Autonomous navigation method and system, medium and navigation robot
The topology diagram is constructed through visual matching logic and prediction models, and the problem of insufficient autonomous navigation capabilities in open environments is solved, and precise navigation in complex environments is achieved.
Patent Information
- Application Number
- CN202411138624.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-19
AI Technical Summary
The existing technology lacks autonomous navigation capabilities in open environments, mainly due to the limited description of complex and changeable terrain characteristics of semantic labels, which greatly reduces the navigation capabilities in open environments.
Through visual matching logic, the passable distance prediction model and the relative pose prediction model are used to construct a topology map, match the positional relationship between the current scene and the impression scene, plan the shortest path and control the steering of the navigation robot to achieve autonomous navigation.
It significantly improves the navigation capabilities of autonomous robots in open environments, can achieve accurate long-distance navigation in complex and changeable environments, simplifies environmental representation and provides fine-grained posture control.
Smart Images

Figure CN119022931A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of intelligent navigation technology, and in particular to an autonomous navigation method, system, medium and navigation robot. Background Art
[0002] Traditional robot navigation is usually regarded as a geometric problem, and the path is planned by building a three-dimensional model of the environment. However, this method has limitations in the application of open environments, mainly because it needs to solve the collision problem in open environments. For example, the robot needs to know that grass can be crossed, but mud cannot be crossed.
[0003] In an open environment, navigation requires not only geometric information, but also the perception of the traversability of the terrain. To this end, existing technologies usually rely on map navigation with semantic information, using maps with semantic labels to help robots identify the characteristics of different terrains, and directly learn to understand the scene through training models, thereby achieving end-to-end navigation.
[0004] However, semantic labels have limited descriptions of complex and changing terrain characteristics, which will greatly reduce the autonomous navigation capabilities of existing technologies in open environments. Summary of the invention
[0005] In order to solve or partially solve the technical problem of low autonomous navigation capability of the prior art in an open environment, the present invention provides an autonomous navigation method, system, medium and navigation robot. Considering that the perception of terrain traversability mainly depends on visual observation, the "map" of the current scene and the impression scene can be matched to complete navigation. Therefore, the present invention simulates such visual matching logic, and matches the positional relationship between the target observation of the current position when it is reachable and observable, thereby improving the autonomous navigation capability.
[0006] In order to solve the above technical problems, the first aspect of the present invention discloses an autonomous navigation method, the method comprising:
[0007] Use a navigation robot to patrol and shoot in an open scene to obtain a set of patrol space images;
[0008] A topological map for navigation is constructed using a traversable distance prediction model and the patrol space image set; in the topological map, each node corresponds to a space image in the patrol space image set, and an edge between two nodes represents a traversable path between the two space images, wherein the traversable path carries a traversable probability, is predicted using the traversable distance prediction model, and is used for navigation;
[0009] According to the navigation requirements, a navigation starting point and a navigation end point are determined in the topological map, and the shortest path from the navigation starting point to the navigation end point is planned; each adjacent path node in the shortest path has a minimum cost value;
[0010] Predicting the position change between two adjacent path nodes in the shortest path using a relative posture prediction model;
[0011] Based on the change in posture, the navigation robot is controlled to turn at each path node in the shortest path to achieve autonomous navigation.
[0012] Optionally, the drivable distance prediction model is trained in the following manner:
[0013] In the open scene, a navigation robot is used to move and shoot on the same trajectory to obtain a spatial image data set, and a positive observation pair consisting of passable adjacent spatial image samples is obtained as a positive sample; wherein the positive observation pair is marked with a measured frame spacing and a measured pose; the measured frame spacing is used to characterize the spatial position of adjacent spatial images; the measured pose of the positive observation pair is the coordinates and heading angle of each spatial image in the adjacent spatial images relative to the position of the navigation robot at the previous moment in its own coordinate system;
[0014] Sampling negative observation pairs consisting of two inaccessible and non-adjacent frames of spatial images on different trajectories as negative samples; wherein the negative observation pairs are marked with inaccessible frame spacing and corresponding poses;
[0015] With the goal of minimizing the cross entropy loss between the measured frame spacing and the predicted frame spacing, the positive samples and the negative samples are used to jointly train a traversable distance basic model to obtain the traversable distance prediction model; wherein the traversable distance basic model includes an image encoder and a fully connected layer; wherein the fully connected layer is used to calculate the latent variable projection obtained by the image encoder processing the positive samples and the negative samples as the predicted frame spacing.
[0016] Optionally, in the training of the relative posture prediction model, the positive samples and the negative samples are used to jointly train a relative posture basic model with the goal of minimizing the cross entropy loss between the measured posture and the predicted posture to obtain the relative posture prediction model; wherein the relative posture basic model includes the image encoder and the fully connected layer; wherein the fully connected layer is used to calculate the latent variable projection obtained by the image encoder processing the positive samples and the negative samples as the predicted posture.
[0017] Optionally, in the training of the traversable distance prediction model and the relative posture prediction model, training is performed with the goal of simultaneously minimizing the cross entropy loss between the measured frame spacing and the predicted frame spacing, and the cross entropy loss between the measured pose and the predicted pose.
[0018] Optionally, the constructing a topological map for navigation using the traversable distance prediction model and the patrol space image set specifically includes:
[0019] Extracting adjacent frame spatial images from the patrol spatial image set in chronological order, and inputting them into the traversable distance prediction model for traversable prediction, to obtain the corresponding traversable path and its traversable probability;
[0020] The topological map M is constructed using the adjacent frame spatial images, the corresponding traversable paths and their traversable probabilities; wherein no intraversable paths are planned in the topological map.
[0021] Optionally, constructing the topological map by using the adjacent frame spatial images, corresponding traversable paths and traversable probabilities thereof specifically includes:
[0022] According to δ1 <T(o i , o j )<δ2, and select the passable paths used to construct the topological graph; wherein T(o i , o j ) refers to the passable probability predicted by the passable distance prediction model for the adjacent frame space images; δ1 and δ2 represent the passable lower limit threshold and the passable upper limit threshold, respectively. If it is less than the passable lower limit threshold, it means that there is a passable path between the adjacent frame space images and it is very easy to traverse; if it is greater than the passable upper limit threshold, it means that the adjacent frame space images are not passable.
[0023] Optionally, planning the shortest path from the navigation starting point to the navigation end point specifically includes:
[0024] Initializing the pass cost values of all nodes in the topological graph to the maximum cost value, and initializing the pass cost value of the navigation starting point to the minimum cost value;
[0025] The passability probability between adjacent nodes is converted into a passability cost according to the radiation order of the navigation starting point, each adjacent node corresponding to the navigation starting point, and each adjacent node corresponding to each adjacent node;
[0026] According to the radial order, the actual cost value between each node is calculated in combination with the pass cost, and is used to replace the corresponding maximum cost value until the navigation end point; wherein the passability probability and the pass cost are inversely related;
[0027] Between the navigation starting point and the navigation end point, a path node with the smallest cost value is selected to form the shortest path.
[0028] A second aspect of the present invention discloses an autonomous navigation system, comprising:
[0029] A shooting module is used to use the navigation robot to patrol and shoot in an open scene to obtain a patrol space image set;
[0030] A construction module, for constructing a topological map for navigation using a traversable distance prediction model and the patrol space image set; in the topological map, each node corresponds to a space image in the patrol space image set, and an edge between two nodes represents a traversable path between the two space images, and the traversable path carries a traversable probability, which is predicted by the traversable distance prediction model and used for navigation;
[0031] A determination module, used to determine a navigation start point and a navigation end point in the topological map according to navigation requirements, and plan a shortest path from the navigation start point to the navigation end point; each adjacent path node in the shortest path has a minimum cost value;
[0032] A prediction module, used to predict the posture change between two adjacent path nodes in the shortest path by using a relative posture prediction model;
[0033] The control module is used to control the steering of each path node of the navigation robot in the shortest path based on the posture change amount to achieve autonomous navigation.
[0034] According to a third aspect of the present invention, a computer-readable storage medium is disclosed, on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.
[0035] According to a fourth aspect of the present invention, a navigation robot is disclosed, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.
[0036] Through one or more technical solutions of the present invention, the present invention has the following beneficial effects or advantages:
[0037] The technical solution in the present invention utilizes the continuous scene understanding capabilities of the traversable distance prediction model and the relative posture prediction model to learn the traversable probabilities and relative positional relationships between the complex and changeable terrains in an open environment, and constructs a topological map M for navigation based on this, in which the traversable spatial positions, traversable paths, and their traversable probabilities are recorded. Therefore, in actual navigation deployment, the navigation starting point and the navigation end point and the shortest path between them can be located from the topological map M, and then the navigation robot can be controlled to turn at each path node in the shortest path to achieve autonomous navigation. It can be seen that the technical solution can achieve autonomous navigation in a complex and changeable environment, which not only effectively simplifies the environmental representation with a topological map and reduces the amount of calculation, but also provides fine-grained posture control, thereby achieving accurate long-distance navigation and significantly improving the navigation capability of the autonomous robot in an open environment.
[0038] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:
[0040] Figure 1 A diagram showing an implementation process of an autonomous navigation method according to an embodiment of the present invention is shown;
[0041] Figure 2 A schematic diagram of a trajectory generated by a navigation robot patrolling the edge of an impassable terrain according to an embodiment of the present invention is shown;
[0042] Figure 3 A module diagram of an autonomous navigation system according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0043] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.
[0044] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.
[0045] First, as Figure 1 As shown, an autonomous navigation method provided by an embodiment of the present invention at least includes the following steps:
[0046] S101, using a navigation robot to patrol and shoot in an open scene to obtain a patrol space image set.
[0047] Among them, in the open scene, there are complex and changeable terrains. For example, it includes mud, pools, walls and other impassable terrains, as well as grass, roads and other passable terrains. In the open scene, the remote control robot patrols and shoots in advance to obtain the patrol space image set {o t},o t Refers to each space image in the set. t}According to the chronological order, it is obtained through one or more inspections and photography.
[0048] S102, constructing a topological map used for navigation using a traversable distance prediction model and a patrol space image set.
[0049] In this embodiment, there are two prediction models: a traversable distance prediction model and a relative posture prediction model, both of which need to be trained.
[0050] The following first introduces the training process of the traversable distance prediction model and the relative posture prediction model.
[0051] Data collection: Various data sets are obtained by remotely controlling the navigation robot and recording the scenes passed by it in combination with the odometer. The navigation robot captures continuous RGB image stacks with a resolution of 1920×720 pixels. At the same time, the odometer records the distance between adjacent frames and the measured pose of the navigator robot when taking the spatial image. The measured pose is the position coordinates and heading angle of the navigation robot in its own coordinate system when taking the spatial image.
[0052] Furthermore, considering that there is traversable terrain in the open scene, in order to enable the traversable distance prediction model to perceive the traversable terrain, in the open scene, a navigation robot is used to move and shoot on the same trajectory to obtain a spatial image dataset, and a positive observation pair consisting of traversable adjacent spatial image samples is obtained as a positive sample. The trajectory is a traversable trajectory generated when the navigation robot patrols and passes through the traversable terrain in a navigation task. At this time, the navigation robot patrols and shoots in the traversable terrain, collects a large amount of spatial image data in chronological order, and constructs several positive observation pairs as positive samples.
[0053] In each positive observation pair, the measured frame spacing and the measured pose are marked.
[0054] The measured frame spacing of the observation pair is used to characterize the spatial position of adjacent spatial images. The measured frame spacing can be expressed as the time step d ij Indicates the number of time steps d ij Obtained by transformation of measured positions, but this representation does not constitute a limitation. ij = ji, j and i represent the spatial images respectively, the two frames of spatial images in the observation pair are adjacent, and the actual measured frame spacing is small. ij ∈(1,δ2), where δ2 represents the upper threshold of the passability, which is obtained by the measured frame spacing of each positive observation pair. The measured pose of the positive observation pair is the coordinates (x, y, z) and heading angle of each space image in the adjacent space image relative to the position of the navigation robot at the previous moment in its own coordinate system.
[0055] Taking into account that there are also impassable terrains in the open scene, in order to enable the traversable distance prediction model to perceive the impassable terrain, negative observation pairs consisting of two impassable and non-adjacent frames of spatial images are sampled on different trajectories as negative samples. The spatial position between the two frames of spatial images in the negative observation pair is impassable. Specifically, the impassable terrain will hinder the movement of the navigation robot, prompting the navigation robot to avoid the impassable terrain and move to produce a trajectory. The impassable terrain is then represented by negative samples composed of images in different trajectories. In the negative observation, the pair is marked with an impassable frame spacing and corresponding pose. It is worth noting that the impassable frame spacing of the negative observation pair can also be expressed in terms of the time step d. ij In order to distinguish positive samples from negative samples, the frame spacing of negative observation pairs is usually set to a larger value: d ij ∈[δ2, d max ]. For example, the frame spacing of the negative observation pair is calibrated according to the maximum similarity value d ij ≈d max .
[0056] by Figure 2As an example, it is a schematic diagram of the trajectory generated by the navigation robot patrolling the edge of an impassable terrain. Trajectory 1 and trajectory 2 are disclosed. Among them, in chronological order, two adjacent frames of spatial images are extracted from trajectory 1 as positive observation pairs, and the number of time steps marked is d ij The measured position of two adjacent frames of spatial images is converted, and track 2 is similar. In addition, one frame of image is sampled from track 1 and track 2 respectively to form a negative observation pair for data expansion, and marked d ij ≈d max .
[0057] Data preprocessing: The average inter-frame distance is calculated through the frame rate, and the average frame spacing normalization is used to standardize positive and negative samples for subsequent model processing.
[0058] Furthermore, positive and negative samples are used to perform supervised learning on the basic model of drivable distance, and a drivable distance prediction model is trained.
[0059] The architecture of the basic model of traversable distance mainly consists of an image encoder and a fully connected layer. The image encoder is used to process the latent variables obtained from positive samples and negative samples. Exemplarily, the image encoder uses a mamba encoder to extract rich feature representations from the input data, but the present application is not limited to the use of a mamba encoder, and other encoders may also be used. The fully connected layer is used to calculate the projection of the latent variables output by the image encoder as the predicted frame spacing. The predicted frame spacing may also be calculated using the time step d i ' j indicates, but does not constitute a limitation.
[0060] In the process of training the traversable distance prediction model, the goal is to minimize the cross entropy loss between the measured frame spacing and the predicted frame spacing, and the traversable distance basic model is trained with positive samples and negative samples to obtain the traversable distance prediction model. max =50, this example sets 50 class labels, namely: 1,...,50(d max ), the image encoder processes the positive and negative samples to obtain several 1024-dimensional latent variables, and uses the fully connected layer to process them to obtain several predicted frame spacings d i ' j , and discretized into 50 class labels. With the goal of minimizing the cross entropy loss between the measured frame spacing and the predicted frame spacing, the passable distance basic model is trained according to the gradient descent training method until a passable distance prediction model that meets the requirements is obtained.
[0061] In the passable distance prediction model, the predicted frame spacing is used to characterize the passability of the terrain. For example, in practical applications, when the predicted frame spacing between two frames of spatial images is less than δ2, it means that the two frames of spatial images are passable, otherwise they are not passable. In addition, the frame spacing predicted by the passable distance prediction model can be used as the passability probability between two frames of images, which will be introduced later and will not be repeated here.
[0062] The architecture of the relative posture basic model is similar to that of the traversable distance basic model, including: mainly composed of an image encoder and a fully connected layer. Among them, the image encoder is used to process the latent variables obtained by positive samples and negative samples. Exemplarily, the image encoder uses a mamba encoder to extract rich feature representations from the input data, but this application is not limited to the use of a mamba encoder, and other encoders can also be used. The fully connected layer is used to calculate the projection of the latent variables output by the image encoder into a predicted pose.
[0063] In the training of the relative posture prediction model, the goal is to minimize the cross entropy loss between the measured posture and the predicted posture, and the relative posture basic model is trained with positive and negative samples to obtain the relative posture prediction model. By training it with positive and negative samples, the model's discrimination ability, robustness and overall performance can be improved.
[0064] In an optional implementation, the traversable distance prediction model and the relative posture prediction model may be jointly trained. Specifically, in the training of the traversable distance prediction model and the relative posture prediction model, the cross entropy loss between the measured frame spacing and the predicted frame spacing, and the cross entropy loss between the measured pose and the predicted pose are simultaneously minimized to enhance the prediction accuracy between the two models.
[0065] The above is the training process of the two models in this embodiment. In actual use, the traversable distance prediction model is used to train the patrol space image set {o t} is processed to construct a topological map M used for navigation.
[0066] Specifically, in chronological order, from the patrol space image set {o t}, and input the passable distance prediction model to perform passable prediction, and obtain the corresponding passable path and its passable probability. The topology map M is constructed using the adjacent frame space images, the corresponding passable paths and their passable probabilities. Among them, if the passable probability is above the passable upper threshold, the corresponding adjacent frame space images are not passable, and the corresponding edge will not appear in the topology map M.
[0067] In the topological graph M, nodes and edges represent spatial images and traversable paths, respectively. Each node corresponds to a spatial image in the patrol spatial image set, and the edge between two nodes represents a traversable path between the two spatial images. The traversable path carries a traversable probability, which is the predicted frame spacing, which is predicted by the traversable distance prediction model and used for navigation.
[0068] In an optional implementation, in order to prevent the traversable paths between nodes in the topological graph M from becoming redundant and infeasible, the edges added to the topological graph M are thresholded to thin out the running trajectories between the spatial images.
[0069] Specifically, according to δ1 <T(o i , o j )<δ2, and the passable paths used to construct the topological graph M are screened out to make the running trajectories between the spatial images in the open scene sparse. i , o j ) refers to the passable probability predicted by the passable distance prediction model of the adjacent frame space images, which is equivalent to the predicted frame spacing d i ' j ; δ1 and δ2 represent the lower and upper passable thresholds respectively.
[0070] Specifically, if the value is less than the passable lower limit threshold, it means that the path between adjacent frame space images is passable and very easy to traverse, so it will not be added to the topology map M. If the value is above the passable upper limit threshold, it means that the path between adjacent frame space images is not passable and the frame spacing is too large, so it will not be added to the topology map M. For example, the predicted frame spacing d i ' j ≥δ2, indicating that the spatial position between the two frames of spatial images is inaccessible, and the predicted frame spacing d is ignored. i ' j , and the edge between the two frames of spatial images will not be added in the topological map M. The inaccessible paths are filtered out using the passable upper threshold δ2. Therefore, all the paths planned in the topological map M are passable paths, and no inaccessible paths will be planned in the topological map M.
[0071] The topology map M records all navigable paths in the open scene and is used for navigation in practical applications.
[0072] S103, according to the navigation requirements, determine the navigation starting point and the navigation end point in the topology map, and plan the shortest path from the navigation starting point to the navigation end point; each adjacent path node in the shortest path has a minimum cost value.
[0073] The navigation requirement can be controlled by the user or generated by the navigation robot according to the actual situation. The navigation start point and the navigation end point are any two nodes in the topological graph M.
[0074] Among them, the passable probability carried by the passable path in the topological graph is integrated into it, the passable cost value of each node is calculated, and the shortest path is planned based on it. The specific process is as follows:
[0075] The pass cost values of all nodes in the topology graph M are initialized to the maximum cost value, and the pass cost value of the navigation starting point is initialized to the minimum cost value.
[0076] Exemplarily, the travel cost values of all nodes in the topological graph M are initialized to infinity, and the travel cost value of the navigation starting point is initialized to zero.
[0077] Furthermore, the passability probability between adjacent nodes is converted into a passability cost based on the navigation starting point, the adjacent nodes corresponding to the navigation starting point, and the radiation order of the adjacent nodes corresponding to each adjacent node.
[0078] Extract the node with the smallest travel cost from the processing queue: the navigation starting point.
[0079] For each adjacent node in the navigation starting point, the passable probability between the navigation starting point and each adjacent node is converted into a passable cost, wherein the passable probability and the passable cost are inversely related.
[0080] For example, if the passability probability of the edge e between the navigation starting point and a certain adjacent node is pe, the passability probability pe is converted into the passability cost ce by using the inverse ce=1 / pe or the logarithm ce=-log(pe).
[0081] In this way, the travel costs between all adjacent nodes are calculated based on the navigation starting point, the adjacent nodes corresponding to the navigation starting point, and the radiation order of the adjacent nodes corresponding to each adjacent node.
[0082] Furthermore, according to the radial order, the actual cost between each node is calculated in combination with the travel cost to replace the corresponding maximum cost until the navigation end point.
[0083] Each node has one or more adjacent nodes, and each adjacent node calculates the actual cost value relative to the node and replaces the corresponding maximum cost value until the navigation end point is calculated. In this way, all adjacent nodes in the topology graph M have their own pass cost value.
[0084] Furthermore, between the navigation start point and the navigation end point, the path node with the smallest cost value is selected to form the shortest path.
[0085] In an optional embodiment, if the environment of the open scene changes, such as the appearance of a new obstacle, the topology map M is updated according to the change, and the nodes and edges affected by the change in the topology map M are updated in the above manner to avoid calculating the path from scratch and improve the efficiency of path planning.
[0086] S104, using the relative posture prediction model to predict the posture change between two adjacent path nodes in the shortest path.
[0087] After the shortest path is determined, two adjacent path nodes in the shortest path are sequentially input into the relative posture prediction model to predict the posture change between the two adjacent path nodes. The architecture and training process of the relative posture prediction model have been introduced above, so they will not be repeated here.
[0088] S105, controlling the steering of each path node of the navigation robot in the shortest path based on the posture change amount to achieve autonomous navigation.
[0089] Specifically, referring to the posture change amount, the PD controller is used to calculate the steering control parameters u(t) of the navigation robot, for example, the linear velocity and angular velocity of the navigation robot, and the motion parameters of the navigation robot, such as speed and direction, are adjusted according to the path execution error and the rate of change of the error, so as to perform turning actions toward each path node and drive the navigation robot to move toward the navigation end point.
[0090] In the technical solution of the present invention, based on the visual perception of the spatial position accessibility of the spatial image and the relative posture prediction, the PD controller is used to realize the navigation of each path node, thereby ensuring the accuracy and stability of autonomous navigation in a complex environment.
[0091] In the technical solution of the present invention, by combining the traversable distance prediction model and the relative posture prediction model, the spatial images of the navigation path are matched and analyzed under the conditions of reachability and observability, and the relative frame spacing and relative posture between the spatial images are obtained, thereby realizing an end-to-end control solution from image to control directly and realizing autonomous navigation. In addition, in the migration of different environments, only scene matching is needed to make predictions, and the portability is strong.
[0092] In a second aspect, based on the same inventive concept as the autonomous navigation method provided in the embodiment of the first aspect, the embodiment of the present invention further provides an autonomous navigation system, see Figure 3 ,include:
[0093] The shooting module 301 is used to use the navigation robot to patrol and shoot in an open scene to obtain a patrol space image set;
[0094] A construction module 302 is used to construct a topological map M for navigation using a traversable distance prediction model and the patrol space image set; in the topological map M, each node corresponds to a space image in the patrol space image set, and an edge between two nodes represents a traversable path between the two space images, and the traversable path carries a traversable probability, which is predicted by the traversable distance prediction model and used for navigation;
[0095] The determination module 303 is used to determine the navigation starting point and the navigation end point in the topological map according to the navigation requirements, and plan the shortest path from the navigation starting point to the navigation end point; each adjacent path node in the shortest path has a minimum cost value;
[0096] A prediction module 304 is used to predict the position change between two adjacent path nodes in the shortest path by using a relative posture prediction model;
[0097] The control module 305 is used to control the steering of each path node of the navigation robot in the shortest path based on the posture change amount to achieve autonomous navigation.
[0098] It should be noted that the autonomous navigation system provided by the embodiment of the present invention, wherein the specific manner in which each module performs operations has been described in detail in the method embodiment provided in the above-mentioned first aspect. The specific implementation process can refer to the method embodiment provided in the above-mentioned first aspect, and will not be elaborated in detail here.
[0099] In the third aspect, based on the same inventive concept as the autonomous navigation method provided in the embodiment of the first aspect mentioned above, the embodiment of the present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.
[0100] In the fourth aspect, based on the same inventive concept as the autonomous navigation method provided in the embodiment of the first aspect mentioned above, the embodiment of the present invention also discloses a navigation robot, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the methods described above when executing the program.
[0101] Through one or more embodiments of the present invention, the present invention has the following beneficial effects or advantages:
[0102] The technical solution in the present invention utilizes the continuous scene understanding capabilities of the traversable distance prediction model and the relative posture prediction model to learn the traversable probabilities and relative positional relationships between the complex and changeable terrains in an open environment, and constructs a topological map M for navigation based on this, in which the traversable spatial positions, traversable paths, and their traversable probabilities are recorded. Therefore, in actual navigation deployment, the navigation starting point and the navigation end point and the shortest path between them can be located from the topological map M, and then the navigation robot can be controlled to turn at each path node in the shortest path to achieve autonomous navigation. It can be seen that the technical solution can achieve autonomous navigation in a complex and changeable environment, which not only effectively simplifies the environmental representation with a topological map and reduces the amount of calculation, but also provides fine-grained posture control, thereby achieving accurate long-distance navigation and significantly improving the navigation capability of the autonomous robot in an open environment.
[0103] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.
[0104] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.
Claims
1. An autonomous navigation method, It is characterized in that The method comprises: Use a navigation robot to patrol and shoot in an open scene to obtain a set of patrol space images; A topological map for navigation is constructed using a traversable distance prediction model and the patrol space image set; in the topological map, each node corresponds to a space image in the patrol space image set, and an edge between two nodes represents a traversable path between the two space images, wherein the traversable path carries a traversable probability, is predicted using the traversable distance prediction model, and is used for navigation; According to the navigation requirements, a navigation starting point and a navigation end point are determined in the topological map, and the shortest path from the navigation starting point to the navigation end point is planned; each adjacent path node in the shortest path has a minimum cost value; Predicting the position change between two adjacent path nodes in the shortest path using a relative posture prediction model; Based on the change in posture, the navigation robot is controlled to turn at each path node in the shortest path to achieve autonomous navigation.
2. The method according to claim 1, It is characterized in that The drivable distance prediction model is trained in the following manner: In the open scene, a navigation robot is used to move and shoot on the same trajectory to obtain a spatial image data set, and a positive observation pair consisting of passable adjacent spatial image samples is obtained as a positive sample; wherein the positive observation pair is marked with a measured frame spacing and a measured pose; the measured frame spacing is used to characterize the spatial position of adjacent spatial images; the measured pose of the positive observation pair is the coordinates and heading angle of each spatial image in the adjacent spatial images relative to the position of the navigation robot at the previous moment in its own coordinate system; Sampling negative observation pairs consisting of two inaccessible and non-adjacent frames of spatial images on different trajectories as negative samples; wherein the negative observation pairs are marked with inaccessible frame spacing and corresponding poses; With the goal of minimizing the cross entropy loss between the measured frame spacing and the predicted frame spacing, the positive samples and the negative samples are used to jointly train a traversable distance basic model to obtain the traversable distance prediction model; wherein the traversable distance basic model includes an image encoder and a fully connected layer; wherein the fully connected layer is used to calculate the latent variable projection obtained by the image encoder processing the positive samples and the negative samples as the predicted frame spacing.
3. The method according to claim 2, It is characterized in that In the training of the relative posture prediction model, the positive samples and the negative samples are used to jointly train the relative posture basic model with the goal of minimizing the cross entropy loss between the measured posture and the predicted posture, so as to obtain the relative posture prediction model; wherein the relative posture basic model includes the image encoder and the fully connected layer; wherein the fully connected layer is used to calculate the latent variable projection obtained by the image encoder processing the positive samples and the negative samples as the predicted posture.
4. The method according to claim 3, It is characterized in that In the training of the traversable distance prediction model and the relative posture prediction model, the training is performed with the goal of simultaneously minimizing the cross entropy loss between the measured frame spacing and the predicted frame spacing, and the cross entropy loss between the measured posture and the predicted posture.
5. The method according to claim 1, It is characterized in that The use of the traversable distance prediction model and the patrol space image set to construct a topological map used for navigation specifically includes: Extracting adjacent frame spatial images from the patrol spatial image set in chronological order, and inputting them into the traversable distance prediction model for traversable prediction, to obtain the corresponding traversable path and its traversable probability; The topological map M is constructed using the adjacent frame spatial images, the corresponding traversable paths and their traversable probabilities; wherein no intraversable paths are planned in the topological map.
6. The method according to claim 5, It is characterized in that The constructing the topological map by using the adjacent frame spatial images, the corresponding traversable paths and their traversable probabilities specifically includes: According to δ 1 <T(o i , o j )<δ 2 The screening criteria are used to screen out the passable paths used to construct the topological graph; wherein T(o i , o j ) refers to the passable probability predicted by the passable distance prediction model for the adjacent frame spatial image; δ 1 , δ 2 They respectively represent the passable lower limit threshold and the passable upper limit threshold. A value less than the passable lower limit threshold indicates that there is a passable path between the adjacent frame space images and it is very easy to traverse. A value greater than the passable upper limit threshold indicates that the adjacent frame space images are not passable.
7. The method according to claim 1, It is characterized in that The planning of the shortest path from the navigation starting point to the navigation end point specifically includes: Initializing the pass cost values of all nodes in the topological graph to the maximum cost value, and initializing the pass cost value of the navigation starting point to the minimum cost value; The passability probability between adjacent nodes is converted into a passability cost according to the radiation order of the navigation starting point, each adjacent node corresponding to the navigation starting point, and each adjacent node corresponding to each adjacent node; According to the radial order, the actual cost value between each node is calculated in combination with the pass cost, and is used to replace the corresponding maximum cost value until the navigation end point; wherein the passability probability and the pass cost are inversely related; Between the navigation starting point and the navigation end point, a path node with the smallest cost value is selected to form the shortest path.
8. An autonomous navigation system, It is characterized in that include: A shooting module is used to use the navigation robot to patrol and shoot in an open scene to obtain a patrol space image set; A construction module, for constructing a topological map for navigation using a traversable distance prediction model and the patrol space image set; in the topological map, each node corresponds to a space image in the patrol space image set, and an edge between two nodes represents a traversable path between the two space images, and the traversable path carries a traversable probability, which is predicted by the traversable distance prediction model and used for navigation; A determination module, used to determine a navigation start point and a navigation end point in the topology map according to navigation requirements, and plan the shortest path from the navigation start point to the navigation end point; There is a minimum cost value between each adjacent path node in the shortest path; A prediction module, used to predict the posture change between two adjacent path nodes in the shortest path by using a relative posture prediction model; The control module is used to control the steering of each path node of the navigation robot in the shortest path based on the posture change amount to achieve autonomous navigation.
9. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.
10. A navigation robot comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.
Citation Information
Patent Citations
Sparse map for autonomous vehicle navigation
CN107438754A
Navigation method and system based on vision and laser radar fusion
CN110849366A
Adder circuit with binary complement conversion
CN113268219A
Visual navigation method based on weed removal
CN115560754A
Map modification using ground-truth measurements
US20130267260A1
Cited By
Unmanned aerial vehicle autonomous exploration method for large-scale unknown scene
CN119642818A
Autonomous navigation and obstacle avoidance system of fire-fighting robot
CN120949764A
