A mobile robot active identification method and device

Abstract

This invention proposes a mobile robot active recognition method and apparatus, belonging to the field of target recognition technology. The method includes: having the mobile robot acquire an RGB image and a depth image of its current position and orientation within a workspace containing a pre-defined target object; inputting the RGB image and depth image into a pre-defined robot active recognition network, which outputs the mobile robot's action at the next moment, including the direction and distance of the action; the mobile robot moving accordingly according to the network's output until it reaches the position closest to the pre-defined target object, at which point recognition is complete. This invention, based on reinforcement learning technology and combining depth and target object information, predicts the optimal field of view using an active vision algorithm. This overcomes the problem of insufficient target object information in the camera's field of view and significantly improves recognition reliability caused by target object occlusion or excessive distance.

Description

Technical Field

[0001] This invention belongs to the field of target recognition technology, and specifically proposes a method and device for active recognition of mobile robots. Background Technology

[0002] In the process of human environmental perception, when objects are too far away or obstructed, the most intuitive method is to change one's position by moving one's feet. This alters the perspective, allowing for the perception of more information about the object and improving the reliability of object recognition. Active vision technology mimics this human strategy when encountering uncertain objects. It uses information theory to analyze the optimal field of view and adjust the camera's pose. This technology is widely used in various fields, including path planning, real-time localization and mapping (SLAM), face tracking, vehicle tracking, and human interaction.

[0003] Active vision algorithms are generally categorized into several types: active camera systems, master-slave camera systems, and actively controlled camera systems. However, these traditional active vision algorithms employ information theory analysis to evaluate whether the input image's field of view is optimal, i.e., whether it can fully acquire information about the target object. These methods struggle to meet the challenges of modern mobile robots recognizing targets in situations such as occlusion, distant targets, and limited field of view, leading to low reliability of the recognition results, inconsistencies in time efficiency, and the inability to predict optimal positions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose an active recognition method and device for mobile robots. Based on reinforcement learning technology, this invention combines depth and target object information, and predicts the optimal field of view through an active vision algorithm. This overcomes the problem of insufficient target object information in the camera's field of view and significantly improves the low recognition reliability caused by target object occlusion or excessive distance.

[0005] A first aspect of this invention provides a method for active identification of a mobile robot, comprising:

[0006] The mobile robot acquires an RGB image and a depth image of its current position and orientation within a workspace containing a pre-defined target object.

[0007] The RGB image and depth image are input into a preset robot active recognition network, and the network outputs the action of the mobile robot at the next moment, including: action direction and action distance;

[0008] The mobile robot moves accordingly according to the output of the network until it moves to the position closest to the preset target object, at which point the identification is complete.

[0009] In a specific embodiment of the present invention, the method further includes: the initial position of the mobile robot is any intersection point of a chessboard grid obtained by dividing the workspace plane according to a set interval; the initial orientation of the mobile robot is any orientation from a preset orientation set; the action direction is an action direction selected from a preset action direction set; the action distance is the action distance corresponding to the action direction selected from a preset action distance set; and the action direction and the action distance ensure that the mobile robot remains at the intersection point of the chessboard grid after movement.

[0010] In one specific embodiment of the present invention, before inputting the RGB image and depth image into a preset robot active recognition network, the method further includes:

[0011] Train the robot's active recognition network;

[0012] Training the robot's active recognition network includes:

[0013] A training set is obtained, wherein each sample in the training set is obtained by collecting a depth map and an RGB map from the same location and orientation in the workspace by a mobile robot.

[0014] A robot recognition network is constructed, comprising a feature extraction subnetwork and a reinforcement learning subnetwork connected sequentially. The input to the feature extraction subnetwork is the depth map and RGB image collected by the mobile robot at the same location and orientation, and the output is the target object parameters and depth feature vector at that location and orientation. The input to the reinforcement learning subnetwork is the target object parameters and depth feature vector output by the feature extraction subnetwork, and the output is the action of the mobile robot at the next moment, including: action direction and action distance.

[0015] The feature extraction subnetwork is first trained using the training set, and then the reinforcement learning subnetwork is trained based on the output of the trained feature extraction subnetwork to obtain the trained robot active recognition network.

[0016] In one specific embodiment of the present invention, obtaining the training set includes:

[0017] A target object that is not completely obscured is placed in the workspace;

[0018] Set the orientation set, motion direction set, and motion distance set;

[0019] The workspace plane is divided into rectangular checkerboard patterns at predetermined intervals. At each intersection of the checkerboard patterns, the mobile robot collects the corresponding RGB image and depth image according to each orientation set. If the target object is present in the RGB image collected by the mobile robot, the type of the target object is recorded, and the bounding box and image difficulty of the target object are labeled. The bounding box of the target object is represented by x. min x max y min y max Indicates that x min x is the minimum value of the row pixels of the figure frame in the RGB image. max The maximum value of the row pixels of the figure frame in the RGB image, y min The minimum value of the column pixels of the figure frame in the RGB image, y max This represents the maximum value of the column pixels of the figure frame in the RGB image.

[0020] In one specific embodiment of the present invention, the feature extraction sub-network adopts a Mask R-CNN network with pre-trained weights; the feature extraction sub-network is used to obtain the corresponding feature map of the input depth map, and convert the feature map into the corresponding depth feature vector and target object parameters;

[0021] The target object parameters include the target object's center coordinates, width, height, and depth (x, y, y). 目标 y 目标 The parameters w, h, d) represent the relationship between the center coordinates, width, and height of the target object and the parameters of the target object's graphic frame:

[0022]

[0023]

[0024] w = x max -x min

[0025] h = y max -y min

[0026] Where, x 目标 y 目标 These are the x and y coordinates of the target object's center point in the RGB image, respectively, and the depth d is the target object's center coordinate (x, y) in the depth image. 目标 y 目标 The corresponding depth value.

[0027] In one specific embodiment of the present invention, the reinforcement learning sub-network employs a two-layer deep Q-network; the Q-value of the reinforcement learning sub-network is represented as follows:

[0028] Q(s,a,θ)=V d (s,a d ,θ f ,θ d )+V r (s,a r ,θ f ,θ r )

[0029] Where s represents the state; a represents the action, including: the direction of the action a d and action distance a r θ f For general decision parameters, θ d With θ r These are the direction decision parameters and the distance decision parameters, respectively, V d V r These represent the motion direction set and the motion distance set, respectively.

[0030] In a specific embodiment of the present invention, the loss function in the reinforcement learning sub-network is:

[0031]

[0032] Where θ represents the parameters of the reinforcement learning subnetwork, γ is the loss factor, and s t Let a be the state value at time t. t Let r be the action value at time t, and Q(s,a,θ) be the action value function with state s, action a, and parameter θ. t Let be the reward value function at time t.

[0033] A second aspect of the present invention provides a mobile robot active recognition device, comprising:

[0034] The image acquisition module is used to enable the mobile robot to acquire RGB and depth images of its current position and orientation in a workspace with a pre-set target object.

[0035] The motion generation module is used to input the RGB image and depth image into a preset robot active recognition network, and the network outputs the motion of the mobile robot at the next moment, including: motion direction and motion distance;

[0036] The execution module is used to make the mobile robot move accordingly according to the output of the network until the mobile robot moves to the position closest to the preset target object, and the recognition is completed.

[0037] A third aspect of the present invention provides an electronic device comprising:

[0038] At least one processor; and a memory communicatively connected to said at least one processor;

[0039] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to perform the above-described mobile robot active recognition method.

[0040] A fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute the above-described mobile robot active recognition method.

[0041] The features and beneficial effects of this invention are as follows:

[0042] This invention utilizes an active vision algorithm based on reinforcement learning that combines depth and RGB information. It takes depth and target information as the state input of the environment and adopts a dual DQN architecture to make more accurate optimal field-of-view predictions for mobile robots.

[0043] This invention can be used to assist mobile robots in detecting targets by overcoming the problem of insufficient target information through the pose of the mobile sensing system. It has a significant effect on improving the low recognition reliability caused by target occlusion or the target being too far away. Attached Figure Description

[0044] Figure 1 This is an overall flowchart of a mobile robot active recognition method according to an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram of the workspace in a specific embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the structure of a robot active recognition network in a specific embodiment of the present invention. Detailed Implementation

[0047] This invention proposes a method and apparatus for active recognition of mobile robots. The invention is further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not limited to the scope of the invention.

[0048] A first aspect of this invention provides a method for active identification of a mobile robot, comprising:

[0049] The mobile robot acquires an RGB image and a depth image of its current position and orientation within a workspace containing a pre-defined target object.

[0050] The RGB image and depth image are input into a preset robot active recognition network, and the network outputs the action of the mobile robot at the next moment, including: action direction and action distance;

[0051] The mobile robot moves accordingly according to the output of the network until it moves to the position closest to the preset target object, at which point the identification is complete.

[0052] In a specific embodiment of the present invention, the overall process of the mobile robot active recognition method is as follows: Figure 1 As shown, it includes the following steps:

[0053] 1) The training phase includes:

[0054] 1-1) Construct the training set;

[0055] 1-1-1) Set up a workspace for the mobile robot (this embodiment has no special requirements for the workspace, and it can be indoors or outdoors), and set up the target object in the workspace.

[0056] In this embodiment, the workspace may also include obstacles, and there may be one or more obstacles and one or more target objects. Starting from any point in the workspace, a path can be found to reach the target object, meaning that the target object will not be completely blocked by obstacles.

[0057] In one specific embodiment of the present invention, a mobile vehicle (i.e., a mobile robot) is used as an experimental platform to search for a specific target in a real-world scenario with obstacles. The workspace is as follows: Figure 2 As shown, an ellipsoid is used as the target object, and two obstacles are placed in the camera's field of view to interfere with the recognition of the target object. The workspace is 10m*10m in size, and the obstacles are 1m*1m square blocks.

[0058] 1-1-2) Divide the workspace plane into rectangular checkerboard patterns according to the set interval d, with each small square having a side length of d. Select any vertex of the rectangular checkerboard as the origin. The two perpendicular sides of the rectangle emanating from the origin are the x-axis and y-axis, respectively. The actual distances from the intersection of any coordinate (x, y) on the checkerboard to the x-axis and y-axis of the checkerboard are denoted as xd and yd, respectively. Define the orientation w on the checkerboard as the angle between the mobile robot's orientation and the x-axis. In this embodiment, this angle is an integer multiple of 30 degrees. Therefore, the mobile robot has a set of 12 orientations.

[0059] During data collection, a human moves the robot to collect data at each intersection of the chessboard grid, taking in data every 30 degrees from the robot's perspective. The robot collects 12 sets of data at each intersection, each set containing the RGB and depth maps acquired by the robot in that direction at that intersection. If an obstacle or target object is present at any intersection, that intersection is skipped. When a target object appears in the collected RGB map, the type of the target object is recorded, and its bounding box and image difficulty are labeled. The bounding box of the target object is represented by x. min x max y min y max It means that x min x is the minimum value of the row pixels of the figure frame in the RGB image. max The maximum value of the row pixels of the figure frame in the RGB image, y min The minimum value of the column pixels of the figure frame in the RGB image, y max This represents the maximum value of the column pixels of the shape frame in the RGB image. In this embodiment, the RGB image and the depth image are the same size, both 512*512; the image difficulty is divided into six levels based on the pixel area of ​​the shape frame, as shown in Table 1:

[0060] Table 1. Image difficulty classification table in a specific embodiment of the present invention.

[0061]

[0062] Here, "box" represents the pixel area of ​​the target object's bounding box. The bounding box area and the corresponding difficulty can be adjusted according to different tasks.

[0063] In one specific embodiment of the present invention, d = 0.05m is selected and the corresponding data is collected and labeled to form a training set.

[0064] It should be noted that this embodiment only requires one workspace to acquire the training set, and the workspace remains unchanged once set. If the robot does not collect the target object at its current location and orientation, it still constitutes a training sample, in which the parameters of the target object are all 0.

[0065] 1-1-3) Define the state S of the mobile robot in the workspace, which includes the robot's current position, orientation, and the RGB and depth maps acquired at the current position and orientation. If there is a target object in the acquired image, the state also includes the target object type and the target object's bounding box. The current position and orientation are represented by x, y, and w.

[0066] 1-1-4) The actions of a mobile robot in the workspace are defined as including action direction and action distance. In this embodiment, a set of action directions and a set of action distances are constructed respectively. Each action direction in the action direction set includes forward, backward, left, right, clockwise rotation, and counterclockwise rotation. Each action distance in the action distance set refers to the range of movement of the robot in that direction after the action direction is determined. In a specific embodiment of the present invention, the distance in each direction is divided into 3 levels: 1d, 2d, and 3d for forward, backward, left, and right directions, respectively; and 30 degrees, 60 degrees, and 90 degrees for clockwise and counterclockwise rotation directions, respectively.

[0067] 1-2) Construct a robot active recognition network;

[0068] In one specific embodiment of the present invention, the robot active recognition network is as follows: Figure 3 As shown, the robot active recognition network consists of a feature extraction subnetwork and a reinforcement learning subnetwork connected in sequence.

[0069] In this embodiment, the feature extraction sub-network includes a convolutional neural network layer connected in sequence (i.e., Figure 3 (convolutional layer) and two downsampling layers, such as Figure 3 The feature extraction subnetwork is shown in the diagram. The input to this subnetwork is the depth map and RGB image acquired by the robot at the same location and orientation in step 1-1), and the output is the target object parameters and depth feature vector at that location and orientation. Specifically, the depth map and RGB image are processed through a convolutional neural network layer to obtain the target object parameters in the robot-captured image. These parameters include the target object's center coordinates, width, height, depth, and (x, y) coordinates. 目标 y 目标 , w, h, d) represent, where the relationship between the center coordinates, width, and height of the target object and the parameters of the target object's graphic frame is as follows:

[0070]

[0071]

[0072] w = x max -x min

[0073] h = y max -y min

[0074] Where, x 目标 y 目标 These are the x and y coordinates of the target object's center point in the RGB image, respectively, and the depth d is the target object's center coordinate (x, y) in the depth image. 目标 y 目标 The corresponding depth value.

[0075] The depth map obtains feature vectors through convolutional neural network layers, two downsampling layers, and normalized feature extraction.

[0076] In one specific embodiment of the present invention, the convolutional neural network layer can be an existing feature extraction network or a 3-layer fully connected network.

[0077] In one specific embodiment of the present invention, MaskR-CNN with pre-trained weights is used as a feature extraction sub-network to output the feature map of the depth map, and is converted by the downsampling layer into a depth feature vector and the center coordinates, width, height and depth parameters of the target object.

[0078] In this embodiment of the invention, the reinforcement learning sub-network structure is as follows: Figure 3 As shown, the DQN algorithm is adopted, using a fully connected neural network of deep learning to replace the original Q-value table in reinforcement learning, that is, using a fully connected network layer (i.e. Figure 3 The Q-value is calculated using fully connected layers in the subnetwork. The input to this subnetwork is the target object parameters and depth feature vector output by the feature extraction subnetwork. To increase the efficiency of the policy in exploring the action space while ensuring the discreteness of the action space, this embodiment adopts a dual DQN architecture (dual-layer deep Q-network). The reinforcement learning subnetwork includes two parallel fully connected layers, each with the same input. Each fully connected layer is connected to an action set, namely the action direction set V. d With action distance set V r By using two action sets to output the robot's action direction and distance corresponding to the input image, and combining the outputs of these two action sets when finally outputting the action, complex actions with both direction and distance can be predicted simultaneously. Therefore, the Q-value of the reinforcement learning sub-network can be expressed as follows:

[0079] Q(s,a,θ)=V d (s,a d ,θ f ,θ d )+V r (s,a r ,θ f ,θ r )

[0080] Where s represents the state, and a represents the action, including: the direction of the action a d and action distance a r θ f For general decision parameters, θ d With θ r These are the direction decision parameters and the distance decision parameters, respectively.

[0081] The loss function for the reinforcement learning subnetwork is defined as follows:

[0082]

[0083] Where θ represents the parameters of the reinforcement learning subnetwork, γ is the loss factor, and s t Let a be the state value at time t. t Let r be the action value at time t, and Q(s,a,θ) be the action value function with state s, action a, and parameter θ. t The reward value function at time t

[0084] 1-3) Train the robot's active recognition network using the training set;

[0085] 1-3-1) For the feature extraction sub-network, the input during training is the RGB image and depth map of each training sample collected in 1-1), and the output is the target object parameters calculated based on the labeled target object bounding box. This sub-network is trained using a common convolutional network training method. In this embodiment, the loss function is defined as the root mean square error value, and training stops when it is less than 0.2.

[0086] In one specific embodiment of the present invention, training is performed using data from an ellipsoid (i.e., a target object) collected by a mobile robot in the workspace, to obtain a trained feature extraction subnetwork.

[0087] 1-3-2) For the reinforcement learning sub-network, this embodiment uses an ε-greedy exploration strategy and trains it using the training set. The training of this sub-network stops when a set upper limit for the number of training iterations is reached. In a specific embodiment of this invention, this is after 3 × 10 5 Rounds, each round using the training set at least 5 × 10 5 The training is repeated to obtain a fully trained reinforcement learning subnetwork.

[0088] In this embodiment, the feature extraction subnetwork is trained first, and then the reinforcement learning subnetwork is trained based on the output of the trained feature extraction subnetwork.

[0089] In a specific embodiment of the present invention, the training configuration is as follows: the hardware uses a desktop computer with a GPU (GTX1070), the training set of the feature extraction network is trained using data from an ellipsoid collected in the working environment, the learning rate of the reinforcement learning sub-network is set to α = 0.01, the probability of the exploration strategy is ∈-greedy = 0.9, the loss factor is γ = 0.9, and the reward value function is defined based on the size of the bounding box of the labeled target object in the current state as the input of the reward function, as shown in Table 2.

[0090] Table 2 Reward Value Function Table in a Specific Embodiment of the Present Invention

[0091]

[0092] After the reinforcement learning subnetwork and the feature extraction subnetwork are trained, the trained robot active recognition network is obtained.

[0093] 2) Testing phase;

[0094] 2-1) Equip the mobile robot with an RGB camera and a depth camera (the robot in the test phase and the robot in the training phase can be different) and place it in the same workspace as the training set data collection. Make the initial orientation of the mobile robot coincide with any orientation defined in step 1-1-2). It is not necessary to guarantee that there is a target object in the initial orientation.

[0095] It should be noted that the workspace environment during the testing phase is the same as that during the training phase, including obstacles and targets. (If a new target is to be selected, the feature extraction network needs to be retrained. The target is related to the feature extraction network; if the feature extraction network has multiple targets during training, one can be selected from these targets during testing.) The initial position of the mobile robot is any intersection point in the chessboard grid of the workspace plane.

[0096] 2-2) The RGB camera and depth camera acquire RGB and depth maps of the workspace at the current position and orientation and input them into the trained robot active recognition network. The network outputs the next action of the mobile robot, which includes: the action direction selected from the preset action direction set and the action distance corresponding to the current action direction selected from the preset action distance set, and controls the mobile robot to move accordingly.

[0097] 2-3) Repeat step 2-2) until the mobile robot moves to the position closest to the target object, and the recognition is complete.

[0098] In one specific embodiment of the present invention, the mobile robot is pre-positioned in the workspace at a location where the target object can be observed, and an ellipsoid is set as the target. Environmental images and depth maps are collected and input into a trained network, and the algorithm controls the robot to move to the optimal position.

[0099] To achieve the above embodiments, a second aspect of the present invention provides a mobile robot active recognition device, comprising:

[0100] The image acquisition module is used to enable the mobile robot to acquire RGB and depth images of its current position and orientation in a workspace with a pre-set target object.

[0101] The motion generation module is used to input the RGB image and depth image into a preset robot active recognition network, and the network outputs the motion of the mobile robot at the next moment, including: motion direction and motion distance;

[0102] The execution module is used to make the mobile robot move accordingly according to the output of the network until the mobile robot moves to the position closest to the preset target object, and the recognition is completed.

[0103] It should be noted that the foregoing explanation of an embodiment of a mobile robot active recognition method also applies to a mobile robot active recognition device in this embodiment, and will not be repeated here. According to an embodiment of the present invention, a mobile robot active recognition device acquires an RGB image and a depth image of its current position and orientation within a workspace containing a preset target object. The RGB image and depth image are input into a preset robot active recognition network, which outputs the mobile robot's action at the next moment, including the direction and distance of the action. The mobile robot moves accordingly according to the network's output until it reaches the position closest to the preset target object, at which point recognition is complete. This allows for the combination of depth and target object information, predicting the optimal field of view using an active vision algorithm. This overcomes the problem of insufficient target object information in the camera's field of view and significantly improves the low recognition reliability caused by target object occlusion or excessive distance.

[0104] To implement the above embodiments, a third aspect of the present invention provides an electronic device, comprising:

[0105] At least one processor; and a memory communicatively connected to said at least one processor;

[0106] The memory stores instructions that can be executed by the at least one processor, and the instructions are configured to perform the above-described mobile robot active recognition method.

[0107] To implement the above embodiments, a fourth aspect of the present invention provides a computer-readable storage medium storing computer instructions for causing the computer to execute the above-described mobile robot active recognition method.

[0108] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0109] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform a mobile robot active identification method according to the above embodiments.

[0110] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0111] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0112] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0113] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0114] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0115] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0116] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0117] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0118] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A mobile robot active recognition method, characterized in that, include: The mobile robot acquires an RGB image and a depth image of its current position and orientation within a workspace containing a pre-defined target object. The initial position of the mobile robot is any intersection point in a checkerboard pattern obtained by dividing the workspace plane according to a set interval. The initial orientation of the mobile robot is any orientation from a pre-defined set of orientations. The movement direction is a movement direction selected from a pre-defined set of movement directions. The movement distance is the movement distance corresponding to the movement direction selected from a pre-defined set of movement distances. The movement direction and the movement distance ensure that the mobile robot remains at the intersection point of the checkerboard pattern after movement. The RGB image and depth image are input into a preset robot active recognition network. The network outputs the robot's action at the next moment, including the action direction and action distance. The robot active recognition network includes a feature extraction sub-network and a reinforcement learning sub-network connected in sequence. The reinforcement learning sub-network uses a two-layer deep Q-network. The Q-value of the reinforcement learning sub-network is represented as follows: ； in, The parameters represent the reinforcement learning subnetwork, where s is the state and a is the action, including the action direction. and action distance , For general decision parameters, and These are the direction decision parameters and the distance decision parameters, respectively. These represent the set of motion directions and the set of motion distances, respectively. The mobile robot moves accordingly according to the output of the network until it moves to the position closest to the preset target object, at which point the identification is complete.

2. The method according to claim 1, characterized in that, Before inputting the RGB image and depth map into a preset robot active recognition network, the method further includes: Train the robot's active recognition network; Training the robot's active recognition network includes: A training set is obtained, wherein each sample in the training set is obtained by collecting a depth map and an RGB map from the same location and orientation in the workspace by a mobile robot. A robot recognition network is constructed, wherein the input of the feature extraction subnetwork is the depth map and RGB image collected by the mobile robot at the same location and in the same orientation, and the output is the target object parameters and depth feature vector at that location and in that orientation; the input of the reinforcement learning subnetwork is the target object parameters and depth feature vector output by the feature extraction subnetwork, and the output is the action of the mobile robot at the next moment, including: action direction and action distance. The feature extraction subnetwork is first trained using the training set, and then the reinforcement learning subnetwork is trained based on the output of the trained feature extraction subnetwork to obtain the trained robot active recognition network.

3. The method according to claim 2, characterized in that, The acquisition of the training set includes: A target object that is not completely obscured is placed in the workspace; Set the orientation set, motion direction set, and motion distance set; The workspace plane is divided into rectangular checkerboard patterns at set intervals. At each intersection of the checkerboard patterns, the mobile robot collects the corresponding RGB image and depth image according to each orientation of the orientation set. If the target object is included in the RGB image collected by the mobile robot, the type of the target object is recorded, and the graphic frame and image difficulty of the target object are labeled. The graphic frame of the target object uses... , , , express, This represents the minimum number of pixels in a row of the figure frame in the RGB image. This represents the maximum value of the row pixels of the figure frame in the RGB image. This represents the minimum value of the column pixels of the figure frame in the RGB image. This represents the maximum value of the column pixels of the figure frame in the RGB image.

4. The method according to claim 3, characterized in that, The feature extraction subnetwork uses a Mask R-CNN network with pre-trained weights; the feature extraction subnetwork is used to obtain the corresponding feature map of the input depth map and convert the feature map into the corresponding depth feature vector and target object parameters; The target object parameters include the target object's center coordinates, width, height, and depth. The relationship between the center coordinates, width, and height of the target object and the parameters of the target object's graphic frame is as follows: ； in, These represent the x and y coordinates of the target object's center point in the RGB image, respectively, and the depth. The coordinates of the center of the target object in the depth map The corresponding depth value.

5. The method according to claim 4, characterized in that, The loss function in the reinforcement learning subnetwork is: ； in, As a loss factor, for Time-state value, for Moment action value, In the state ,action ,parameter Action value function at time, for The reward value function at each time step.

6. A mobile robot active recognition device, characterized in that, include: The image acquisition module is used to enable the mobile robot to acquire RGB and depth images of its current position and orientation in a workspace with a preset target object. The initial position of the mobile robot is any intersection point of a chessboard grid obtained by dividing the workspace plane according to a set interval. The initial orientation of the mobile robot is any orientation from a preset orientation set. The movement direction is a movement direction selected from a preset movement direction set. The movement distance is the movement distance corresponding to the movement direction selected from a preset movement distance set. The movement direction and the movement distance ensure that the mobile robot remains at the intersection point of the chessboard grid after movement. An action generation module is used to input the RGB image and depth image into a preset robot active recognition network. The network outputs the action of the mobile robot at the next moment, including the action direction and action distance. The robot active recognition network includes a feature extraction sub-network and a reinforcement learning sub-network connected in sequence. The reinforcement learning sub-network adopts a two-layer deep Q-network. The Q-value of the reinforcement learning sub-network is represented as follows: ； in, The parameters represent the reinforcement learning subnetwork, where s is the state and a is the action, including the action direction. and action distance , For general decision parameters, and These are the direction decision parameters and the distance decision parameters, respectively. These represent the set of motion directions and the set of motion distances, respectively. The execution module is used to make the mobile robot move accordingly according to the output of the network until the mobile robot moves to the position closest to the preset target object, and the recognition is completed.

7. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method according to any one of claims 1-5.

Images