Mobile robot dead reckoning method based on Trans-GCN

Abstract

The invention relates to a mobile robot dead reckoning method based on Trans-GCN. The method comprises the following steps: S1, collecting data of a wheel speed sensor and an inertial sensor of the wheeled mobile robot, intercepting time sequence data through a sliding window and an adaptive sliding step length, and representing the time sequence data of the sensor as graph structure data; s2, constructing a Laplacian position code for each node in the graph, and aggregating the Laplacian position code into a new embedding representation of the node after linear pooling; s3, embedding the graph feature signal and the coding information into Transform, and inputting the output feature of the overall model into a full connection layer to obtain an output vector; s4, constructing a training truth value, training the model according to the sample data, and stopping training and storing the model when the training truth value is smaller than a threshold value; and S5, inputting the data under the condition that the satellite signal is missing into the trained model to obtain the predicted two-dimensional position increment within the sampling time period of the mobile robot so as to realize dead reckoning of the basic body perception sensor signal without the GNSS signal.

Description

Technical Field

[0001] The present invention belongs to the field of dead reckoning of mobile robots, and particularly relates to a dead reckoning method based on a graph convolutional neural network and a Transformer structure. Background Art

[0002] In the context of the transformation of modern technology from automation to intelligence, robot technology, as a model of the integration of multiple disciplines such as mechanical design, sensing technology, electronic information, automatic control theory, and artificial intelligence, is widely used in various fields. Among them, the role of mobile robots in various production and daily life scenarios has become increasingly prominent, and the efficiency and safety of tasks such as power inspection, security search and rescue, and material distribution have been significantly improved. However, the core prerequisite for a mobile robot to successfully handle various challenging tasks is its autonomous positioning ability, and autonomous positioning is also a prerequisite for decision-making and path planning. Wheeled mobile robots, with their flexible movement mode, high stability, and excellent load-bearing capacity, are the most widely used mobile robots. The dead reckoning system based on the wheel speed sensor and inertial sensor of wheeled mobile robots has been widely studied. Among them, the odometer measures the number of pulses of the wheel per unit time through the wheel encoder to calculate the wheel speed and the driving distance. However, due to factors such as uneven ground friction and unstable heading calculation of differential speed, pure odometer positioning may have various cumulative errors. The IMU (Inertial Measurement Unit) usually includes two types of sensors, an accelerometer and a gyroscope, which are used to measure and output the acceleration and angular velocity information of the carrier. The IMU measurement is affected by multiple noises, and these noises gradually amplify during integration, resulting in the gradual accumulation of attitude angle and displacement errors, and an exponential growth. Traditional filtering methods rely on accurate system models and noise statistical characteristics, and may be difficult to adapt to the changes of the model with the dynamic environment, resulting in an increase in estimation errors. With the continuous progress of big data and artificial intelligence technologies, data-driven methods mainly based on machine learning are widely used to solve navigation and positioning-related problems such as sensor calibration, suppressing the divergence of positioning errors, and multi-sensor fusion. The dead reckoning of mobile robots based on multi-sensor fusion can be regarded as a prediction problem between multivariate time series and sequences. Not only from the time perspective, the internal connection and interaction between sensor data will also affect the prediction accuracy.

[0003] The present invention proposes a position prediction model Trans-GCN integrating GCN and Transformer, uses GCN to learn complex topological structures to capture the spatial dependence between data, proposes a node embedding method that fuses node position features with Laplace vectors, embeds graph feature signals and encoded information into the Transformer to enhance the global position perception ability of nodes in the graph, further improves the expression ability of the graph structure, makes the model more globally perceptive during feature extraction, and models time series data from multiple scales. Summary of the Invention

[0004] The present invention aims to solve the above problems of the prior art. A dead reckoning method for a mobile robot based on Trans-GCN is proposed. The technical solution of the present invention is as follows:

[0005] A dead reckoning method for a mobile robot based on Trans-GCN, comprising the following steps:

[0006] S1: Collect the data of the wheel speed sensor and the inertial sensor of the wheeled mobile robot, and intercept the time series data through a sliding window and an appropriate sliding step. In each window of data, construct the sensor data of a single dimension into graph nodes, and construct a bidirectional edge connection structure for some nodes to form complete graph structure data;

[0007] S2: Construct a graph convolutional network GCN to extract the dependency relationship of the data in the non-Euclidean space in the graph, including three steps of feature transfer, aggregation, and update, to obtain node features and edge features;

[0008] S3: Decompose the adjacency matrix of the constructed graph, construct the position encoding matrix of the nodes, and after integration through a linear layer and an L2 pooling layer, perform a linear transformation with the features extracted in S2 and input them into the Transformer encoder structure.

[0009] S4: Calculate the correlation between the target node and the neighbor nodes in each layer to obtain the attention weights, and then connect the attention features and the original features residually and integrate them through a feed-forward neural network FFN, and finally input them into a fully connected network to obtain the predicted value of the network;

[0010] S5: Train the model according to the training sample data, stop training when the verification loss is less than the threshold, save the model, and then input the data of the wheel speed sensor and the inertial sensor into the model to obtain the dead reckoning result without GNSS signal.

[0011] Further, the step S1 specifically includes:

[0012] Use the inertial sensor of the mobile robot body to record the time series data of the three-axis angular velocity and the three-axis acceleration, and the two-dimensional data of the left and right wheels respectively recorded by the wheel speed sensor, and set a sliding window and a sliding step of a certain length, and perform dimensionality reduction processing on the data, take the mean and standard deviation of the single-dimensional window data, align the true value labels according to the time stamp, and then model the single-dimensional data of the training data without labels into nodes, and set the edge connection between the nodes to generate a recorded adjacency matrix, and construct a complete graph structure training data.

[0013] Further, the construction of the position encoding matrix in step S3 includes three-dimensional feature data: after obtaining the feature matrix from the Laplacian matrix of the graph, the normalized eigenvectors and standardized eigenvectors corresponding to the first m smallest eigenvalues are selected, and together with the original eigenvectors, a three-dimensional initial position encoding matrix is constructed. Then, a new embedding representation is obtained through linear layer aggregation, and finally, the position encoding is obtained through L2 pooling layer for propagation learning.

[0014] Further, in step S4, the input data passes through four fully connected layers to obtain four matrices of Query (Q), Key (K), Value (V), and Edge (E), projecting the feature data into the attention space. The calculation formula of the Attention-score matrix for the h-th attention head at the k-th layer is:

[0015]

[0016] where i is the target node number, j is the neighbor node, and d K represents the dimension of the Key vector. The Softmax function calculates each row of the matrix, and the formula is as follows:

[0017]

[0018] where y a represents the value of the a-th column in the a-th row of the Attention-score matrix, y b represents the value of the b-th column in the a-th row of the matrix, and w represents the number of columns of the matrix.

[0019] Further, when calculating the Attention-score, a residual connection is introduced. At the same time, Layernorm normalizes the input data. The features of each node pass through a two-layer feedforward neural network for further non-linear mapping of feature representation. The specific formula is as follows:

[0020]

[0021] where is the Layernorm layer, is the multi-head concatenated self-attention, W is the weight matrix, and Relu is the activation function. represents the hidden state after normalization by LayerNorm, is the hidden state after passing through the feedforward neural network and the ReLu activation function, adding non-linear transformation to enable the model to learn more complex features.

[0022] Further, in S4, the attention feature and the original feature are concatenated residually and then integrated through a feed-forward neural network FFN, and finally input into a fully connected network to obtain the predicted value of the network, which specifically includes:

[0023] After passing through two layers of feed-forward neural networks, an adaptive learning rate adjustment factor is designed, that is, the ratio of the current gradient norm to the weight norm. The local learning rate scaling factor is calculated as follows:

[0024]

[0025] where η represents the global learning rate, is the gradient of the current k-th layer on the current batch of data, and λ represents the weight decay coefficient. Finally, the node feature output after Trans-GCN is:

[0026]

[0027] Then, after passing through the fully connected layer, anti-normalization is performed to obtain the predicted output of the model.

[0028] Further, in step S5, the training ground truth is derived from the latitude and longitude data output by the RTK differential positioning system. The conversion of the latitude and longitude data is as follows. The latitude and longitude coordinates in the geographic coordinate system are converted into the local coordinate system. Let r e represent the equatorial radius of the earth, r p represent the polar radius of the earth, lat 0 be the radian latitude of the reference point. The effective radius at the reference point is expressed as:

[0029]

[0030] r ns 、r ew are used to describe the radius of curvature of the earth in different directions and affect the calculation of position information:

[0031] r ns represents the radius of curvature in the north-south direction (meridian direction) and aids in the calculation of displacements in the auxiliary latitude direction;

[0032] r ew represents the radius of curvature in the east-west direction (parallel direction) and aids in the calculation of displacements in the auxiliary longitude direction.

[0033] Thus, the change in latitude and longitude coordinates can be converted into the displacement change in the plane coordinate system in the local coordinate system. Let lat rad and lon rad represent the radian values of the target latitude and longitude, and N and E represent the northward displacement and eastward displacement:

[0034]

[0035] After aligning the timestamp after dimensionality reduction of the input data with the longitude-latitude timestamp, the model can be trained and feature extraction can be performed.

[0036] An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the program, it implements the Trans-GCN-based dead reckoning method for mobile robots as described in any one of the above.

[0037] A non-transitory computer-readable storage medium, having stored thereon a computer program, wherein when the computer program is executed by a processor, it implements the Trans-GCN-based dead reckoning method for mobile robots as described in any one of the above.

[0038] The advantages and beneficial effects of the present invention are as follows:

[0039] The Trans-GCN model of the present invention is a novel neural network model with multi-scale spatial search. It overcomes the limitation of traditional dead reckoning methods that only rely on time series modeling.

[0040] 1. Break through the traditional time series modeling limitation and achieve multi-scale spatial feature learning: By constructing graph-structured data (Claim 1, S1-S2), the present invention models sensor data at different time steps as graph nodes, and extracts their associated information in the non-Euclidean space through a graph convolutional network (GCN), thereby increasing the spatial structure modeling ability beyond the time dimension.

[0041] 2. Innovative node position encoding method to improve feature learning ability, combined with the global modeling ability of Transformer to enhance the node feature fusion effect: In traditional GCN methods, nodes can only aggregate information from local neighborhoods and it is difficult to capture long-distance dependencies. The present invention proposes a three-dimensional position encoding method that combines Laplacian features and normalized eigenvectors (Claim 3). Compared with traditional methods that only rely on absolute position encoding, this method can more accurately represent the relationship between nodes, and further optimizes the embedded features in the L2 pooling layer to improve feature learning ability. When calculating the attention scores of the graph, the present invention adopts the Query-Key-Value-Edge (QKVE) mechanism (Claims 4-5), considering both node features and edge features in the calculation. And through residual connection and LayerNorm, the model can be trained more stably, preventing the problem of gradient disappearance.

[0042] 3. Introduce an adaptive learning rate adjustment factor to accelerate model convergence: Traditional deep learning models often have difficulty selecting an appropriate learning rate during training, resulting in slow convergence or getting stuck in local optima. The present invention designs an adaptive learning rate adjustment factor (Claim 6), which dynamically adjusts the learning rate based on the ratio of the current gradient norm to the weight norm, improves the training stability of the model, and accelerates the convergence speed.

[0043] The present invention can provide a new perspective and new thinking for the dead reckoning method of mobile robots, and promote the development of high-precision autonomous positioning of unmanned systems from the perspective of more intelligent and diversified methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the overall model learning data information flow designed in the preferred embodiment provided by the present invention.

[0045] Figure 2 It is a schematic diagram of the graph structure construction method in the embodiment of the present invention.

[0046] Figure 3 It is a schematic diagram of the implementation of the overall dead reckoning.

[0047] Figure 4 It is a flowchart of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0049] The technical solution of the present invention to solve the above technical problems is:

[0050] First, the present invention provides a dead reckoning method for mobile robots based on Trans-GCN, as Figure 4 shown, including collecting data and processing, inputting the data into the model, training the model, updating the parameters, and performing dead reckoning of the robot when the GNSS signal is lost and the absolute position cannot be obtained. The specific steps of model training verification and testing are as Figure 3 shown, including the following steps:

[0051] Step 1: Use the inertial sensor located inside the robot and the wheel angular velocity recorded by the wheel encoder to record the odometry data and attitude change data of the robot's movement, and set a sliding window with a certain length to intercept the corresponding length of data and the true displacement change amount corresponding to each sliding window;

[0052] Step 1.1: Sliding window segmentation

[0053] Suppose there are a total of L sampling time steps, the sliding window size is M, and the step size is S. Then the data can be divided into H parts, where H is also the number of graphs. The calculation of H is as follows:

[0054]

[0055] By configuring the size of the sliding window, the time-step data of length M is modeled as a single graph, and the features of each node simultaneously include the time features of M time steps.

[0056] Step 1.2: Conversion of longitude and latitude to displacement

[0057] Convert the longitude and latitude coordinates in the geographic coordinate system into the local coordinate system. Let r e represent the equatorial radius of the earth, r p represent the polar radius of the earth, and lat 0 be the latitude in radians of the reference point. With reference to the longitude and latitude reference starting point, the effective radius of the reference point is calculated as:

[0058]

[0059] Thus, the change in longitude and latitude coordinates can be converted into the displacement change in the plane coordinate system of the local coordinate system. Let lat rad and lon rad represent the radian values of the target longitude and latitude, and N and E represent the northward displacement and eastward displacement. The calculation process is as follows:

[0060]

[0061] After the above steps, the true label values for training are obtained. Then, the training data is aligned according to the timestamps and label values, and the dimensions are kept consistent. After dividing the training set and the validation set, the complete training data is formed.

[0062] Step Two: Construct the Trans-GCN model;

[0063] As Figure 1 shown, the Trans-GCN model includes three steps.

[0064] Step 2.1: Feature representation and encoding embedding

[0065] First, it is necessary to construct the Laplacian matrix of the graph to obtain the eigenvectors:

[0066]

[0067] where U ∈ R N×N is the eigenvector matrix of the Laplacian matrix, N is the number of nodes, Λ is the diagonalized eigenvalue matrix, and the columns corresponding to the first m smallest eigenvalues in U are represented as:

[0068] P (i) =(U i,1 ,U i,2 ,...,U i,m )∈R m

[0069] Let λ i represent the i-th smallest eigenvalue, φ i and represent the normalized eigenvector and the standardized eigenvector respectively. Pair the m smallest eigenvalues with their corresponding φ and to obtain the position encoding matrix of node v i . Then, through a linear layer on the 3D position encoding matrix, it is aggregated into a new embedding representation of size k. Finally, it is further simplified through an L2 pooling layer to obtain the final node encoding representation input to the Transformer Encoder.

[0070] For each graph structure, first, by transforming the node features of the basic graph, the d-dimensional node features and edge features are embedded into the graph Transformer layer. The pre-computed encoding information of dimension k is added to the node features through a linear transformation. The specific calculation representation of the entire input embedding is as follows:

[0071]

[0072] where and represent the basic features of the graph nodes, where B 0 ∈R d×k , and a 0 ,b 0 ∈R d are the parameters of the linear projection layer.

[0073] Step 2.2: Attention calculation

[0074] Each node v i and v j will generate Query (Q), Key (K), and Value (V) respectively, where d h is the dimension after the vector is projected into the attention space. The Query of node v i , and the Key of node v j can be used to calculate the attention score, which reflects the similarity of the two vectors.

[0075]

[0076] After the dot product calculation in the high-dimensional space, Scaling scales the dot product result by the dimension d of the Key to handle the numerical stability problem. The formula is as follows: K Scale it according to the following formula:

[0077]

[0078] The attention scores can be controlled within a reasonable range through the Softmax function, avoiding excessive gradient problems in subsequent calculations, enabling the model to be trained stably.

[0079]

[0080] After obtaining the attention weights for each neighbor node weighted sum of the Values of each neighbor node using these weights to obtain a new feature representation of node v i

[0081]

[0082] To continue to deepen feature extraction, multiple different attention heads operate simultaneously. Each head corresponds to different query, key, and value vectors (where H represents the number of heads). The role of these heads is to capture the relationships between nodes from multiple different subspaces. The specific process is as follows:

[0083]

[0084] Step 2.3: Feature Integration and Propagation

[0085] Residual connections can help retain the original input features, prevent over-updating of information, and solve the problem of vanishing gradients in deep networks. Such a setting also ensures the propagation of the original input features in each layer. Layer normalization standardizes the mean and variance of the input to converge faster and stabilize the gradients. The features of each node pass through a two-layer FFN for further non-linear mapping of feature representations, enhancing the expressive power of the model and learning more complex feature patterns. The specific formula is as follows:

[0086]

[0087] To prevent the model from being overly smoothed and affecting the overall prediction effect, the model is improved as follows: By calculating the scaling factor for each layer, that is, the ratio of the current gradient norm to the weight norm, it is ensured that layers with larger gradients have smaller learning rates, while layers with smaller gradients have larger learning rates. This adaptive adjustment helps to stabilize the training process and effectively guides parameter updates. The local learning rate scaling factor is calculated as follows:

[0088] ​

[0089] where η represents the global learning rate, and W (k) represents the weight matrix of the k-th layer, is the gradient of the k-th layer on the current batch of data, and λ represents the weight decay coefficient. The final node feature output is as follows:

[0090]

[0091] The steps of edge feature update are similar to those of node feature update, and also include residual connection and layer normalization. Through these steps, the node features and edge features in Trans-GCN will be continuously updated in each layer, combining the graph structure and edge information to capture the complex relationships between nodes in the graph.

[0092] Step 3: Training and validation of the model

[0093] Continue to train the model designed in Step 2 according to the data processed in Step 1. Set a certain data ratio and the parameters of the neural network model, and start the training of the model. The loss function uses the mean square error loss, and the Adam optimizer is adopted. When the value of the loss function is lower than the set threshold, stop the training and save the model.

[0094] Step 4: When the GNSS signal is briefly interrupted or occluded and drifts, use the trained Trans-GCN model to perform dead reckoning on the mobile robot to complete the position compensation when the absolute position is lost.

[0095] The systems, devices, modules or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions.

[0096] It should also be noted that the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, commodity or device including the element.

[0097] The above embodiments should be understood as being only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the content recorded in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A mobile robot dead reckoning method based on Trans-GCN, characterized in that: The following steps are involved: S1: Collect wheel speed sensor and inertial sensor data of the wheeled mobile robot, and intercept time series data through sliding windows and adaptive sliding steps. In each window data, the sensor data of a single dimension is constructed as a graph node, and a bidirectional edge connection structure is constructed for some nodes to form a complete graph structure data; S2: Build a graph convolutional network (GCN) to extract the dependency relationship of data in the graph in non-Euclidean space, including three steps of feature transfer, aggregation, and update to obtain node features and edge features; S3: Decompose the adjacency matrix of the constructed graph, construct the node position encoding matrix, integrate it through the linear layer and L2 pooling layer, and perform linear transformation with the features extracted in S2 and input it into the Transformer encoder structure; S4: In each layer, the correlation between the target node and the neighboring nodes is calculated to obtain the attention weight. Then, the attention feature is connected to the original feature residual and integrated through the feedforward neural network FFN. Finally, it is input into the fully connected network to obtain the network prediction value. S5: Train the model according to the training sample data, stop the training when the verification loss is less than the threshold, save the model, and then input the wheel speed sensor and inertial sensor data into the model to obtain the dead reckoning result without GNSS signal.

2. A mobile robot dead reckoning method based on Trans-GCN according to claim 1, characterized in that: The step S1 specifically includes: The inertial sensor of the mobile robot body is used to record the time series data of three-axis angular velocity and three-axis acceleration, as well as the two-dimensional data of the left and right wheels recorded by the wheel speed sensor respectively. A sliding window and sliding step of a certain length are set, and the data is subjected to dimensionality reduction processing. The mean and standard deviation of the single-dimensional window data are taken, and the labels of the true values ​​are aligned by timestamp. Then, the single-dimensional data of the training unlabeled data is modeled into nodes, and the edge connections between the nodes are set to generate a record adjacency matrix to construct a complete graph structure training data.

3. The mobile robot dead reckoning method based on Trans-GCN according to claim 1, characterized in that: The position encoding matrix constructed in step S3 contains feature data of three dimensions: after obtaining the feature matrix from the Laplacian matrix of the graph, the normalized feature vectors and the standardized feature vectors corresponding to the first m smallest eigenvalues ​​are selected, and together with the original feature vectors, a three-dimensional initial position encoding matrix is ​​constructed, and then a new embedding representation is obtained by aggregation through a linear layer, and then simplified through an L2 pooling layer to obtain the final position encoding and propagate learning together.

4. The mobile robot dead reckoning method based on Trans-GCN according to claim 3, characterized in that: In step S4, the input data is passed through four fully connected layers to obtain four matrices: Query (Q), Key (K), Value (V), and Edge (E). The feature data is projected into the attention space. The calculation formula of the Attention-score matrix of the kth layer of the hth attention head is: Where i is the target node number, j is the neighbor node, and d K Represents the dimension of the Key vector. The Softmax function calculates each row of the matrix. The formula is as follows: where y a Represents the value of the ath column of a row of the Attention-score matrix, y b represents the value of the bth column of a row of the matrix, and w represents the number of matrix columns.

5. The mobile robot dead reckoning method based on Trans-GCN according to claim 4, characterized in that: After calculating the Attention-score, the residual connection is introduced, and Layernorm standardizes the input data. The features of each node pass through a two-layer feedforward neural network for further nonlinear mapping feature representation. The specific formula is expressed as follows: in is the Layernorm layer, is the self-attention of multi-head splicing, W is the weight matrix, Relu is the activation function, Represents the hidden state after LayerNorm normalization, It is the hidden state after the feedforward neural network and ReLu activation function. Nonlinear transformation is added to enable the model to learn more complex features.

6. The mobile robot dead reckoning method based on Trans-GCN according to claim 5, characterized in that: The S4 connects the attention feature with the original feature residual and integrates it through the feedforward neural network FFN, and finally inputs it into the fully connected network to obtain the network's predicted value, which specifically includes: After two layers of feedforward neural network, an adaptive learning rate adjustment factor is designed, that is, the ratio of the current gradient norm to the weight norm. The local learning rate proportional factor is calculated as follows: Where η represents the global learning rate, is the gradient of the current k-th layer on the current batch of data, λ represents the weight attenuation coefficient, and finally the node feature output after Trans-GCN is: Then, after passing through the fully connected layer, the model’s predicted output is obtained through denormalization.

7. The mobile robot dead reckoning method based on Trans-GCN according to claim 6, characterized in that: In step S5, the training truth value comes from the latitude and longitude data output by the RTK differential positioning system. The latitude and longitude data are converted as follows: the latitude and longitude coordinates in the geographic coordinate system are converted into the local coordinate system. e represents the equatorial radius, r p Represents the polar radius of the Earth, lat0 is the latitude in radians of the reference point, and the effective radius at the reference point is expressed as: r ns 、r ew Used to describe the radius of curvature of the earth in different directions, affecting the calculation of position information: r ns Indicates the radius of curvature in the north-south direction, i.e. the meridian direction, to assist in calculating displacement in the latitude direction; r ew It represents the radius of curvature in the east-west direction, i.e. the direction of latitude, and assists in the calculation of displacement in the longitude direction. Thus, the change of longitude and latitude coordinates can be converted into the displacement change in the plane coordinate system in the local coordinate system, with lat rad and lon rad Indicates the arc value of the target longitude and latitude, N and E represent the north displacement and east displacement, calculated as follows: By aligning the timestamps of the input data after dimensionality reduction with the longitude and latitude timestamps, the model can be trained and feature extracted.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the mobile robot dead reckoning method based on Trans-GCN as claimed in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the mobile robot dead reckoning method based on Trans-GCN is implemented as described in any one of claims 1 to 7.

Images