Robot repositioning method and device based on laser radar
Through the robot relocation method based on lidar, convolutional neural network and Re-Attention module are used to process point cloud data, which solves the problems of insufficient accuracy and poor generalization during relocation of neural networks, and achieves higher positioning accuracy and Environmental robustness.
Patent Information
- Application Number
- CN202411543383.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-31
- Publication Date
- 2025-02-07
AI Technical Summary
Neural networks have problems with insufficient accuracy and poor generalization when relocating robots, especially in spatial position perception.
Using the robot relocation method based on lidar, the initial point cloud data is obtained, and the trained convolutional neural network is used to convert it into point cloud data under the coordinate system of the target scene, and dynamic position encoding is performed, and input it to the Re-Attention module. and MLP networks to obtain the position of the lidar in the target scenario.
It improves the expression and generalization ability of the model, enhances the understanding and utilization of location information, reduces the complexity of the network for data processing, and improves the positioning accuracy and environmental robustness.
Smart Images
Figure CN119388422A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of robot positioning technology, and in particular to a robot repositioning method and device based on laser radar. Background Art
[0002] In recent years, neural networks have demonstrated powerful information processing capabilities and have achieved many successful explorations in the field of robotics, especially in computer vision such as image recognition, object detection, image segmentation, and face recognition, and natural language processing such as machine translation and natural language understanding. However, compared with traditional methods, neural networks still lack the performance in spatial position perception, mainly in terms of insufficient accuracy.
[0003] Robot localization and mapping is one of the most important technologies in mobile robot applications, because mobile robots need to obtain their own positions in space in real time during the execution of tasks. Traditional robot localization and mapping solutions have achieved good positioning and mapping accuracy, but their generalization ability and robustness to the environment are still insufficient. Summary of the invention
[0004] In view of the above problems, the present invention provides a robot repositioning method and device based on laser radar to solve the problems of insufficient generalization and precision in the repositioning of neural network robots.
[0005] According to a first aspect of the present invention, a robot relocation method based on laser radar is provided, comprising:
[0006] Obtain the initial point cloud data of the target scene collected at the current moment;
[0007] According to the initial point cloud data, based on the trained convolutional neural network, target point cloud data in the target scene coordinate system is obtained;
[0008] Encoding the target point cloud data to obtain a dynamic position code of the target point cloud data;
[0009] Input the dynamic position encoding of the target point cloud data into the trained Re-Attention module to output the target features;
[0010] The target features are input into the trained MLP network to obtain the position and posture of the laser radar in the target scene.
[0011] Optionally, obtaining target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data includes:
[0012] Determine the corresponding transformed point cloud data in the spherical coordinate system according to the initial point cloud data;
[0013] The converted point cloud data is input into the trained convolutional neural network to obtain the target point cloud data in the target scene coordinate system.
[0014] Optionally, the converted point cloud data is input into a trained convolutional neural network in the form of image data.
[0015] Optionally, the posture includes a translation vector and a rotation vector of the laser radar in the target scene coordinate system.
[0016] Optionally, the process of training a convolutional neural network includes:
[0017] Constructing a plurality of input point cloud data about the target scene, wherein the input point cloud data is data in a laser radar coordinate system;
[0018] For each input point cloud data, the input point cloud data is converted into output point cloud data in the target scene coordinate system through a homogeneous transformation matrix;
[0019] According to multiple sets of input point cloud data and output point cloud data, the initial convolutional neural network is trained and the parameters are adjusted to obtain a trained convolutional neural network.
[0020] Optionally, training the Re-Attention module includes:
[0021] For each target scene, train the convolutional neural network corresponding to the target scene;
[0022] Encoding the output point cloud data of the convolutional neural network to obtain dynamic position encoding of the output point cloud data, and using the dynamic position encoding of the output point cloud data as sample input data;
[0023] According to the sample input data, based on the high-precision combined inertial navigation system and system external parameters, the position and posture of the target scene in the coordinate system are obtained and used as sample output data;
[0024] According to the sample input data and the sample output data, the initial Re-Attention module is trained and the parameters are adjusted to obtain a trained Re-Attention module.
[0025] Optionally, the Re-Attention module also includes a regression head, and the regression head adopts an MLP multi-layer perceptron.
[0026] According to a second aspect of the present invention, there is provided a robot repositioning device based on laser radar, comprising:
[0027] An acquisition module is used to obtain the initial point cloud data of the target scene collected at the current moment;
[0028] A conversion module, used to obtain target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data;
[0029] An encoding module, used for encoding the target point cloud data to obtain a dynamic position encoding of the target point cloud data;
[0030] A feature module, used to input the dynamic position encoding of the target point cloud data into the trained Re-Attention module and output the target feature;
[0031] The pose module is used to input the target features into the trained MLP network to obtain the pose of the laser radar in the target scene.
[0032] According to a third aspect of the present invention, an electronic device is provided, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the aforementioned laser radar-based robot repositioning method when executing the computer program.
[0033] According to a fourth aspect of the present invention, a robot is provided, characterized in that the robot is equipped with a laser radar, and the robot uses the aforementioned laser radar-based robot repositioning method when repositioning in a target scene.
[0034] The above one or more technical solutions in the embodiments of this specification have at least the following technical effects:
[0035] The embodiments of this specification provide a laser radar-based robot repositioning method and device, which obtains the initial point cloud data of the target scene collected at the current moment; obtains the target point cloud data in the target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data; encodes the target point cloud data to obtain the dynamic position encoding of the target point cloud data; inputs the dynamic position encoding of the target point cloud data into the trained Re-Attention module to output the target features; inputs the target features into the trained MLP network to obtain the position and posture of the laser radar in the target scene. The present invention can enhance the model's understanding and utilization of position information through dynamic position encoding, thereby improving the model's expressive power; in addition, it can also help the model better understand and process new and unseen position information, thereby improving the model's generalization ability. On the other hand, the use of an unconstrained rotation description method reduces the complexity of the network's data processing, enabling the network to achieve better performance during training.
[0036] 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
[0037] 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. Also, throughout the accompanying drawings, the same reference figures are used to represent the same components. In the drawings:
[0038] Figure 1 A flowchart of a laser radar-based robot repositioning method in an embodiment of the present invention is shown.
[0039] Figure 2 A schematic diagram of a network model related to relocation in an embodiment of the present invention is shown.
[0040] Figure 3 A block diagram of a laser radar-based robot repositioning device in an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0041] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0042] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0043] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0044] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0045] In recent years, neural networks have demonstrated powerful information processing capabilities and have achieved many successful explorations in the field of robotics, especially in computer vision such as image recognition, object detection, image segmentation, and face recognition, and natural language processing such as machine translation and natural language understanding. However, compared with traditional methods, neural networks still lack the performance in spatial position perception, mainly in terms of insufficient accuracy.
[0046] Robot localization and mapping is one of the most important technologies in mobile robot applications, because mobile robots need to obtain their own position in space in real time during the execution of tasks. In some scenarios, such as tunnels or remote areas, the GPS signal is weak and the robot needs other ways to locate. In traditional solutions, most of them need to load and store the global map into the robot. In addition, traditional robot localization and mapping solutions have achieved good positioning and mapping accuracy, but the generalization ability and robustness of the environment are still insufficient.
[0047] Based on the above situation, we consider the following two aspects:
[0048] Neural networks have powerful information processing capabilities and can give full play to the advantages of data-driven, gaining good generalization capabilities and robustness to scenarios;
[0049] Using a neural network to regress the pose change can achieve centimeter-level error. If a neural network is used to regress the odometer, there will be cumulative errors. In order to ensure the accuracy of positioning and mapping, loop detection and backend optimization are required. The benefits of the neural network to the entire solution are limited. However, if a neural network is used to regress the absolute pose of the robot in a specified scene, it can ensure robustness to the environment while having good positioning accuracy.
[0050] Therefore, the present invention provides a robot relocalization method based on laser radar, which has guaranteed generalization ability and robustness to the environment and good positioning accuracy. Because the main calculations are deployed on hardware devices such as GPU, the burden on the CPU can be greatly reduced, which also provides a guarantee for the robot system to perform multiple tasks.
[0051] Specifically, combined with Figure 1 , Figure 2 As shown, the present invention provides a robot repositioning method based on laser radar, and the robot repositioning method based on laser radar includes steps 101 to 105:
[0052] Step 101: Acquire the initial point cloud data of the target scene collected at the current moment;
[0053] In this embodiment, the target scene refers to a known scene, which includes multiple sub-scenes, such as streets, forest trails, and other sub-scenes.
[0054] The robot is equipped with detection sensors, which can obtain data with depth information in the effective field of view.
[0055] The detection sensor may be a laser radar or a depth camera, etc. In this example, a laser radar is selected as the detection sensor for illustration.
[0056] The initial point cloud data refers to the point cloud data about the target scene collected by the lidar at the current moment.
[0057] Step 102: Obtain target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data;
[0058] In this embodiment, the initial point cloud data is recorded as k is the time index of the point cloud of the current frame, and M is the number of points in the point cloud.
[0059] According to the initial point cloud data, the corresponding transformed point cloud data transformed into the spherical coordinate system is determined, which is recorded as The converted point cloud data is input into the trained convolutional neural network in the form of image data to obtain the target point cloud data in the target scene coordinate system, denoted as P k ′. The spatial 3D coordinates of the converted point cloud data correspond to the three channels of the image. The purpose is to use mature image processing technology and reduce the amount of calculation and memory consumption. The specific conversion mapping formula is:
[0060]
[0061] Among them, (u,v) T It is the coordinate of the pixel point in the 2D image. The four channels of each pixel point are the 3D coordinate (x, y, z) of the point in space and the reflection intensity of the point. U represents the horizontal axis of the image, V represents the vertical axis of the image, the origin of the coordinate is located in the upper left corner of the image, and the depth d = (p x2 +p y2 ) 1 / 2 , f h and f vThey are the horizontal field of view and the vertical field of view. The vertical field of view f v =f vu +f vl , where f vu The part above the horizontal plane, f vl is the part below the horizontal line, δ h and δ v are the horizontal and vertical resolutions represented by pixels respectively. All of the above data can be obtained through the known parameters of the LiDAR.
[0062] Step 103: performing encoding processing on the target point cloud data to obtain dynamic position encoding of the target point cloud data;
[0063] In this embodiment, the expressiveness of the model is enhanced. Dynamic position coding can enhance the model's understanding and use of position information, thereby improving the model's expressiveness; improve the model's generalization ability. Dynamic position coding can help the model better understand and process new, unseen position information, thereby improving the model's generalization ability. The specific implementation method is to map the 3D point cloud data to a higher dimensional space, and the calculation formula is:
[0064] PE 3D (p) = [p, sin(2 0 πp),cos(2 0 πp),…,sin(2 m-1 πp),cos(2 m-1 πp)]
[0065] Among them, p belongs to a point in the target point cloud data, And p∈P k ′. m is the dimension of the embedding space. In order to enable the network to capture more location information, and considering the noise factor of the input data, m is taken as 10. PE 3D (p) is the result of dynamic position encoding of point p. In addition, PE 3D (p) Use 1x1 convolution to convert the number of channels to 128.
[0066] Step 104: input the dynamic position encoding of the target point cloud data into the trained Re-Attention module, and output the target features;
[0067] In this embodiment, the target feature is recorded as It should be noted that the Re-Attention module has multiple attention submodules, and the submodules are connected by residual connections. The reason for using residual connections is that as the network depth increases, the information flow will weaken. In experiments, it can be verified that adding residual connections not only makes the network more generalizable, but also speeds up the network convergence.
[0068] Step 105: Input the target features into the trained MLP network to obtain the position and posture of the laser radar in the target scene.
[0069] In this embodiment, MLP refers to multi-layer perceptron. Target features The fully connected two-layer MLP network regresses t and n respectively, and obtains the estimated result t infer and n infer , t infer and n infer It is the inferred position and posture of the laser radar in the target scene coordinate system, that is, the translation vector and rotation vector of the laser radar in the target scene coordinate system. It should be pointed out that the common neural network schemes for robot repositioning usually use quaternions to describe rotations. It is well known that quaternions have unit vector constraints, and processing constrained problems places high demands on neural networks. For this reason, the present invention adopts an unconstrained rotation description method, namely the rotation vector. Through experiments, it can be verified that the method adopted by the present invention can not only improve the generalization of the network, but also improve the accuracy of the network regression odometer.
[0070] So far, the process from inputting the initial point cloud data to outputting the position and posture of the laser radar in the target scene has been described. In order to facilitate understanding of the solution described in the present invention, the following supplementary explanation is required.
[0071] The neural network in the present invention consists of three parts: (1) a convolutional neural network that transforms the input point cloud from the lidar coordinate system to the target scene coordinate system; (2) a Transformer-based Re-Attention module; and (3) a regression head based on MLP that regresses the lidar's position in the target scene.
[0072] It should be noted that the convolutional neural network is trained separately from the other two neural networks, and is trained for the target scene to transform the point cloud data from the laser radar coordinate system to the target scene coordinate system; in addition, although it is trained for the target scene, it is combined with the preset Re-Attention module to enable the entire network to show good generalization ability. The purpose of this is to avoid the need to retrain the entire system when the scene changes, making the entire system deployment more efficient.
[0073] For the Re-Attention module, this module is trained on a large amount of different scene data. It can be generalized to new scenes by combining dynamic position encoding, attention mechanism and Re-Attention network design strategy. When the target scene changes, it is only necessary to retrain the convolutional neural network according to the new scene, while keeping the preset Re-Attention module unchanged. The Re-Attention module is trained simultaneously with the regression head MLP. The output of the preset convolutional neural network is dynamically encoded, the encoding result is used as input data, and the position of the lidar in the target scene coordinate system is used as output.
[0074] In addition, the present invention does not directly store the global map like the traditional method. Instead, the geometric information of the target scene is implicitly encoded by training the convolutional neural network, which means that when the target scene changes, the convolutional neural network needs to be retrained to accurately transform the initial point cloud data from the lidar coordinate system to the target scene coordinate system.
[0075] In one embodiment, the process of training a convolutional neural network includes:
[0076] Constructing a plurality of input point cloud data about the target scene, wherein the input point cloud data is data in a laser radar coordinate system;
[0077] For each input point cloud data, the input point cloud data is converted into output point cloud data in the target scene coordinate system through a homogeneous transformation matrix;
[0078] According to multiple sets of input point cloud data and output point cloud data, the initial convolutional neural network is trained and the parameters are adjusted to obtain a trained convolutional neural network.
[0079] Specifically, you need to prepare the input point cloud data and output point cloud data first The data comes from the target scene and is all point cloud data. The input point cloud data is in the laser radar coordinate system, and it is converted to the target scene coordinate system through the homogeneous transformation matrix T as the output point cloud data, that is, the label data. Among them, T is usually obtained by traditional methods with high accuracy or combined with high-precision combined inertial navigation system and system external parameters.
[0080] The input point cloud data is then transformed into a spherical coordinate system to obtain the transformed input point cloud data, which is then input into the initial convolutional neural network in the form of image data.
[0081] The transformed input point cloud data is input into the initial convolutional neural network in the form of image data, and the input point cloud data in the target scene coordinate system is output.
[0082] About the loss function calculation. The specific calculation formula is:
[0083]
[0084] Among them, p i ∈P k ,i=1,…,M k ; M k is the number of point clouds.
[0085] Then update the network weights. Use a gradient descent algorithm (such as Adam or SGD) to optimize the loss function and update the network weights and biases.
[0086] Finally, iterate the training. Repeat the training process until the network's performance on the test set no longer improves significantly or reaches the predetermined number of iterations.
[0087] In one embodiment, training the Re-Attention module includes:
[0088] For each target scene, train the convolutional neural network corresponding to the target scene;
[0089] Encoding the output point cloud data of the convolutional neural network to obtain dynamic position encoding of the output point cloud data, and using the dynamic position encoding of the output point cloud data as sample input data;
[0090] According to the sample input data, based on the high-precision combined inertial navigation system and system external parameters, the position and posture of the target scene in the coordinate system are obtained and used as sample output data;
[0091] According to the sample input data and the sample output data, the initial Re-Attention module is trained and the parameters are adjusted to obtain a trained Re-Attention module.
[0092] Specifically, prepare sample input data and sample output data first:
[0093]
[0094] The training data needs to contain as many target scene types and target scene numbers as possible so that the Re-Attention module can obtain a good generalization effect. For the input sample data, it is necessary to train a convolutional neural network specifically for each specific target scene, and then perform dynamic position encoding on the output of this convolutional neural network, and record the encoding result as:
[0095]
[0096] As input sample data, the LiDAR pose in the specific target scene coordinate system As output sample data. Usually, the LiDAR position and posture are obtained by traditional methods with high accuracy or by combining high-precision combined inertial navigation system and system external parameters.
[0097] Will Input to the Re-Attention module, output the position of the lidar in the target scene coordinate system And the intermediate supervision value
[0098] Regarding the calculation of loss function, according to Calculate the loss and get In order to improve the overall performance of the solution, the present invention adds supervision in the regression intermediate layer. Combined with 2, the supervision results and Used to calculate auxiliary losses and The calculation formula can be substituted into the following formula:
[0099]
[0100] Among them, λ is the coefficient of the loss term, here λ = 100.0. Overall loss as follows:
[0101]
[0102] in, For the overall loss, is the inference loss value, and is the intermediate supervision loss value.
[0103] Then update the network weights. Use a gradient descent algorithm (such as Adam or SGD) to optimize the loss function and update the network weights and biases.
[0104] Then iterate the training. Repeat the training process until the network's performance on the test set no longer improves significantly or reaches the predetermined number of iterations.
[0105] It should be noted that the Re-Attention module also includes a regression head, and the regression head adopts an MLP multi-layer perceptron.
[0106] In summary, the laser radar-based robot repositioning method provided by the embodiments of this specification has stronger generalization and higher accuracy. As mentioned in the scheme description, dynamic position coding can enhance the model's understanding and utilization of position information, thereby improving the model's expressiveness; in addition, dynamic position coding can help the model better understand and process new and unseen position information, thereby improving the model's generalization ability. In addition, the heavy attention mechanism with residual connection, its residual connection can enhance the network's sensitivity to information flow, so that the effective information in the information flow can be more efficiently summarized after passing through the multi-layer network, thereby reducing the network's dependence on data volume and data type. An unconstrained rotation description method is also used. Obviously, this method reduces the complexity of the network's data processing, allowing the network to obtain better performance during training.
[0107] The use of an unconstrained rotation description method reduces the complexity of the network's data processing, enabling the network to achieve better performance during training. On the other hand, the use of auxiliary loss introduces supervision in the regression intermediate layer, improving the overall performance of the solution.
[0108] Based on the same inventive concept, Figure 3 As shown, an embodiment of the present invention further provides a robot repositioning device based on laser radar, comprising:
[0109] An acquisition module is used to obtain the initial point cloud data of the target scene collected at the current moment;
[0110] A conversion module, used to obtain target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data;
[0111] An encoding module, used for encoding the target point cloud data to obtain a dynamic position encoding of the target point cloud data;
[0112] A feature module, used to input the dynamic position encoding of the target point cloud data into the trained Re-Attention module and output the target feature;
[0113] The pose module is used to input the target features into the trained MLP network to obtain the pose of the laser radar in the target scene.
[0114] Optionally, obtaining target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data includes:
[0115] Determine the corresponding transformed point cloud data in the spherical coordinate system according to the initial point cloud data;
[0116] The converted point cloud data is input into the trained convolutional neural network to obtain the target point cloud data in the target scene coordinate system.
[0117] Optionally, the converted point cloud data is input into a trained convolutional neural network in the form of image data.
[0118] Optionally, the posture includes a translation vector and a rotation vector of the laser radar in the target scene coordinate system.
[0119] Optionally, the process of training a convolutional neural network includes:
[0120] Constructing a plurality of input point cloud data about the target scene, wherein the input point cloud data is data in a laser radar coordinate system;
[0121] For each input point cloud data, the input point cloud data is converted into output point cloud data in the target scene coordinate system through a homogeneous transformation matrix;
[0122] According to multiple sets of input point cloud data and output point cloud data, the initial convolutional neural network is trained and the parameters are adjusted to obtain a trained convolutional neural network.
[0123] Optionally, training the Re-Attention module includes:
[0124] For each target scene, train the convolutional neural network corresponding to the target scene;
[0125] Encoding the output point cloud data of the convolutional neural network to obtain dynamic position encoding of the output point cloud data, and using the dynamic position encoding of the output point cloud data as sample input data;
[0126] According to the sample input data, based on the high-precision combined inertial navigation system and system external parameters, the position and posture of the target scene in the coordinate system are obtained and used as sample output data;
[0127] According to the sample input data and the sample output data, the initial Re-Attention module is trained and the parameters are adjusted to obtain a trained Re-Attention module.
[0128] Optionally, the Re-Attention module also includes a regression head, and the regression head adopts an MLP multi-layer perceptron.
[0129] In summary, the embodiment of this specification provides a robot repositioning device based on laser radar, which obtains the initial point cloud data of the target scene collected at the current moment; obtains the target point cloud data in the target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data; encodes the target point cloud data to obtain the dynamic position encoding of the target point cloud data; inputs the dynamic position encoding of the target point cloud data into the trained Re-Attention module to output the target features; inputs the target features into the trained MLP network to obtain the position and posture of the laser radar in the target scene. The present invention can enhance the model's understanding and utilization of position information through dynamic position encoding, thereby improving the model's expressive power; in addition, it can also help the model better understand and process new and unseen position information, thereby improving the model's generalization ability. On the other hand, the use of an unconstrained rotation description method reduces the complexity of the network's data processing, enabling the network to obtain better performance during training.
[0130] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the laser radar-based robot repositioning device described above can refer to the corresponding process in the aforementioned method, and will not be elaborated here.
[0131] Based on the same inventive concept, this embodiment provides an electronic device, including a laser radar-based robot repositioning device, a memory, a processor and a communication unit. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor and the memory communicate through a bus, the processor executes the machine-readable instructions, and executes the laser radar-based robot repositioning method.
[0132] The memory, processor and communication unit are electrically connected to each other directly or indirectly to achieve signal transmission or interaction. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines. The laser radar-based robot repositioning device includes at least one software function module that can be stored in the memory in the form of software or firmware. The processor is used to execute the executable module stored in the memory (for example, the software function module or computer program included in the laser radar-based robot repositioning device).
[0133] Among them, the memory can be, but is not limited to, random access memory (Random Access Memory, RAM), read only memory (Read Only Memory, ROM), programmable read-only memory (Programmable Read-Only Memory, PROM), erasable programmable read-only memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable read-only memory (Electric Erasable Programmable Read-Only Memory, EEPROM), etc.
[0134] In some embodiments, the processor is used to perform one or more functions described in the present embodiment. In some embodiments, the processor may include one or more processing cores (e.g., a single-core processor (S) or a multi-core processor (S)). By way of example only, the processor may include a central processing unit (CPU), an application specific integrated circuit (ASIC), an application specific instruction set processor (ASIP), a graphics processing unit (GPU), a physical processing unit (PPU), a digital signal processor (DSP), a field programmable gate array (FPGA), a programmable logic device (PLD), a controller, a microcontroller unit, a reduced instruction set computer (RISC) or a microprocessor, or any combination thereof.
[0135] For ease of explanation, only one processor is described in the electronic device. However, it should be noted that the electronic device in this embodiment may also include multiple processors, so the steps performed by one processor described in this embodiment may also be performed jointly or individually by multiple processors. For example, if the processor of the server performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually in one processor. For example, the processor performs step A and the second processor performs step B, or the processor and the second processor perform steps A and B together.
[0136] In this embodiment, the memory is used to store the program, and the processor is used to execute the program after receiving the execution instruction. The process definition method disclosed in any implementation of this embodiment can be applied to the processor, or implemented by the processor.
[0137] The communication unit is used to establish a communication connection between the electronic device and other devices through a network, and to send and receive data through the network.
[0138] In some embodiments, the network may be any type of wired or wireless network, or a combination thereof. By way of example only, the network may include a wired network, a wireless network, a fiber optic network, a telecommunications network, an intranet, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a wide area network (WAN), a public switched telephone network (PSTN), a Bluetooth network, a ZigBee network, or a near field communication (NFC) network, or any combination thereof.
[0139] In this embodiment, the electronic device may be, but is not limited to, a laptop computer, an ultra-mobile personal computer (UMPC), a netbook, a personal digital assistant (PDA), and the like. This embodiment does not impose any restrictions on the specific type of the electronic device.
[0140] Based on the above, this embodiment provides a robot equipped with a laser radar, and the robot uses the aforementioned laser radar-based robot repositioning method when repositioning in a target scene.
[0141] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the robot described above can refer to the corresponding process in the aforementioned method, and will not be elaborated herein.
[0142] The above are only various embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A robot relocation method based on laser radar, characterized in that: include: Obtain the initial point cloud data of the target scene collected at the current moment; According to the initial point cloud data, based on the trained convolutional neural network, target point cloud data in the target scene coordinate system is obtained; Encoding the target point cloud data to obtain a dynamic position code of the target point cloud data; Input the dynamic position encoding of the target point cloud data into the trained Re-Attention module to output the target features; The target features are input into the trained MLP network to obtain the position and posture of the laser radar in the target scene.
2. The method according to claim 1, characterized in that The step of obtaining target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data includes: Determine the corresponding transformed point cloud data in the spherical coordinate system according to the initial point cloud data; The converted point cloud data is input into the trained convolutional neural network to obtain the target point cloud data in the target scene coordinate system.
3. The method according to claim 2, characterized in that The converted point cloud data is input into the trained convolutional neural network in the form of image data.
4. The method according to claim 1, characterized in that: The posture includes the translation vector and rotation vector of the laser radar in the target scene coordinate system.
5. The method according to claim 1, characterized in that The process of training a convolutional neural network involves: Constructing a plurality of input point cloud data about the target scene, wherein the input point cloud data is data in a laser radar coordinate system; For each input point cloud data, the input point cloud data is converted into output point cloud data in the target scene coordinate system through a homogeneous transformation matrix; According to multiple sets of input point cloud data and output point cloud data, the initial convolutional neural network is trained and the parameters are adjusted to obtain a trained convolutional neural network.
6. The method according to claim 1, characterized in that Training the Re-Attention module includes: For each target scene, train the convolutional neural network corresponding to the target scene; Encoding the output point cloud data of the convolutional neural network to obtain dynamic position encoding of the output point cloud data, and using the dynamic position encoding of the output point cloud data as sample input data; According to the sample input data, based on the high-precision combined inertial navigation system and system external parameters, the position and posture of the target scene in the coordinate system are obtained and used as sample output data; According to the sample input data and the sample output data, the initial Re-Attention module is trained and the parameters are adjusted to obtain a trained Re-Attention module.
7. The method according to claim 1, characterized in that The Re-Attention module also includes a regression head, and the regression head adopts an MLP multi-layer perceptron.
8. A robot repositioning device based on laser radar, characterized in that: include: An acquisition module is used to obtain the initial point cloud data of the target scene collected at the current moment; A conversion module, used to obtain target point cloud data in a target scene coordinate system based on the trained convolutional neural network according to the initial point cloud data; An encoding module, used for encoding the target point cloud data to obtain a dynamic position encoding of the target point cloud data; A feature module, used to input the dynamic position encoding of the target point cloud data into the trained Re-Attention module and output the target feature; The pose module is used to input the target features into the trained MLP network to obtain the pose of the laser radar in the target scene.
9. An electronic device, characterized in that: The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the laser radar-based robot repositioning method described in any one of claims 1 to 7 is implemented.
10. A robot, characterized in that: The robot is equipped with a laser radar, and when the robot is repositioned in a target scene, the laser radar-based robot repositioning method described in any one of claims 1 to 7 is used.