Autonomous Exploration Method, Device and Medium for Tensor Field-Driven Hierarchical Path Planning

Abstract

The present invention discloses an autonomous exploration method, device and medium for tensor field-driven hierarchical path planning. The present invention includes constructing a scene tensor field based on depth point cloud information and obtaining a set of degenerate points P; solving the online traveling salesman problem based on the set of degenerate points P to obtain a sequence of degenerate points Q d ; grouping the boundary set F of the unexplored space to obtain a boundary grouping set Q F and splitting to filter the boundaries located outside the room; according to the boundary grouping set Q F determine the next boundary point to be visited, calculate a global path T through the A* algorithm, use the global path T to influence and update the tensor field; along the particle advection of the tensor field, calculate the linear velocity and angular velocity of the robot, and guide the robot to the target. The present invention can eliminate a large amount of time-consuming recalculation of paths, is suitable for robot exploration in complex indoor scenes, and can greatly improve the efficiency of autonomous exploration while ensuring the integrity of exploration.

Description

Technical Field

[0001] The present invention relates to the technical field of robot path planning, and in particular to an autonomous exploration method, device and medium for tensor field driven hierarchical path planning. Background Art

[0002] Autonomous mobile robots are an important research topic in robotics and have been widely used in industrial and agricultural production, transportation, services, medical and military fields. With the complexity of the application environment, humans have put forward higher requirements for the autonomy and intelligence of mobile robots. Especially in unknown and complex environments, the autonomous exploration of robots has become a difficult and important issue. There are several common methods of autonomous exploration of robots:

[0003] 1) LiDAR-based methods: LiDAR sensors can provide robots with a 360-degree field of view (FOV), but they are relatively expensive. Moreover, most ground robots need expensive LiDAR for navigation and exploration for autonomous exploration. Since RGB-D cameras only provide one perspective of the environment, rather than 360-degree coverage like LiDAR, autonomous exploration based on RGB-D cameras is inherently complex and brings a series of problems. If these methods are directly used on robots with RGB-D, they usually fail.

[0004] 2) RGB-D-based methods: With the widespread use of RGB-D cameras such as RealSense and Kinect in robots, robots can perceive the environment at a low cost. Compared with LiDAR, RGB-D can be more compact, lightweight and cheaper. However, RGB-D cameras have the following problems: Route detour problem: RGB-D-based robots can only perceive environmental information in a certain direction at the same time. If the route planning is unreasonable, it is very easy to turn back on a large scale, resulting in low exploration efficiency. Since RGB-D has a smaller field of view than LiDAR, it is more difficult for robots to avoid local obstacles, especially in complex scenes or narrow spaces, where collisions are very likely to occur. Therefore, due to the limitation of FOV, there is a problem of small perception range, low efficiency and easy collision. At present, most RGB-D-based methods cannot overcome their shortcomings in perception range, resulting in frequent turns and detours to ensure the integrity of the exploration scene, and the exploration efficiency is generally not high.

[0005] 3) Boundary-based exploration method: It guides the robot's exploration through the boundary between the free space and the unexplored space, which can help the robot conduct a more complete exploration. However, due to the lack of global optimization, the greedy nature of these methods often leads to wrong decisions. Generally, boundary-based exploration methods prefer the boundary closer to the robot as the next target without considering subsequent actions. Due to its greedy strategy, it will generate overly long trajectories. Especially in complex indoor scenes, it often enters and exits the same room multiple times, causing the route to detour, thus affecting the exploration efficiency.

[0006] On this basis, Xu et al. (Xu K, Zheng L, Yan Z, Yan G, Zhang E, Niessner M, Deussen O, Cohen-Or D, Huang H. Autonomous reconstruction of unknown indoor scenes guided by time-varying tensor fields. ACM Transactions on Graphics (TOG), 2017, 36(6): 1–15.) disclosed a tensor field method for smooth scene reconstruction. This method regards the topological skeleton of the tensor field as an undirected graph, which contains all degenerate points and the separatrices connecting them. The time-varying tensor field can be updated in real time to directly guide the robot's movement. It is proved that although the FOV of RGB-D is limited, the topological structure of the tensor field is sufficient to achieve effective global path routing in a partially reconstructed scene. However, this method has poor effects when the scene is large and complex, because the routing on the static degenerate points in the tensor field cannot support global path optimization and easily leads to incomplete final scene exploration. Summary of the Invention

[0007] The technical problem to be solved by the present invention: Aiming at the above problems of the prior art, the present invention provides an autonomous exploration method, device and medium for tensor field-driven hierarchical path planning. Based on this, the present invention improves the robot navigation framework based on the tensor field, can better utilize the local topological structure for global path optimization, can eliminate a large number of time-consuming recalculations of the path, is suitable for robot exploration in complex indoor scenes, and can greatly improve the efficiency of autonomous exploration while ensuring the exploration integrity.

[0008] To solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0009] An autonomous exploration method for tensor field-driven hierarchical path planning, including:

[0010] S101. Construct a scene tensor field according to the depth point cloud information sensed by the input RGB-D camera, and detect degenerate points on the scene tensor field to obtain a set of degenerate points P;

[0011] S102. Solve the online traveling salesman problem based on the set of degenerate points P to obtain a sequence of degenerate points Q corresponding to the optimal visiting order d ;

[0012] S103. Rely on the sequence of degenerate points Q d Group the boundary set F of the unexplored space to obtain a boundary grouping set Q F ;

[0013] S104. For each boundary grouping F that spans rooms in the boundary grouping set Q F , split it to obtain the boundaries outside the room, filter the boundaries outside the room and regard them as new degenerate points and insert them into the set of degenerate points P; k

[0014] S105. Determine the next boundary point to be visited according to the boundary grouping set Q F , calculate a global path T through the A* algorithm, and use the global path T to influence and update the tensor field;

[0015] S106. Along the particle advection of the tensor field, calculate the linear velocity and angular velocity of the robot to guide the robot to the target.

[0016] Optionally, the set of degenerate points P obtained by detecting degenerate points on the scene tensor field in step S101 is obtained by merging two parts: the set of degenerate points obtained on the scene tensor field and the extended set of degenerate points obtained by extending the degenerate points to cover the entire scene.

[0017] Optionally, step S102 includes:

[0018] S201. Find the degenerate point p0 closest to the robot position c r in the set of degenerate points;

[0019] S202. Based on the degenerate point p0 closest to the robot position c r , solve the online traveling salesman problem shown in the following formula based on the set of degenerate points P to obtain a sequence of degenerate points Q corresponding to the optimal visiting order d :

[0020]

[0021] In the above formula, represents the online traveling salesman problem, p0 is the degenerate point closest to the robot position c r , is the sequence of degenerate points Q dthe first degenerate point in is the i-th degenerate point in the degenerate point sequence Q d ; is the (i + 1)-th degenerate point in the degenerate point sequence Q d , dist is the path length between two degenerate points, and n is the number of degenerate points in the degenerate point sequence Q d , where the degenerate points and both come from the degenerate point set P.

[0022] Optionally, step S103 includes:

[0023] S301. For any boundary point f in the boundary set F m , calculate the Euclidean distance between the boundary point f m and any degenerate point d in the degenerate point sequence Q , and determine the degenerate point

[0024] with the closest Euclidean distance; m S302. Based on the grouping function of the following formula, group the boundary point f into the corresponding group F k of the degenerate point

[0025]

[0026] In the above formula, is the grouping function, d f is the threshold, f m is the boundary point, and the function expression of the minimum distance d k (f m ) is:

[0027]

[0028] In the above formula, is the degenerate point with the closest Euclidean distance to the boundary point f m , k is an element in the set , and the set represents the index set in the degenerate point sequence Q d ;

[0029] S303. Add the boundary point f m satisfying the following formula to the degenerate point set P:

[0030] |f m -d k (f m )|>d f ,

[0031] In the above formula, df is the threshold value.

[0032] Optionally, the boundary outside the room obtained by splitting in step S104 is {F k -S(F k )}, where S(F k ) is the splitting function, and:

[0033] S(F k )={f k |f k ∈F k ,ν≥d g ∨(μ <d p ∧ν <d g )},

[0034] In the above formula, f k Group F for the boundary k The boundary point in the figure, ν is the robot's position c r With group F k The distance between them, μ is the boundary point f k The path length between the robot and the p and d g is the threshold, ∨ represents the OR operation, and ∧ represents the AND operation.

[0035] Optionally, in step S105, using the global path T to influence and update the tensor field includes:

[0036] S401, projecting the grid cells in the input grid map onto the floor, and performing a specified distance d on the center of the projected boundary points s The farthest point sampling is used to select a set of plane constraint points and sort them by path direction to obtain the constraint point set Pc, and the global path T is projected to the constraint point set Pc;

[0037] S402: for any two adjacent constraint points p in the constraint point set Pc, c-1 (x c-1 ,y c-1 ) and p c (x c ,y c ) Calculate the constraint point p c (x c ,y c ) of the regular elements (S c ,T c ), where S c is the x-direction vector of the general element affecting the field, T c is the y-direction vector of the general element affecting the field;

[0038] S403, based on any constraint point p c(x c , y c ) of the conventional elements (S c , T c ) to calculate the constraint point p c (x c , y c ) of the base tensor field T c (p);

[0039] S404, sum the base tensor fields T c (p) of all constraint points through Gaussian radial basis functions to obtain the final tensor field.

[0040] Optionally, the calculation function expression of the conventional elements (S c , T c ) of the constraint point p c (x c , y c ) in step S402 is:

[0041] S c = x c - x c-1 ,

[0042] T c = y c - y c-1 ,

[0043] In the above formula, (x c , y c ) is the plane coordinate of the constraint point p c (x c , y c ), and (x c-1 , y c-1 ) is the plane coordinate of the constraint point p c-1 (x c-1 , y c-1 ).

[0044] Optionally, the function expression for summing to obtain the final tensor field in step S404 is:

[0045]

[0046] In the above formula, T(p) is the final tensor field, T c (p) is the base tensor field T c (x c , y c ) of each constraint point p c (p), d is the attenuation coefficient, σ is the Gaussian bandwidth, p is the passable point on the robot motion plane, and p c is the constraint point.

[0047] In addition, the present invention also provides an autonomous exploration device for tensor field-driven hierarchical path planning, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the autonomous exploration method for tensor field-driven hierarchical path planning.

[0048] In addition, the present invention also provides a computer-readable storage medium, in which a computer program is stored, and the computer program is used to be programmed or configured by a microprocessor to execute the autonomous exploration method for tensor field-driven hierarchical path planning.

[0049] Compared with the prior art, the present invention mainly has the following advantages:

[0050] 1. Low equipment cost: When using the method of the present invention, the robot only needs to use a robot equipped with an RGB-D camera for navigation and exploration, and the cost is lower compared with the Lidar-based solution.

[0051] 2. Strong global perception ability: The method of the present invention only uses an RGB-D camera for positioning, navigation and exploration operations, overcomes the field of view (Fov) limitation, and enhances the global perception ability of the robot.

[0052] 3. High exploration efficiency: The method of the present invention uses dynamically expanding degenerate points to provide a good reference for forming an optimized scene structure topology, enhances the global perception ability of the robot, and thus improves the efficiency. At the same time, an exploration strategy based on boundary grouping is introduced to perform fine-grained local scanning. These two strategies together constitute our hierarchical exploration method, which can better utilize the local topological structure for global path optimization, can eliminate a large number of time-consuming recalculations of paths, is suitable for robot exploration in complex indoor scenes, and can greatly improve the efficiency of autonomous exploration while ensuring the exploration integrity. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the basic process of the method in the embodiment of the present invention.

[0054] Figure 2 It is a schematic diagram of the set of degenerate points P obtained in the embodiment of the present invention.

[0055] Figure 3 It is a schematic diagram of the access order of the set of degenerate points obtained in the embodiment of the present invention.

[0056] Figure 4 It is a schematic diagram of the boundary set (yellow triangle) in the embodiment of the present invention.

[0057] Figure 5 It is a comparison schematic diagram of non-grouping splitting (a) and grouping splitting (b) in the embodiment of the present invention.

[0058] Figure 6 Schematic diagram of calculating the global path by the A* algorithm in the embodiment of the present invention.

[0059] Figure 7 Schematic diagram of the particle advection path planning along the tensor field in the embodiment of the present invention.

[0060] Figure 8 Schematic diagram of the exploration path comparison with and without using the method of the embodiment of the present invention.

[0061] Figure 9 Schematic diagram of the path obtained by using the method of this embodiment in different scenarios. Specific implementation manner

[0062] As Figure 1 shown, this embodiment provides an autonomous exploration method for tensor field-driven hierarchical path planning, including:

[0063] S101. Construct a scene tensor field according to the depth point cloud information sensed by the input RGB-D camera, and detect degenerate points on the scene tensor field to obtain a degenerate point set P; use the degenerate points in the tensor field as anchor points, and generate a rough topology by solving the online TSP problem, which significantly improves the effectiveness of the tensor field-driven method in complex environments;

[0064] S102. Solve the online traveling salesman problem based on the degenerate point set P to obtain a degenerate point sequence Q corresponding to the optimal visit order d ;

[0065] S103. Rely on the degenerate point sequence Q d to group the boundary set F of the unexplored space to obtain a boundary grouping set Q F ; Relying on the dynamic grouping of degenerate points can improve the exploration efficiency while ensuring the scanning integrity;

[0066] S104. For each boundary grouping F F that spans a room in the boundary grouping set Q k , split it to obtain the boundary outside the room, filter the boundary outside the room and regard it as a new degenerate point and insert it into the degenerate point set P;

[0067] S105. Determine the next boundary point to be visited according to the boundary grouping set Q F , calculate a global path T by the A* algorithm, and use the global path T to affect and update the tensor field;

[0068] S106. Along the particle advection of the tensor field, calculate the linear velocity and angular velocity of the robot to guide the robot to reach the target. After step S106, steps S101 to S106 can be continuously looped until the robot scans the entire scene.

[0069] In step S101, constructing a scene tensor field based on the depth point cloud information sensed by the input RGB-D camera and detecting degenerate points on the scene tensor field are existing methods. For details, please refer to the literature (Xu K, Zheng L, Yan Z, Yan G, Zhang E, Niessner M, Deussen O, Cohen-Or D, Huang H. Autonomous reconstruction of unknown indoor scenes guided by time-varying tensor fields. ACM Transactions on Graphics (TOG), 2017, 36(6): 1–15.). Therefore, the implementation details will not be elaborated here. The tensor field T on the two-dimensional plane D is a smooth tensor-valued function that associates a two-dimensional tensor T(p) with each point p ∈ D, which can be expressed as:

[0070]

[0071] In the above formula, τ 11 (p) and τ 22 (p) are the elements of the two-dimensional tensor respectively, and can be expressed by the general formula τ ij . When and only when τ ij = τ ji , the tensor is symmetric. The tensor T mentioned in the method of this embodiment is symmetric and can be uniquely decomposed into an isotropic part S and an anisotropic part A:

[0072]

[0073] In the above formula, λ is the coefficient of the unit vector corresponding to the isotropic part, μ is the eigenvalue of A, and θ is the tangent direction angle of a point on the projected contour of the obstacle plane, where μ > 0, and the eigenvalues of the anisotropic part A are ±μ.

[0074] Denote the anisotropic part A of point p as A(p). When A(p) ≠ 0, A(p) is equivalent to two orthogonal eigenvector fields E1(p) and E2(p):

[0075] E1(p) = μ(p)e1(p),

[0076] E2(p) = μ(p)e2(p),

[0077] In the above formula, μ(p) is the point on the obstacle plane projection contour of point p, and e1(p) and e2(p) are the unit eigenvectors corresponding to the eigenvalues μ and -μ respectively. Therefore, E1(p) and E2(p) are the main eigenvector and the secondary eigenvector of the anisotropic part A. In this method, the robot is guided by the main eigenvector E1(p). A point p is a degenerate point of the tensor field T(p) if and only if the anisotropic part A of point p satisfies A(p) = 0. The rough topology, i.e., the set of degenerate points, can be constructed using the degenerate points. In this embodiment, the set of degenerate points P obtained by detecting degenerate points on the scene tensor field in step S101 consists of the set of degenerate points obtained on the scene tensor field (as shown by the green circles in Figure 2 ) and the set of extended degenerate points obtained by extending the degenerate points to cover the entire scene (as shown by the red circles in Figure 2 ). Suppose the set of degenerate points obtained on the scene tensor field is P = p0, p1, …, p i , p i ∈ D, and the set of extended degenerate points obtained by extension is P * = p0, p1, …, p j , and there is If a degenerate point p i is visited, we delete these points from P ← P - p i . As shown in Figure 3 , where the gray circles represent the visited degenerate points, the green circles represent the degenerate points obtained on the scene tensor field, the red circles represent the extended degenerate points, and the numbers in the circles indicate the order of the degenerate points. Therefore, the merged set of degenerate points P can be expressed as:

[0078] P ← P ∪ P * = p0, p1, …, p n , n = i + j,

[0079] During the exploration process, a set of degenerate points (the set of degenerate points) is obtained from the scene tensor field, and the degenerate points are extended (the set of extended degenerate points) to cover the entire scene. Secondly, based on the degenerate points, the online TSP is solved by dynamic expansion to obtain an optimal sequence of degenerate points. We denote the path length between degenerate points i and j as dist(p i , p j ). Given the degenerate points p0, p1, …, p n , the online traveling salesman problem (TSP problem) is defined as finding a sequence Π = π1, …, π n for I = 1, …, n, starting from p0, such that the total access route of the permutation is the shortest. Then the degenerate points Is selected as the next target position. Note that in the online traveling salesman problem (TSP problem), the optimal visiting order is found and it ends at any point instead of returning to the starting point, which can prevent the solver of the online traveling salesman problem (TSP problem) from generating overly long trajectories. In this embodiment, step S102 solves the online traveling salesman problem (TSP problem) based on the set of degenerate points P to obtain the sequence of degenerate points Q corresponding to the optimal visiting order d ; The online traveling salesman problem (TSP problem) refers to solving the traveling salesman problem (TSP problem) for continuously updated degenerate points to maintain their optimal visiting order, so as to dynamically generate a rough topology for navigation. Specifically, step S102 includes:

[0080] S201, find the degenerate point p0 in the set of degenerate points that is closest to the robot position c r ; This strategy can prevent the robot from moving a long distance to start visiting.

[0081] S202, based on the degenerate point p0 that is closest to the robot position c r , solve the online traveling salesman problem shown in the following formula based on the set of degenerate points P to obtain the sequence of degenerate points Q corresponding to the optimal visiting order d :

[0082]

[0083] In the above formula, represents the online traveling salesman problem, p0 is the degenerate point closest to the robot position c r , is the first degenerate point in the sequence of degenerate points Q d , is the i-th degenerate point in the sequence of degenerate points Q d , is the (i + 1)-th degenerate point in the sequence of degenerate points Q d , dist is the path length between two degenerate points, n is the number of degenerate points in the sequence of degenerate points Q d , where the degenerate points and both come from the set of degenerate points P. After selecting the starting point, we need to calculate the distance from the current point to the next point as the distance cost for optimizing the objective function. Usually, the Euclidean distance is used to calculate the spatial distance. However, when there are obstacles (such as walls, furniture) between two points, measuring the path length is inaccurate. Therefore, in this embodiment, specifically, A *The algorithm (P.E. Hart, N.J. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum cost paths in graphs. IEEE Trans. Syst. Sci. and Cybernetics, SSC-4(2):100-107, 1968) calculates the path length dist to ensure the shortest path for the entire access sequence. It should be noted that using the A * algorithm to calculate the path length between two points is an existing method. In this embodiment, only the application of this method is involved, and no improvement of this method is involved. Therefore, the implementation details are not elaborated here.

[0084] For exploration calculation, our method is based on a rough topology and scans the scene at a fine-grained level. We have planned an optimal access topology based on degenerate points and can explore along this topology. At the same time, to fully cover the scene, we introduce a boundary-based approach. The boundary is the boundary between the free space and the unexplored space. Eliminating the boundary in the map means that the scene scanning is complete. Therefore, we dynamically group the boundaries around the degenerate points. The key idea is that the global path generated by the topological skeleton of the degenerate points covers the entire scene, providing a reasonable basis for boundary grouping. In addition, the exploration efficiency is improved by accessing groups instead of individual boundaries.

[0085] As Figure 4 shown, step S103 is used to group the boundary set F (yellow triangles) based on the obtained degenerate point sequence Q d (circles with numbers) to obtain a boundary grouping sequence Q F (the boundaries corresponding to each circle with numbers). In this embodiment, step S103 includes:

[0086] S301. For any boundary point f m in the boundary set F, calculate the Euclidean distance between the boundary point f m and any degenerate point d in the degenerate point sequence Q , and determine the degenerate point

[0087] with the closest Euclidean distance. m S302. Based on the grouping function of the following formula, group the boundary point f into the group F k corresponding to the degenerate point k (which can be expressed as F k ← F m ∪ f

[0088]

[0089] In the above formula, is a grouping function, and d f is a threshold value (which can be set according to actual needs. For example, in this embodiment, the specific value is 2m), and f m is a boundary point. The function expression of the minimum distance d k (f m ) is:

[0090]

[0091] In the above formula, is the degenerate point closest to the boundary point f m in terms of Euclidean distance. k is an element in the set , and the set represents the index set in the degenerate point sequence Q d ;

[0092] S303. Add the boundary point f m that satisfies the following formula to the degenerate point set P (which can be expressed as P←P∪f m ):

[0093] |f m -d k (f m )|>d f ,

[0094] In the above formula, d f is the threshold value.

[0095] After grouping, we scan each group F k along the degenerate points. During the scanning process, if the boundary f k ∈F k is visited, it will be deleted from the group F k (F k ←F k -f k ). When the group F k is empty, the degenerate point is also regarded as visited and deleted from the degenerate point list P Then we find the next degenerate point according to the boundary grouping sequence Q F and scan the corresponding group.

[0096] However, simply grouping the boundaries is not enough because obstacles (such as walls) are sometimes located between two boundaries of the same group, and the robot will make a long detour when approaching the two boundaries. As shown in Figure 5 (a), when the robot explores the boundary group F iWhen it is time, go out to scan F i for the remaining boundaries (blue trajectory) and return to the room to access the next group of F j (pink trajectory) is not advisable. To solve this problem, we will perform dynamic splitting when the group spans two rooms. Step S104 is for the set Q of boundary groups F For each boundary group F that spans a room in k , split it to obtain the boundaries outside the room, filter the boundaries outside the room and regard them as new degenerate points and insert them into the set P of degenerate points; specifically, in this embodiment, the boundaries obtained by splitting in step S104 and located outside the room are {F k -S(F k )}, where S(F k ) is the splitting function, and there is:

[0097] S(F k )={f k |f k ∈F k , ν≥d g ∨(μ<d p ∧ν<d g )},

[0098] In the above formula, f k is the boundary point in the boundary group F k , ν is the distance between the position c r of the robot and the group F k , μ is the path length between the boundary point f k and the robot, d p and d g are thresholds, ∨ represents the OR operation, and ∧ represents the AND operation.

[0099] Among them, the calculation function expression of the distance ν between the position c r of the robot and the group F k is:

[0100] ν=||c r ,p k ||,

[0101] Onuo, c r represents the position of the robot, p k represents the position of the corresponding degenerate point of the group F k , and the Euclidean distance between the position c r of the robot and the position p k of the corresponding degenerate point of the group F k is measured. f k ∈F k The path length between and the robot is μ=dist(cr , f k ), similar to the previous text, can be calculated by the existing A * algorithm. As Figure 5 shown in (b) of , after grouping (green circles), the planned trajectory is shorter, which reduces the exploration time.

[0102] In this embodiment, step S105 is used to determine the next boundary point to be visited according to the boundary grouping set Q F and calculate a global path T through the A* algorithm, and use the global path T to affect and update the tensor field. The input of this process is the boundary grouping set Q F , the tensor field, and the grid map as inputs, and the output is the changed tensor field.

[0103] The existing method directly uses the path defined by the particles advected by the field as the robot's motion path. However, when the scene is complex, the path is usually discontinuous and ambiguous. We use the A * algorithm (P.E. Hart, N.J. Nilsson, and B. Raphael. A formal basis for the heuristic determination of minimum cost paths in graphs. IEEE Trans. Syst. Sci. and Cybernetics, SSC-4(2): 100-107, 1968) to find the shortest collision-free path between the current position and the next exploration target point. As Figure 6 shown, a global path (green arrow) is calculated through the A * algorithm, and the yellow triangles in the figure are still boundary points. The fact is that when it faces a large scene, the A * calculation is very time-consuming.

[0104] To reduce the frequency of A* algorithm calculations, we found that the path planned by the A* algorithm can change the surrounding tensor field, and the robot can update the trajectory path along the new advection. As Figure 7 shown, (a) represents a path planned based on the A * algorithm, but blocked by obstacles in the unexplored area. (b) represents the robot moving along the path, and the surrounding tensor field changes, providing a new advection for obstacle avoidance. This positive feedback between the planned path and the tensor field helps to further guide the robot to determine a better motion. Therefore, the frequency of A* algorithm calculations can be significantly reduced.

[0105] In this embodiment, the step of using the global path T to affect and update the tensor field in step S105 includes:

[0106] S401. Project the grid cells in the input raster map onto the floor and perform sampling at the farthest points at a specified distance d (which can be set according to actual needs, for example, d = 0.2m in this embodiment) at the centers of the projected boundary points to select a set of planar constraint points, sort them in the path direction to obtain the constraint point set Pc, and project the global path T onto the constraint point set Pc. s (The value can be taken according to actual needs. For example, in this embodiment, d s = 0.2m).

[0107] S402. For any two adjacent constraint points p(x, y) and p(x, y) in the constraint point set Pc respectively, calculate the conventional elements (S, T) of the constraint point p(x, y), where S is the x-direction vector of the conventional element of the influence field, and T is the y-direction vector of the conventional element of the influence field. c-1 (x c-1 , y c-1 ) and p c (x c , y c ) to calculate the conventional elements (S c (x c , y c ) of the constraint point p c , T c ). Here, S c is the x-direction vector of the conventional element of the influence field, and T c is the y-direction vector of the conventional element of the influence field.

[0108] S403. Based on the conventional elements (S, T) of any constraint point p(x, y), calculate the basis tensor field T(p) of the constraint point p(x, y). c (x c , y c ) of the conventional elements (S c , T c ), calculate the basis tensor field T c (x c , y c ) of the constraint point p c (p).

[0109] S404. Sum up the basis tensor fields T(p) of all constraint points through the Gaussian radial basis function to obtain the final tensor field. c (p).

[0110] In this embodiment, the calculation function expressions for the conventional elements (S, T) of the constraint point p(x, y) in step S402 are: c (x c , y c ) are: c , T c ) are:

[0111] S c = x c - x c-1 ,

[0112] T c = y c - y c-1 ,

[0113] In the above formula, (x c , y c ) is the planar coordinate of the constraint point p c (x c , y c ), and (x c-1 , y c-1 ) is the planar coordinate of the constraint point p c-1 (x c-1 , y c-1 ).

[0114] In this embodiment, the function expression for obtaining the final tensor field by summation in step S404 is:

[0115]

[0116] In the above formula, T(p) is the final tensor field, T c (p) is the basis tensor field T c (x c , y c ) of each constraint point p c (p), d is the attenuation coefficient, σ is the Gaussian bandwidth, p is a passable point on the robot motion plane, and p c is the constraint point. The global path will affect the field update process, and the degenerate points on the path will disappear, thus solving the problem of discontinuity or ambiguity. In addition, when an obstacle blocks the global path, the robot can advect along the particles of the field without performing the A* algorithm calculation again. In this embodiment, the attenuation constant d = 1. The Gaussian bandwidth σ can be used to control the influence range of the basic field, and its value is set to σ = 2.5ds.

[0117] To verify the method of this embodiment, the autonomous exploration method using the tensor field-driven hierarchical path planning of this embodiment and the autonomous exploration method without using the tensor field-driven hierarchical path planning are compared respectively. As Figure 8 shown, where (a) is the search path of the autonomous exploration method using the tensor field-driven hierarchical path planning of this embodiment, and (b) is the search path of the autonomous exploration method without using the tensor field-driven hierarchical path planning of this embodiment. The autonomous exploration method of the tensor field-driven hierarchical path planning of this embodiment provides a rough topological structure (boundary grouping set Q F ) matching the scene structure for the robot to navigate, and provides a fine-grained level scan (obtained by splitting and grouping in step S104). When the integrity of the scanned scene is similar, Figure 8 the path in (a) is 52.734m, but Figure 8The path in (b) is 105.282 m, which doubles the exploration efficiency. In addition, the autonomous exploration method using the tensor field-driven hierarchical path planning in this embodiment for autonomous exploration in different scenarios results in paths as Figure 9 shown. Tests reveal that the autonomous exploration method using the tensor field-driven hierarchical path planning in this embodiment achieves an impressive high-quality coverage of the scanning scene and provides reasonable planned paths for all these scenarios. Compared with other RGB-D camera-based methods, the autonomous exploration method using the tensor field-driven hierarchical path planning in this embodiment obtains a more complete scanning result than theirs, which benefits from our tensor field-driven hierarchical exploration.

[0118] In summary, the method in this embodiment utilizes boundary-based and tensor field-driven methods. To further unleash the global optimization ability of the tensor field, we adopt sparse degenerate points as anchor points to better utilize the local topological structure for global path optimization. Specifically, we find that degenerate points usually occur at the connection points of structures, such as entrances and porches. Since these anchor points are dynamically generated during the exploration process, we represent them as extended nodes in the traveling salesman problem (TSP) to measure the optimal path to form a rough topology, which can help the robot avoid traveling during long-distance cyclic exploration. In addition, a scanning strategy based on boundary grouping is proposed for fine-grained exploration. We dynamically group the boundaries around the anchor points and generate a local fine-grained exploration path for a complete scan. Therefore, a hierarchical exploration strategy is introduced to maintain the exploration efficiency while ensuring the scanning integrity. In addition, the tensor field of the method in this embodiment also supports a movement strategy to avoid collisions based on particle advection. A *The algorithm is the most commonly used method to find the shortest collision-free path to the next exploration target point. However, in a large environment, it is time-consuming. Previous LiDAR-scan-based work used techniques such as offline search strategies to find local motion paths, which were highly efficient. However, it highly depends on the complete FOV provided by LiDAR. In contrast, our framework can eliminate a large amount of time-consuming recalculation through a tensor field, which can be updated in real time and does not require FOV. Therefore, the depth scan movement strategy based on the tensor field in the method of this embodiment can be similar to or even better than the latest alternatives to LiDAR scan. The method of this embodiment uses the degenerate points in the tensor field as anchor points and generates a rough topology by solving the online TSP problem, which significantly improves the effectiveness of the tensor field-driven method in complex environments; the method of this embodiment is based on a scan strategy of frontier grouping for exploration at a fine-grained level. The dynamic grouping strategy relying on degenerate points can improve the exploration efficiency while ensuring the scanning integrity; the method of this embodiment uses dynamically extended degenerate points to provide a good reference for forming an optimized scene structure topology, enhancing the robot's global perception ability, thereby improving the efficiency. At the same time, an exploration strategy based on boundary grouping is introduced to perform fine-grained local scanning. These two strategies together constitute our hierarchical exploration method, which can better utilize the local topology for global path optimization, can eliminate a large amount of time-consuming recalculation of the path, is suitable for robot exploration in complex indoor scenes, and can greatly improve the efficiency of autonomous exploration while ensuring the exploration integrity.

[0119] In addition, this embodiment also provides an autonomous exploration device for tensor field-driven hierarchical path planning, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the autonomous exploration method for tensor field-driven hierarchical path planning. In addition, this embodiment also provides a computer-readable storage medium, and a computer program is stored in the computer-readable storage medium, and the computer program is used to be programmed or configured by the microprocessor to execute the autonomous exploration method for tensor field-driven hierarchical path planning.

[0120] Those skilled in the art should understand that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including instruction means that implement the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0121] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements should also be regarded as within the protection scope of the present invention.

Claims

1. An autonomous exploration method for tensor field-driven hierarchical path planning, characterized in that, Including: S101, constructing a scene tensor field according to the depth point cloud information sensed by the input RGB-D camera, and detecting degenerate points on the scene tensor field to obtain a set of degenerate points P; S102. Solve the online traveling salesman problem based on the degenerate point set P to obtain the degenerate point sequence Q corresponding to the optimal visiting order d ; S103, relying on the sequence of degradation points Q d Group the boundary set F of the unexplored space to obtain the boundary grouping set Q F ; S104, for each boundary grouping F in the boundary grouping set Q that spans a room F obtain the boundary outside the room by splitting it, filter the boundary outside the room and regard it as a new degenerate point and insert it into the degenerate point set P; k ​ S105, group the set Q according to the boundaries F Determine the next boundary point to be visited, calculate a global path T through the A* algorithm, and use the global path T to affect and update the tensor field; S106, advecting particles along the tensor field, calculating the linear velocity and angular velocity of the robot, and guiding the robot to reach the target.

2. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 1, characterized in that The set of degenerate points P obtained by detecting degenerate points on the scene tensor field in step S101 is obtained by combining the set of degenerate points obtained on the scene tensor field and the extended set of degenerate points obtained by extending the degenerate points covering the entire scene.

3. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 1, characterized in that Step S102 includes: S201. Find the degenerate point p0 closest to the robot position c in the degenerate point set r ; S202, based on the degradation point p0 closest to the robot position c r Solve the online traveling salesman problem shown in the following formula based on the set of degradation points P, and obtain the sequence of degradation points Q corresponding to the optimal visit order d : In the above formula, represents the online traveling salesman problem, and p0 is the degenerate point closest to the robot position c r The nearest degenerate point, is the first degenerate point in the degenerate point sequence Q d in, is the i-th degenerate point in the degenerate point sequence Q d in, is the (i + 1)-th degenerate point in the degenerate point sequence Q, dist is the path length between two degenerate points, and n is the number of degenerate points in the degenerate point sequence Q d The number of degenerate points, where the degenerate points d and and both come from the degenerate point set P.

4. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 1, wherein Step S103 includes: S301. For any boundary point f in the boundary set F m , calculate the Euclidean distance between the boundary point f m and any degenerate point in the degenerate point sequence Q d respectively, and determine the degenerate point with the closest Euclidean distance . S302, based on the grouping function of the following formula, group the boundary point f m into the degenerate point corresponding grouping F k ; In the above formula, is a grouping function, d f is a threshold value, f m is a boundary point, and the functional expression of the minimum distance d k (f m ) is: In the above formula, is the degenerate point with the closest Euclidean distance to the boundary point f m , k is an element of the set , and the set represents the index set in the degenerate point sequence Q d ; S303, add the boundary point f that satisfies the following formula m to the degenerate point set P: |f m -d k (f m )|>d f , In the above formula, d f is the threshold value.

5. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 1, characterized in that In step S104, the boundary obtained by splitting and located outside the room is {F k -S(F k )}, where S(F k ) is a splitting function, and there is: S(F k ) = {f k | f k ∈ F k , ν ≥ d g ∨ (μ < d p ∧ ν < d g )}, In the above formula, f k is a boundary point in the boundary group F k , ν is the position c where the robot is located r and the distance between the group F k , μ is the path length between the boundary point f k and the robot, d p and d g are thresholds, ∨ represents the OR operation, and ∧ represents the AND operation.

6. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 1, wherein In step S105, influencing and updating the tensor field with the global path T includes: S401, project the grid cells in the input raster map onto the floor, and perform the farthest point sampling at the centers of the projected boundary points for a specified distance d to select a set of planar constraint points, sort them in the path direction to obtain the constraint point set Pc, and project the global path T onto the constraint point set Pc; s ; project the global path T onto the constraint point set Pc; S402, for any two adjacent constraint points p in the set of constraint points Pc c-1 (x c-1 , y c-1 ) and p c (x c , y c ), calculate the regular elements (S c (x c , y c ) of the constraint point p c , T c ), where S c is the x-direction vector of the regular elements of the influence field, and T c is the y-direction vector of the regular elements of the influence field; S403, based on any constraint point p c (x c , y c ) of the regular elements (S c , T c ) calculates the basis tensor field T c (x c , y c ) of the constraint point p c (p); S404, Summing up the basis tensor fields T(p) of all constraint points through Gaussian radial basis functions to obtain the final tensor field. c (p) to obtain the final tensor field.

7. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 6, characterized in that, The constraint point p in step S402 c (x c ,y c ) of the general element (S c ,T c ) has the calculation function expression as follows: S c = x c - x c-1 , T c = y c -y c-1 , In the above formula, (x c , y c ) is the planar coordinates of the constraint point p c (x c , y c ), and (x c-1 , y c-1 ) is the planar coordinates of the constraint point p c-1 (x c-1 , y c-1 ).

8. The autonomous exploration method for tensor field-driven hierarchical path planning according to claim 6, characterized in that The function expression for summing to obtain the final tensor field in step S404 is: In the above formula, T(p) is the final tensor field, T c (p) is each constraint point p c (x c ,y c )'s basis tensor field T c (p), d is the attenuation coefficient, σ is the Gaussian bandwidth, p is the traversable point on the robot motion plane, p c is the constraint point.

9. An autonomous exploration device for tensor field-driven hierarchical path planning, comprising a microprocessor and a memory connected to each other, characterized in that, The microprocessor is programmed or configured to execute the autonomous exploration method for tensor field-driven hierarchical path planning according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program therein, characterized in that, The computer program is used to be programmed or configured by the microprocessor to execute the autonomous exploration method for tensor field-driven hierarchical path planning according to any one of claims 1 to 8.

Images