map according to the current posture cell network state;
[0010] 4) Save all files of the current experience map and reconstruct all data structures of the historical experience map;
[0011] 5) When the local view cell detects a position loop, the robot's position in the map is optimized; when the local view cell detects that the position loop is the overlapping position of the new map and the historical map, the new map and the historical map are merged.
[0012] The step 1) comprises the following steps:
[0013] 1.1) Use the ORB algorithm to extract features from the scene image, extract N ORB feature points, calculate the BRIEF descriptor of each feature point, and generate a set of N m-dimensional binary feature descriptors of the image;
[0014] 1.2) Image I i All feature descriptors are input into the binary search tree for retrieval, and based on the similarity score, it is determined whether there is a feature that is similar to image I. i Similar images, if image I j and the query image I i If the similarity score between them is greater than the threshold, then image I j For image I i Similar images of image I j As an alternative image, otherwise, insert image I into the binary search tree i All features are used to update the binary search tree;
[0015] 1.3) Collection and Image I i All candidate images similar to I i Sort the similarity from high to low to get an ordered image set C i = {I i1 ,I i2 ,...,I in}, where I ij Indicates that i The jth most similar image, j = 1, 2, ..., n, normalizes the similarity scores of the images in the set, removes images with scores less than the threshold, and obtains an ordered set of candidate images
[0016] 1.4) Using Alternative Image Sets All images I ij , creating dynamically managed islands Among them, the islands Represents images with timestamps between m and n. Each island has a representative image, called the island center, which represents the image most similar to the query image in the timestamp range [m, n]. The island most similar to the query image is selected, and the image corresponding to its island center is used as the final closed-loop decision result.
[0017] The similarity score s(I i ,I j )for:
[0018]
[0019] Among them, n i,j I is the query image i With image I j The number of feature points that are successfully matched in the search tree, N is the total number of feature points in the query image, τ 1 It means I i A dynamic threshold based on the mean similarity score of the previous n frames,
[0020]
[0021] The normalization of the similarity scores of the images in the set is specifically as follows:
[0022]
[0023] Among them, s′(I i ,I ij ) is the normalized similarity score.
[0024] The binary search tree comprises:
[0025] Leaf node: used to store the BRIEF descriptor set calculated from the feature points extracted from the image A subset of j is the jth BRIEF descriptor of the image, a dimensional binary vector, each descriptor d j The index of the image it represents is stored, and different leaf nodes store the feature set Divided into different areas, so that each descriptor belongs to a separate leaf node;
[0026] Root node and intermediate node N i : used to store bit index k i, k i ∈(0,1,…,dim(d)-1)), where dim(d) is the descriptor dimension.
[0027] The binary search tree performs the following operations:
[0028] Insert operation: For any descriptor d j According to d j [k i ] is stored in the left or right subtree respectively, d j [k i ] is d j At index k i If the value of d j [k i ] is 0, then enter the left subtree, otherwise enter the right subtree, and finally store it in the leaf node;
[0029] Retrieval operation: For the descriptor d to be queried q , starting from the root node, according to d q [k i ]Search the branch until the leaf node L * Until then, the descriptor to be queried is compared with all the descriptors stored in the leaf node one by one, and the most similar descriptor d is found according to a certain threshold. * .
[0030] The step 1.4) comprises the following steps:
[0031] 1.4.1) Yes All images in I ij , evaluates whether its timestamp is located on an existing island Within the time interval, if I ij Exists in the existing island, image I ij Set as The center of the island will be ij The first k frames and the last k frames of are added to the island. In order to adapt to the size of the island, when the current island size exceeds the limit, the images far away from the center of the island are moved out of the island and updated. time interval; if I ij If it does not exist in the existing island, a new island is created. The new island consists of I ij and its previous k frames and the next k frames, the center of the island is I ij ;
[0032] 1.4.2) Check whether all islands intersect with other islands. If so, crop the borders of the islands that intersect to ensure that the same image only appears in one island;
[0033] 1.4.3) Perform a global evaluation on each island and calculate the global score
[0034]
[0035] 1.4.4) Select the island with the highest global score and use the image corresponding to the center of the island as the closed-loop decision.
[0036] The fusion of the new map and the historical map in step 5) is specifically as follows:
[0037] 5.1) Before detecting the overlap between the new and old experience maps, the robot takes the last state of the historical experience map as the initial state and starts to rebuild the experience map. The new experience map has no connection with the historical experience map. At this time, the graph relaxation algorithm is called at a certain frequency to correct the map, and the position update coefficient is set to α = 0.5, and standard map optimization is performed;
[0038] 5.2) When the overlap between the new and old experience maps is detected, that is, when the current image is identified as being the same as a scene stored in the previous map, a link is established between the current experience and the experience in the historical map, and α=1 is set in the position update formula to optimize the map;
[0039] 5.3) After the fusion of the new and old maps is completed, the position update rule with α=0.5 is used to perform standard map relaxation optimization.
[0040] The graph relaxation algorithm is specifically:
[0041]
[0042] Among them, N f Representative node e i The number of nodes pointed to, N t Indicates that it points to node e i The number of nodes, α represents the confidence in other nodes pointed by the node, and p i , p j , p k Respectively represent the node e i , e j , e k Position in the empirical map, Δp ij , Δp ki Respectively represent the slave node e i To node e j and from node e k To node e i The required position and orientation changes.
[0043] An experience map reuse system, comprising:
[0044] The local view cell improvement module is used to improve the local view cells using the incremental binary search tree and dynamic island mechanism to obtain environmental perception information;
[0045] The pose cell network construction module is used to construct the pose cell network, convert the environmental perception information and odometer information into the expression of the robot's position in the environment, and output the network state of the pose cell;
[0046] The experience mapping module is used to update the experience map according to the current state of the pose cell network;
[0047] First, the offline storage domain of the experience map loads the reconstruction module, which is used to save all files of the current experience map and reconstruct all data structures of the historical experience map;
[0048] The map fusion optimization module is used to optimize the robot's position in the map when the local view cell detects a position closed loop; when the local view cell detects that the position closed loop is the overlapping position of the new map and the historical map, the new map and the historical map are merged.
[0049] The present invention has the following beneficial effects and advantages:
[0050] 1. The present invention uses an incrementally constructed binary search tree to save image features and provides fast image retrieval, which can quickly and accurately retrieve images similar to the current image, greatly improving the efficiency of position recognition.
[0051] 2. The present invention utilizes a dynamic island mechanism and adopts the idea of sequence matching to complete loop verification, and only selects the most similar island center as the final decision image, thereby avoiding mismatching. At the same time, it also better avoids competition between similar images that are close in time, greatly improving the accuracy of position recognition.
[0052] 3. Compared with the original RatSLAM method, the calculation speed of loop detection in the present invention is greatly improved, and the time complexity of loop detection is reduced from linear to logarithmic.
[0053] 4. The present invention realizes offline storage, loading and reconstruction of experience maps, and fusion of old and new experience maps, so that experience maps can be reused between robots, which can greatly improve the energy efficiency of environmental mapping, and can also enable robots with limited battery capacity to perceive the large environment through environmental zoning. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 A framework diagram of an experience map reuse method of the present invention;
[0055] Figure 2Flow chart of the improved local view cell position recognition method of the present invention;
[0056] Figure 3 Schematic diagram of the incremental binary search tree structure of the present invention;
[0057] Figure 4 Schematic diagram of the dynamic island of the present invention;
[0058] Figure 5 Schematic diagram of offline storage and loading reconstruction of the experience map of the present invention;
[0059] Figure 6 The present invention intends to merge the old and new maps;
[0060] Figure 7 A graph showing the change in single-frame image recognition time versus the number of image frames of the present invention;
[0061] Figure 8 Comparison diagram of the single-machine position estimation error of the present invention with the pure odometer and the original RatSLAM method;
[0062] Fig. 9 A comparison diagram of the environment experience map and the real trajectory of a single machine of the present invention;
[0063] Fig.10 Schematic diagram of map fusion of the present invention;
[0064] Fig.11 A comparison diagram of the experience map obtained by map reuse in the present invention and the experience map obtained by continuous exploration on a single machine. DETAILED DESCRIPTION
[0065] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.
[0066] The present invention discloses an empirical map reuse method, which utilizes an incrementally constructed binary search tree model and a dynamic island mechanism to establish an accurate and fast location recognition method, improves the visual perception part of the RatSLAM of brain-like navigation, and also provides overlapping area detection of new and old maps for map reuse. The improved graph relaxation method is used to update the formula to achieve the fusion of new and old empirical maps at the detected overlapping positions.
[0067] like Figure 1 As shown, a method for reusing empirical maps includes five modules: local view cells improved by incremental binary search tree and dynamic island mechanism, pose cell network, empirical mapping, offline storage and loading reconstruction of prior empirical maps, and map fusion.
[0068] (1) Improved local view cells, using incremental binary search trees and dynamic island mechanisms to improve the retrieval speed and recognition accuracy of local view cells for environmental locations; if the current scene is identified as a new scene, a new local visual perception cell is established to record the ID of the scene, and an excitatory connection is established between the local visual perception cell and the most active pose cell unit in the pose cell network; conversely, if the current scene is identified as a scene that has been experienced before, the corresponding local visual perception cell is found, and energy is injected into the network through the excitatory connection between the local visual perception cell and the pose cell network, thereby changing the energy distribution of the current pose cell network.
[0069] like Figure 2 As shown, the local view cell improved by using the incremental binary search tree and the dynamic island mechanism includes the following steps:
[0070] Step 1: Use the ORB algorithm to extract features from the scene image collected by the robot, extract N ORB feature points, calculate the BRIEF descriptor of each feature point, and generate N m-dimensional binary feature descriptor sets of the image, where m is 256;
[0071] Step 2: Image I i All feature descriptors are input into the binary search tree for retrieval to determine whether there is one that matches image I. i Similar images. If image I does not exist in the binary tree j and the query image I i The similarity score between i ,I j )>τ 1 , then insert the image I into the binary search tree i All features update the binary search tree; otherwise, image I j As candidate images, the similarity scores between images are calculated by the following formula:
[0072]
[0073] Here n i,j I is the query image i With image I j The number of feature points that are successfully matched in the search tree, N is the total number of feature points in the query image. 1 is a dynamic threshold, which is I i The mean similarity score with the previous n frames depends on the frame rate of the image: the higher the frame rate, the larger n is. It is calculated as follows:
[0074]
[0075] This threshold screening is to ensure that Ij It's a familiar scene.
[0076] Step 3: Collect and image I i All candidate images similar to I are screened again to remove the images with lower scores among these similar images. i Sort the similarity from high to low to get an ordered image set C i = {I i1 ,I i2 ,...,I in}, where I ij Indicates that i The jth most similar image. Their similarity scores are normalized and calculated as follows:
[0077]
[0078] If s′(I i ,I ij )<τ 2 If the image is not found, it will be discarded, otherwise it will be retained, and finally an ordered set of candidate images will be obtained. For subsequent analysis. Threshold τ 2 Generally, 0.4 is selected to ensure integrity and efficiency.
[0079] Step 4: Utilize Alternative Sets All images in I ij , creating dynamically managed islands The islands Represents images with timestamps between m and n to complete the final closed-loop decision. Each island has a representative image, called the island center, which represents the image most similar to the query image in the timestamp range [m, n]. Select the island most similar to the query image, and use the image corresponding to its island center as the final closed-loop decision result.
[0080] like Figure 3 As shown, a binary search tree has the following structure and operations:
[0081] Leaf node: stores the BRIEF descriptor set calculated from the feature points extracted from the image A subset of j is the jth BRIEF descriptor of the image, a dimensional binary vector, each descriptor d j The index of the image it represents will be stored for subsequent image retrieval. Different leaf nodes will store the feature set Divided into different areas, so that each descriptor belongs to a separate leaf node;
[0082] Root node and intermediate node N i : Stores bit index k i , k i ∈(0,1,…,dim(d)-1)), where dim(d) is the descriptor dimension. Each index value can only appear once in the tree to ensure the uniqueness of the index and limit the depth of the tree to below the descriptor dimension.
[0083] Insert operation: For any descriptor d j According to d j [k i ] is stored in the left or right subtree respectively, d j [k i ] is d j At index k i If d j [k i ] is 0, then enter the left subtree, otherwise enter the right subtree, and finally store it in the leaf node;
[0084] Retrieval operation: For the descriptor d to be queried q , starting from the root node, according to d q [k i ]Search the branch until the leaf node L * Until then, the descriptor to be queried is compared with all the descriptors stored in the leaf node one by one, and the most similar descriptor d is found according to a certain threshold. * .
[0085] like Figure 4 As shown, the dynamic island mechanism is to use the alternative set All images in I ij , creating dynamically managed islands The islands Represents images with timestamps between m and n to complete the final closed-loop decision. Each island has a representative image, called the island center, which represents the image most similar to the query image in the timestamp range [m, n]. Select the island most similar to the query image, and use the image corresponding to its island center as the final closed-loop decision result. The specific steps for establishing closed-loop verification using the dynamic island mechanism are as follows:
[0086] Step 1: Right All images in I ij , evaluates whether its timestamp is located on an existing island If yes, image I ij Set as The center of the island will be ijThe first k frames and the last k frames of are added to the island. In order to adapt to the size of the island, when the current island size exceeds the limit, the images far away from the center of the island are moved out of the island and updated. If I ij If it does not exist in the existing island, a new island is created. The new island consists of I ij and its previous k frames and the next k frames, the center of the island is I ij Here the island size is limited to 9, and k=2.
[0087] Step 2: Check whether all islands intersect with other islands. If so, crop the border of the island to ensure that the same image only appears in one island.
[0088] Step 3: Perform a global evaluation for each island. The global score is calculated as follows:
[0089]
[0090] Step 4: Select the island with the highest global score and use the image corresponding to the center of the island as the closed-loop decision.
[0091] (2) The pose cell network is a theoretical model established based on the neural coding principles of position cells and head heading cells in the rat brain. It converts environmental perception information and odometry information into the expression of the robot's position in the environment. It is mainly driven by odometry information and estimates the robot's pose through path integration. It also accepts visual perception information from local view cells and corrects the error of pose estimation by injecting energy from local view cells into the pose cell network.
[0092] (3) Experience mapping: Based on the current state of the posture cell network, an environment cognitive map with topological relationships called the experience map is updated; this map consists of experience nodes e i and the connections between nodes ij Composition, experience node e i = {P i ,Vi,p i}, where P i represents the position of the activated pose cell (x′ i ,y′ i ,θ′ i ), p i Indicates the current position in the experience map (x i ,y i ,θ i ), V i Represents the corresponding local view cell, which stores the corresponding environment image information; node e i With node e j Connections between Among them, Δp ij Indicates that from experience i To experience j The required position and orientation changes, From experience i To experience j The time required; calculate the current robot state {P j ,V j} and all experience nodes e in the experience map i The distance between the corresponding robot states:
[0093] S i =μ p |P i -P j |+μ v |B i -V j |,i=1,…,M
[0094] where μ p and μ v They represent the weight coefficient of the degree of stimulation of the pose cells and the weight coefficient of the local visual perception cells, respectively, indicating the proportion of odometer information and visual information in the process of robot map building. Different weight coefficients can be set as needed; M is the number of experience nodes in the current experience map. Find the minimum S i S min If S mim Greater than the threshold τ S , then establish a new experience node e j , and at the new experience node e j With the previous newly created experience node e i Establish a connection ij :
[0095] l ij ={Δp ij ,Δt ij}
[0096] Among them, Δp ij Indicates that from experience i To experience j The required change in position and orientation, Δt ij From experience i To experience j The time required.
[0097] like Figure 5 As shown, we first experience the offline storage and loading reconstruction of the map, and then j and visual perception cell state V jWrite the binary file Map.pcn to the robot's position information in the experience map p i , and all edges l ij Write the binary file Map.bmp, write the entire binary search tree to the binary file Map.tr in a depth-first manner, load all files that save the experience map in the same way, and reconstruct all data structures related to the previous experience map so that the robot can have the previous experience map.
[0098] like Figure 6 As shown, map fusion optimization, map fusion optimization adopts the following improved graph relaxation algorithm position update rule:
[0099]
[0100] Among them, N f Represents e i The number of nodes pointed to, N t Indicates that it points to node e i The number of nodes, α represents the confidence in other nodes pointed to by the node. The larger α is, the higher the confidence in the old experience is. After the map is updated, the new map nodes will be more close to the previous experience. On the contrary, the smaller α is, the more trust is placed in the new experience. After the map is updated, the old experience will be adjusted more to be close to the new experience.
[0101] Before detecting the overlapping position between the new and old experience maps, the robot takes the last state of the old experience map as the initial state and starts to rebuild the experience map. The new experience map has no connection with the old experience map. At this time, the graph relaxation algorithm is called at a certain frequency to correct the map, the position update coefficient is set to α=0.5, and standard map optimization is performed.
[0102] When the overlap of the new and old experience maps is detected, that is, when the current image is identified as the same as a scene stored in the previous map, a link is established between the current experience and the experience in the old map, and α=1 is set in the position update formula to optimize the map. In this case, the previous experience is considered to be completely correct, and the newly generated experience is converted to the coordinate system of the previously saved map to form a unified cognition. The number of iterations is set to the number of newly created experience nodes to ensure that all newly generated experiences are converted to the coordinate system of the previously saved map.
[0103] After the fusion of the old and new maps, the standard map relaxation optimization is performed using the position update rule with α = 0.5. In this case, the robot has a unified perception of the environment. Therefore, α = 0.5, the confidence in the previous experience is equal to the confidence in the current experience, resulting in a balanced position adjustment for the entire map.
[0104] like Figure 7 As shown, the single-frame image recognition time of the present invention is significantly faster than that of the original RatSLAM system. The original RatSLAM single-frame image recognition time theoretically increases linearly with the increase of the number of image frames, while the single-frame image recognition time of the present invention theoretically increases logarithmically with the increase of the number of image frames. The method of the present invention greatly reduces the time complexity of image processing.
[0105] like Figure 8 As shown in FIG. 1 , the position estimation error of the method of the present invention is significantly smaller than that of the pure odometer and the original RatSLAM method; Fig. 9 The cognitive map drawn by the method of the present invention is close to the real trajectory. The method of the present invention greatly improves the accuracy of environmental map drawing.
[0106] like Fig.10 As shown, at t = 0s, the robot loads the previously stored experience map. At this time, it can have the previous experience and reuse the previous experience map. The robot's position is at the position it was at the last moment. Then the robot begins to explore the orange area. From t = 7s to t = 20s, it explores the new environment and generates a new experience map, which is not connected to the previous map. At t = 27s, it can be clearly seen that the newly generated experience map has changed, and the current position of the robot has returned to the position in the previous map, indicating that the robot has recognized the previous scene. At this time, the map is fused and a globally consistent map is obtained. From t = 49s to t = 107s, the robot's newly created experience map has been fused with the previous map, and can be incrementally expanded and optimized on the basis of the fusion. The RatSLAM algorithm based on map reuse proposed in the present invention can not only realize reuse and positioning under offline maps, but also perform incremental expansion on the basis of fusion, and the map building efficiency is significantly improved.
[0107] like Fig.11 As shown in the figure, under the two data sets, the global experience map obtained by map reuse is basically consistent with the map constructed by a single robot, which proves the accuracy of the proposed algorithm in mapping.
Claims
1. A method for reusing an experience map, characterized in that: The following steps are involved: 1) Using incremental binary search tree and dynamic island mechanism to improve local view cells and obtain environmental perception information; 2) Construct a pose cell network to convert the environmental perception information and the collected odometer information into the expression of the robot's position in the environment, and output the network state of the pose cells; 3) Update the experience map according to the current posture cell network state; 4) Save all files of the current experience map and reconstruct all data structures of the historical experience map; 5) When the local view cell detects a position loop, the robot's position in the map is optimized; when the local view cell detects that the position loop is the overlapping position of the new map and the historical map, the new map and the historical map are merged.
2. The method for reusing an experience map according to claim 1, characterized in that: The step 1) comprises the following steps: 1.1) Use the ORB algorithm to extract features from the scene image, extract N ORB feature points, calculate the BRIEF descriptor of each feature point, and generate a set of N m-dimensional binary feature descriptors of the image; 1.2) Image I i All feature descriptors are input into the binary search tree for retrieval, and based on the similarity score, it is determined whether there is a feature that is similar to image I. i Similar images, if image I j and the query image I i If the similarity score between them is greater than the threshold, then image I j For image I i Similar images of image I j As an alternative image, otherwise, insert image I into the binary search tree i All features are used to update the binary search tree; 1.3) Collection and Image I i All candidate images similar to I i Sort the similarity from high to low to get an ordered image set C i = {I i1 , I i2 , ..., I in }, where I ij Indicates that i The jth most similar image, j = 1, 2, ..., n, normalizes the similarity scores of the images in the set, removes images with scores less than the threshold, and obtains an ordered set of candidate images 1.4) Using Alternative Image Sets All images in I ij , creating dynamically managed islands Among them, the islands Represents images with timestamps between m and n. Each island has a representative image, called the island center, which represents the image most similar to the query image in the timestamp range [m, n]. The island most similar to the query image is selected, and the image corresponding to its island center is used as the final closed-loop decision result.
3. The method for reusing an experience map according to claim 2, characterized in that: The similarity score s(I i , I j )for: Among them, n i,j I is the query image i With image I j The number of feature points that are successfully matched in the search tree, N is the total number of feature points in the query image, and τ1 is the number of feature points that are successfully matched in the search tree. i A dynamic threshold based on the mean similarity score of the previous n frames, 4. The method for reusing an experience map according to claim 2, characterized in that: The normalization of the similarity scores of the images in the set is specifically as follows: Among them, s′(I i , I ij ) is the normalized similarity score.
5. The method for reusing an experience map according to claim 2, characterized in that: The binary search tree comprises: Leaf node: used to store the BRIEF descriptor set calculated from the feature points extracted from the image A subset of j is the jth BRIEF descriptor of the image, which is an m-dimensional binary vector. Each descriptor d j The index of the image it represents is stored, and different leaf nodes store the feature set Divided into different areas, so that each descriptor belongs to a separate leaf node; Root node and intermediate node N i : used to store bit index k i , k i ∈(0, 1, ..., dim(d)-1)), where dim(d) is the descriptor dimension.
6. The method for reusing an experience map according to claim 2, characterized in that: The binary search tree performs the following operations: Insert operation: For any descriptor d j According to d j [k i ] is stored in the left or right subtree respectively, d j [k i ] is d j At index k i If the value of d j [k i ] is 0, then enter the left subtree, otherwise enter the right subtree, and finally store it in the leaf node; Retrieval operation: For the descriptor d to be queried q , starting from the root node, according to d q [k i ]Search the branch until the leaf node L * Until then, the descriptor to be queried is compared with all the descriptors stored in the leaf node one by one, and the most similar descriptor d is found according to a certain threshold. * .
7. The method for reusing an experience map according to claim 2, characterized in that: The step 1.4) comprises the following steps: 1.4.1) Yes All images in I ij , evaluates whether its timestamp is located on an existing island Within the time interval, if I ij Exists in the existing island, image I ij Set as The center of the island will be ij The first k frames and the last k frames of are added to the island. In order to adapt to the size of the island, when the current island size exceeds the limit, the images far away from the center of the island are moved out of the island and updated. time interval; if I ij If it does not exist in the existing island, a new island is created. The new island consists of I ij and its previous k frames and the next k frames, the center of the island is I ij ; 1.4.2) Check whether all islands intersect with other islands. If so, crop the borders of the islands that intersect to ensure that the same image only appears in one island; 1.4.3) Perform a global evaluation on each island and calculate the global score 1.4.4) Select the island with the highest global score and use the image corresponding to the center of the island as the closed-loop decision.
8. The method for reusing an experience map according to claim 1, characterized in that: The fusion of the new map and the historical map in step 5) is specifically as follows: 5.1) Before detecting the overlap between the new and old experience maps, the robot takes the last state of the historical experience map as the initial state and starts to rebuild the experience map. The new experience map has no connection with the historical experience map. At this time, the graph relaxation algorithm is called at a certain frequency to correct the map, and the position update coefficient is set to α = 0.5, and standard map optimization is performed; 5.2) When the overlap between the new and old experience maps is detected, that is, when the current image is identified as being the same as a scene stored in the previous map, a link is established between the current experience and the experience in the historical map, and α=1 is set in the position update formula to optimize the map; 5.3) After the fusion of the new and old maps is completed, the position update rule with α=0.5 is used to perform standard map relaxation optimization.
9. The method for reusing an experience map according to claim 8, characterized in that: The graph relaxation algorithm is specifically: Among them, N f Representative node e i The number of nodes pointed to, N t Indicates that it points to node e i The number of nodes, α represents the confidence in other nodes pointed by the node, and p i , p j , p k Respectively represent the node e i , e j , e k Position in the empirical map, Δp ij , Δp ki Respectively represent the slave node e i To node e j and from node e k To node e i The required position and orientation changes.
10. An experience map reuse system, characterized in that: include: The local view cell improvement module is used to improve the local view cells using the incremental binary search tree and dynamic island mechanism to obtain environmental perception information; The pose cell network construction module is used to construct the pose cell network, convert the environmental perception information and odometer information into the expression of the robot's position in the environment, and output the network state of the pose cell; The experience mapping module is used to update the experience map according to the current state of the pose cell network; First, the offline storage domain of the experience map loads the reconstruction module, which is used to save all files of the current experience map and reconstruct all data structures of the historical experience map; The map fusion optimization module is used to optimize the robot's position in the map when the local view cell detects a position closed loop; when the local view cell detects that the position closed loop is the overlapping position of the new map and the historical map, the new map and the historical map are merged.
