A Dynamic SLAM Method and System for Lightweight Real-Time Semantic Mapping

Abstract

A dynamic SLAM method and system for lightweight real-time semantic mapping, including: step S100, training and deploying a target detection neural network; step S200, embedding a tracking and prediction mechanism to remove feature points belonging to movable objects in all input frames; step S300, constructing a semantic octree map through probabilistic fusion of detection and tracking results to perform object-level description of the scene. This visual SLAM system can stably operate in a dynamic scene, maintain a high-precision positioning effect, complete real-time operation only on a CPU device without relying on a GPU, and the running frame rate reaches about 30 FPS; through the added semantic mapping thread, the semantic information generated by the target detection network is fully utilized to generate a static semantic octree map that removes all movable objects.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and particularly relates to a dynamic SLAM method and system for lightweight real-time semantic mapping. Background Art

[0002] Most existing SLAM algorithms in dynamic scenarios rely on object detection, semantic segmentation, or instance segmentation networks to distinguish dynamic and static feature points. Usually, they need to be accelerated by a GPU to complete. Even with a parallel architecture, the running frame rate FPS is only about 10 frames, and it cannot work in real time without a GPU. Therefore, it cannot meet lightweight applications. Most dynamic SLAM algorithms based on ORB-SLAM mostly eliminate dynamic feature points and leave static feature points to construct a sparse landmark map, without making full use of semantic information for mapping. In the prior art, classic visual SLAM algorithms usually assume that the scene is rigid and static. This assumption causes visual SLAM systems to frequently malfunction in dynamic scenarios, and the positioning accuracy drops severely. In the prior art, SLAM systems that can generate octree maps generally delete moving objects by increasing the threshold p. These methods can create an octree map without moving objects, but cannot delete potentially movable objects. Summary of the Invention

[0003] In view of the above problems in the prior art, the present invention proposes a dynamic SLAM method and system for lightweight real-time semantic mapping. Specifically, the present invention is implemented by the following technical solutions:

[0004] The first aspect of the present invention provides a dynamic SLAM method for lightweight real-time semantic mapping, including:

[0005] Step S100, training and deploying a target detection neural network, and step S200, embedding a tracking and prediction mechanism to remove feature points belonging to movable objects in all input frames;

[0006] Step S300, constructing a semantic octree map through probability fusion of detection and tracking results to perform object-level description of the scene.

[0007] Further, the step S100 includes: adopting an ultra-lightweight target detection neural network in key frames, and training the target detection neural network using the COCO dataset.

[0008] Further, the step S200 includes:

[0009] Step S210, the SLAM system inputs data through the RGB-D camera, extracts ORB feature points on the RGB image, uses the image pyramid and quadtree method to evenly sample points, determines whether the frame is added as a key frame, and starts detection if it is a key frame, detects object labels in the image and generates a bounding box, and uses a target tracking algorithm to track multiple detection results of the key frame;

[0010] Step S220, determine whether the label in the image is a dynamic object label. If so, mark the bounding box as a dynamic object box, pass the boundary information into the tracking process, remove the feature points in the dynamic object box, use the remaining static feature points to estimate the pose, match the static feature points of two adjacent frames, and then use the iterative closest point to solve the rotation matrix and translation, that is, the change in the camera's pose. If the current frame is not selected as a key frame, find the nearest key frame and use the constant velocity model to predict the position of multiple detection results of the key frame in the current frame;

[0011] Step S230 , using the constant velocity model to predict the position of the moving object in the latest key frame in the current coordinate system, and obtaining a prediction result of the position of the bounding box of the moving object in the current frame.

[0012] Furthermore, in step S230, if the current frame is a key frame, the coordinates P of the position matrix of the dynamic object bounding box in the world coordinate system are obtained through the target detection network. w (x,y,z), the position matrix P w Project to the two-dimensional coordinate plane (xOy, yOz, zOx) and get the matrix P xy (x,y),P yz (y,z) and P zx (x,z), using the SORT algorithm to simultaneously track and predict P xy , P yz , P zx ; If the current frame is not selected as a key frame, the input is empty, and the constant speed model is used to predict the P of the current frame based on the previous key frame. xy , P yz , P zx And update; traverse the three matrices, calculate their IOU projections on the xOy plane, record the maximum value of the IOU and the corresponding matrix index, and find the P that meets the conditions according to the index xy , P yz , P zx Matrix, thereby obtaining the coordinates P of the bounding box position matrix of the predicted current frame in world coordinates wp (x p ,y p ,z p );

[0013] After detection and prediction, the detection results of key frames are used to predict the positions of moving objects in the current frame, and then the prediction results are used to remove the feature points on the movable objects in the current frame.

[0014] Further, the step S300 includes:

[0015] Step S310, using a point cloud segmentation module on the key frame to generate a point cloud map in combination with the depth map corresponding to the key frame, and then segmenting and aggregating the three-dimensional point cloud to obtain the boundaries of three-dimensional objects;

[0016] Step S320, the 3D object database construction module uses the label information generated by the object detection network to establish a three-dimensional object database;

[0017] Step S330, after the point cloud is segmented, it is passed into the updated semantic octree map module to filter the current point cloud and place it in the candidate occupied point cloud;

[0018] Step S340, the updated semantic octree map module will query whether the candidate occupied point is located in the three-dimensional bounding box. If it is true, it will assign the corresponding color according to its label. At the same time, this module will query whether the candidate occupied point is located in the three-dimensional bounding box of the movable object. If so, it will be marked as a movable point;

[0019] Step S350, update the occupancy probability of the octree map. The octree map provides a point cloud structure inside, which only carries the spatial position information of the points, and uses a probability form to express the fact of whether a certain node is occupied. When there is a new observation result, its occupancy probability will be updated.

[0020] The present invention also provides a dynamic SLAM system for lightweight real-time semantic mapping, including:

[0021] A training module that trains and deploys a feature point removal module for the object detection neural network, and embeds a tracking and prediction mechanism to remove the feature points belonging to movable objects in all input frames;

[0022] An object-level description module that constructs a semantic octree map through the probability fusion of detection and tracking results to perform object-level description of the scene.

[0023] The present invention also relates to an electronic device, and the electronic device includes:

[0024] At least one processor; and,

[0025] A memory communicatively connected to the at least one processor; wherein,

[0026] The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method described above.

[0027] The present invention also relates to a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the method described above.

[0028] The technical solution of the present invention can achieve the following beneficial technical effects:

[0029] The visual SLAM system can operate stably in a dynamic scene, maintain a high-precision positioning effect, and complete real-time operation only on a CPU device without relying on a GPU, with a frame rate of about 30 FPS; through the added semantic mapping thread, the semantic information generated by the object detection network is fully utilized to generate a static semantic octree map that removes all movable objects. Description of the Drawings

[0030] Figure 1 is a schematic flowchart of the method of the present invention;

[0031] Figure 2 is a detailed flowchart of the method of the present invention. Detailed Description of the Embodiments

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0033] The present invention will be described in detail below in conjunction with the accompanying drawings and embodiments.

[0034] The first aspect of the present invention provides a dynamic SLAM method for lightweight real-time semantic mapping. Specifically, the present invention uses a ultra-lightweight object detection network NanoDet-Plus-m-1.5 module to detect "prior" moving objects, that is, objects that may move, in key frames, and then removes the feature points on the dynamic objects as outliers, and uses static feature points for feature matching and motion estimation, so as to solve the problem of the failure of visual SLAM algorithms in dynamic scenarios; uses the ultra-lightweight object detection network NanoDet-Plus and accelerates it with the high-performance neural network forward propagation framework NCNN, so that the algorithm does not rely on GPU and can run at high speed on CPU devices, laying a foundation for subsequent applications on mobile devices; constructs a semantic octree map to show whether a position is occupied, providing a more complete map description of what a occupied point represents at the object level. A detection and prediction module is added to the tracking thread of the ORB-SLAM3 algorithm, and dynamic objects are only detected on key frames, and the position of the movable object in the current frame is predicted through the two-dimensional bounding box detected on the nearest key frame.

[0035] Specifically, the method of the present invention includes the following steps:

[0036] Step S100, training and deploying the object detection neural network. Specifically, a ultra-lightweight object detection neural network - NanoDet-Plus is adopted in key frames, and the object detection neural network is trained using the COCO dataset. Specifically, the COCO dataset has more than 200,000 labeled images, including 80 object categories, covering common objects in life. Among them, people, animal categories, and vehicle categories (a total of 20 categories) can be regarded as dynamic objects and may move. The training images and test images are cropped to a size of 416*416 pixels, and the model is trained using the AdamW optimizer with a learning rate (lr) of 0.001 and a weight decay coefficient of 0.05. The Batchsize is set to 96, and the learning rate is adjusted using the cosine annealing method (CosineAnnealingLR), with the minimum learning rate being 0.00005. A total of 300 epochs are trained, and the network training is completed using parallel computing on 4 NVIDIA RTX 3090 graphics cards. According to the evaluation criteria of the COCO dataset, its accuracy can reach 0.341 mAP (0.5:0.95).

[0037] Deploy the neural network model based on Pytorch through the high-performance neural network forward computing framework (NCNN) to accelerate its subsequent use in the SLAM system. The specific steps are as follows: First, export the pth model file generated by the trained model into an ONNX file through torch.onnx.export, then convert it into an ncnn model through the onnx2ncnn tool in NCNN, and finally perform compilation. After the deployment, the average time consumption for each key frame detection is 22.53 ms, and the running frame rate of the SLAM system can reach 30 FPS.

[0038] Step S200: Embed a tracking and prediction mechanism to remove the feature points belonging to movable objects in all input frames.

[0039] Specifically, the step S200 includes:

[0040] Based on the three threads of ORB-SLAM3 tracking, local mapping, and loop closing detection, add a thread for detecting and constructing an octree map, and insert a "prediction" module into the tracking. Use the key frame decision mechanism as the detection mechanism, and only use the object detection NanoDet-Plus-m-1.5 module in key frames. The input is the RGB-D data of the key frame (including the RGB image and the corresponding depth map), and the output is the two-dimensional bounding box, the depth of the center point of the bounding box, and the object label.

[0041] Step S210: The SLAM system inputs RGB-D data through an RGB-D camera, extracts ORB feature points on the RGB image, uses the image pyramid and quadtree method to evenly sample points, determines whether this frame is newly added as a key frame. If it is a key frame, start detection, detect the object label in the image and generate a bounding box, and use the object tracking algorithm to track multiple detection results of the key frame.

[0042] Specifically, the SLAM system uses the input generated by the RGB-D camera, that is, the data collected by the sensor; the RGB-D camera can collect RGB images and corresponding depth images, and the above RGB refers to the RGB image of the RGB-D camera.

[0043] Specifically, use the SORT algorithm commonly used in object tracking algorithms. SORT is an algorithm that uses a constant velocity Kalman filter framework and the Hungarian algorithm to track two-dimensional bounding boxes. If the current frame is a key frame, use SORT to predict. If the current frame is not a key frame, use the constant velocity model and the nearest key frame to predict.

[0044] Step S220, determine whether the label in the image is a dynamic object label. If so, mark the bounding box as a dynamic object box, pass the boundary information into the tracking process, remove the feature points in the dynamic object box, use the remaining static feature points to estimate the pose, match the static feature points of two adjacent frames, and then use the iterative closest point (ICP) to solve the rotation matrix and translation, that is, the change in the camera's pose. If the current frame is not selected as a key frame, find the nearest key frame and use the constant velocity model to predict the position of multiple detection results of the key frame in the current frame.

[0045] Step S230, using a constant speed model (based on the speed estimated in the previous frame, assuming that the speed remains unchanged, to obtain the initial posture of the current frame) to predict the position of the moving object in the most recent key frame in the current coordinate system, and obtain the predicted result of the bounding box position of the moving object in the current frame.

[0046] Specifically, if the current frame is a key frame, the position matrix of the dynamic object bounding box in the world coordinate system is obtained through the target detection network, and the position matrix P w Project to the two-dimensional coordinate plane (xOy, yOz, zOx) and get the matrix P xy (x,y),P yz (y,z) and P zx (x,z), using the SORT algorithm to simultaneously track and predict P xy , P yz , P zx ; If the current frame is not selected as a key frame, the input is empty, and the constant speed model is used to predict the P of the current frame based on the previous key frame. xy , P yz , P zx Traverse the three matrices, calculate the IOU of their projections on the xOy plane, record the maximum IOU value and the corresponding matrix index, and find the P that meets the conditions according to the index. xy , P yz , P zx Matrix, thereby obtaining the coordinates P of the bounding box position matrix of the predicted current frame in world coordinates wp (x p ,y p ,z p ).

[0047] After detection and prediction, the detection results of the key frame are used to predict the position of the active object in the current frame, and then the prediction results are used to remove the feature points on the movable object in the current frame.

[0048] The present invention does not use object detection in all frames.

[0049] Step S300, construct a semantic octree map through probability fusion of detection and tracking results to perform object-level description of the scene.

[0050] Specifically, the step S300 includes:

[0051] Step S310, use the point cloud segmentation module on the key frame to generate a point cloud map in combination with the depth map corresponding to the key frame, and then segment and aggregate the three-dimensional point cloud to obtain the boundaries of three-dimensional objects. Specifically, use the PointCloud Library to generate the point cloud of the current key frame, and then filter it using a voxel filter, that is, replace the point cloud within a voxel with its centroid, and then decompose the point cloud into independent point cloud clusters according to Euclidean distance segmentation and clustering. Then, back-project the point cloud clusters onto the RGB image plane, calculate the 2D rectangular box of the projection points corresponding to the point cloud clusters, and then calculate the IOU (Intersection over Union) with the 2D rectangular box of the object detection. Each detection box matches a point cloud cluster to calculate the 3D information of the point cloud cluster.

[0052] Step S320, the 3D object database construction module uses the label information generated by the object detection network to establish a three-dimensional object database. Specifically, during the construction of the object database, first generate the 3D target objects existing in each frame of observation according to steps S230 and S310, and then add them to the database. When the same category is detected again next time, first judge whether it is the same object according to the position information. If it is the same object, update the spatial coordinates and other information of the object, otherwise record it as a new object.

[0053] After the point cloud is segmented, it is passed into the updated semantic octree map module to filter the current point cloud and place it in the candidate occupied point cloud.

[0054] Step S340, the updated semantic octree map module will query whether the candidate occupied point is located in the three-dimensional bounding box. If it is true, assign the corresponding color according to its label. At the same time, the module will query whether the candidate occupied point is located in the three-dimensional bounding box of the movable object. If so, it will be marked as a movable point.

[0055] Step S350, update the occupancy probability of the octree map. The octree map provides a point cloud structure internally, which only carries the spatial position information of the points, and uses a probability form to express the fact of whether a certain node is occupied. When there is a new observation result, its occupancy probability will be updated.

[0056] Specifically, assume that n is one of these nodes. To avoid the occupancy probability exceeding the interval, use the probability logarithm value of the occupancy probability P(n) ∈ [0, 1] of node n to describe, When "occupation" is continuously observed, increment the value of L(n); otherwise, decrement the value of L(n). When querying the probability, use the inverse logit transformation to convert L(n) to the probability P(n). Mathematically speaking, let a certain node be n and the observed data be z. Then the logarithm of the probability of a certain node from the start to time t is L(n|z 1:t ) and L(n|z 1:t ) = L(n|z 1:t-1 ) + L(n|z t ), that is, the logarithm of the probability of a certain node from the start to time t is equal to the logarithm of the probability of a certain node from the start to time t - 1 plus the logarithm of the occupation probability of the current node. If a node n is inserted at time t, L(n|z t ) = τ, and if it is not inserted, L(n|z t ) = 0. The occupation probability threshold is p, and the principle for a node n to be considered occupied is

[0057] In the past, SLAM systems that could generate octree maps generally deleted moving objects by increasing the threshold. These methods can create an octree map without moving objects, but they cannot delete potentially movable objects. After increasing the threshold p, when the robot moves fast and the observation time for each scene is limited, less frequently observed static scenes will also be removed.

[0058] To overcome this drawback, the present invention removes all movable objects, including moving objects and potentially movable objects, by inserting points marked with different probabilities as movable or static, and at the same time it does not remove less frequently observed static scenes. The occupation probability threshold is set to the default p = 0.5. For points marked as movable, we set L(n|z t ) = -0.4, and for other points, we set L(n|z t ) = 0.8, and if not observed, L(n|z t ) = 0.

[0059] The present invention also relates to a dynamic SLAM system for lightweight real-time semantic mapping, which is a lightweight and real-time RGB-D vSLAM system based on ORB-SLAM3 in a dynamic scenario. It includes:

[0060] A training module for training and deploying a target detection neural network;

[0061] A feature point removal module that embeds a tracking and prediction mechanism to remove feature points belonging to movable objects in all input frames;

[0062] An object-level description module that constructs a semantic octree map through probability fusion of detection and tracking results to perform object-level description of the scene.

[0063] In summary, the present invention provides a dynamic SLAM method and system for lightweight real-time semantic mapping, including: step S100, training and deploying a target detection neural network; step S200, embedding a tracking and prediction mechanism to remove feature points belonging to movable objects in all input frames; step S300, constructing a semantic octree map through probability fusion of detection and tracking results to perform object-level description of the scene. This visual SLAM system can stably operate in a dynamic scene, maintain a high-precision positioning effect, and complete real-time operation only on a CPU device without the aid of a GPU, with a running frame rate of about 30 FPS; through the added semantic mapping thread, the semantic information generated by the target detection network is fully utilized to generate a static semantic octree map that removes all movable objects.

[0064] It should be understood that the above specific embodiments of the present invention are only used for exemplary illustration or explanation of the principle of the present invention, and do not constitute a limitation to the present invention. Therefore, any modifications, equivalent replacements, improvements, etc. made without departing from the spirit and scope of the present invention shall be included within the protection scope of the present invention. In addition, the appended claims of the present invention are intended to cover all variations and modifications that fall within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A dynamic SLAM method for lightweight real-time semantic mapping, characterized in that, Including: Step S100, training and deploying a target detection neural network; Step S200, embedding a tracking and prediction mechanism to remove the feature points belonging to movable objects in all input frames; including: Step S210, the SLAM system inputs data through an RGB-D camera, extracts ORB feature points on the RGB image, uses the image pyramid and quad-tree method to evenly sample points, determines whether the frame is newly added as a key frame, if it is a key frame, starts detection, detects the object labels in the image and generates bounding boxes, and uses the object tracking algorithm to track multiple detection results of the key frame; Step S220, determining whether the label in the image is a dynamic object label, if so, marking the bounding box as a dynamic object box, passing the boundary information into the tracking process, removing the feature points within the dynamic object box, using the remaining static feature points for pose estimation, performing static feature point matching on two adjacent frames, and then using the iterative closest point to solve for the rotation matrix and translation, that is, the pose change of the camera. If the current frame is not selected as a key frame, find the nearest key frame and use the constant velocity model to predict the positions of multiple detection results of the key frame in the current frame; Step S230: Use a constant velocity model to predict the position of the moving object in the current coordinate system in the nearest key frame, and obtain the prediction result of the position of the moving object bounding box in the current frame; if the current frame is a key frame, obtain the coordinates of the position matrix of the dynamic object bounding box in the world coordinate system through the object detection network , project the position matrix onto the two-dimensional coordinate plane to obtain the matrix , and , and use the SORT algorithm to track and predict simultaneously , , ; if the current frame is not selected as a key frame, the input is the null value empty, and use the constant velocity model to predict the , , of the current frame according to the previous key frame and update; traverse the three matrices, calculate the IOU of their projections on the plane, record the maximum value of the IOU and the corresponding matrix index, and find the , , matrix that meets the conditions, so as to obtain the coordinates of the predicted position matrix of the bounding box in the current frame in the world coordinates ; After detection and prediction, use the detection results of the key frame to predict the positions of active objects in the current frame, and then use the prediction results to remove the feature points on the movable objects in the current frame; Step S300, constructing a semantic octree map through the probability fusion of detection and tracking results to perform object-level description of the scene; Step S310, using the point cloud segmentation module on the key frame to generate a point cloud map in combination with the depth map corresponding to the key frame, and then performing segmentation and aggregation on the three-dimensional point cloud to obtain the boundaries of three-dimensional objects; Step S320, the 3D object database construction module uses the label information generated by the target detection network to establish a three-dimensional object database; Step S330, after the point cloud is segmented, it is passed into the updated semantic octree map module to filter the current point cloud and place it in the candidate occupied point cloud; Step S340, the updated semantic octree map module will query whether the candidate occupied point is located in the three-dimensional bounding box. If true, assign the corresponding color according to its label. At the same time, this module will query whether the candidate occupied point is located in the three-dimensional bounding box of the movable object. If so, it will be marked as a movable point; Step S350, update the occupancy probability of the octree map. The octree map internally provides a point cloud structure that only carries the spatial position information of the points, and uses a probability form to express the fact of whether a certain node is occupied. When there is a new observation result, its occupancy probability will be updated.

2. The dynamic SLAM method for lightweight real-time semantic mapping according to claim 1, characterized in that, The said step S100 includes: adopting an ultra-lightweight target detection neural network in the key frame and training the target detection neural network using the COCO dataset.

3. A dynamic SLAM system for lightweight real-time semantic mapping, characterized in that, Including: A training module for training and deploying a target detection neural network; A feature point removal module for embedding a tracking and prediction mechanism to remove the feature points belonging to movable objects in all input frames; including: Step S210, the SLAM system inputs data through the RGB-D camera, extracts ORB feature points on the RGB image, uses the image pyramid and quadtree method to evenly sample points, determines whether the frame is added as a key frame, and starts detection if it is a key frame, detects object labels in the image and generates a bounding box, and uses a target tracking algorithm to track multiple detection results of the key frame; Step S220, determine whether the label in the image is a dynamic object label. If so, mark the bounding box as a dynamic object box, pass the boundary information into the tracking process, remove the feature points in the dynamic object box, use the remaining static feature points to estimate the pose, match the static feature points of two adjacent frames, and then use the iterative closest point to solve the rotation matrix and translation, that is, the change in the camera's pose. If the current frame is not selected as a key frame, find the nearest key frame and use the constant velocity model to predict the position of multiple detection results of the key frame in the current frame; Step S230: Use the constant velocity model to predict the position of the moving object in the current coordinate system in the nearest key frame, and obtain the prediction result of the position of the moving object bounding box in the current frame; if the current frame is a key frame, obtain the coordinates of the position matrix of the dynamic object bounding box in the world coordinate system through the target detection network , project the position matrix onto the two-dimensional coordinate plane , and obtain the matrix , and . Use the SORT algorithm to track and predict simultaneously , , ; if the current frame is not selected as a key frame, the input is the null value empty, and use the constant velocity model to predict the current frame based on the previous key frame , , and update it; traverse the three matrices, calculate the IOU of their projections on the plane, record the maximum value of the IOU and the corresponding matrix index, and find the , , matrix that meets the conditions, so as to obtain the coordinates of the predicted position matrix of the bounding box in the current frame in the world coordinates ; After detection and prediction, the detection results of the key frame are used to predict the position of the active object in the current frame, and then the prediction results are used to remove the feature points on the movable object in the current frame; The object-level description module constructs a semantic octree map by probabilistic fusion of detection and tracking results to describe the scene at the object level. Step S310: Generate a point cloud map on the key frame using the point cloud segmentation module combined with the depth map corresponding to the key frame, and then segment and aggregate the 3D point cloud to obtain the boundary of the 3D object. Step S320, the 3D object database construction module uses the label information generated by the object detection network to establish a three-dimensional object database; Step S330, after the point cloud is segmented, it is passed to the updated semantic octree map module to filter the current point cloud and put it into the candidate occupied point cloud; Step S340, the update semantic octree map module will query whether the candidate occupied point is located in the three-dimensional bounding box. If so, it will be assigned a corresponding color according to its label. At the same time, the module will query whether the candidate occupied point is located in the three-dimensional bounding box of the movable object. If so, it will be marked as a movable point; Step S350, updating the occupancy probability of the octree map. The octree map provides a point cloud structure inside, which only carries the spatial position information of the point and uses probability form to express whether a node is occupied. When there is a new observation result, its occupancy probability will be updated.

4. An electronic device, characterized in that, The electronic device comprises: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of the preceding claims 1 to 2. 5 . A non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the method of claim 1 .

Images