Mobile robot positioning and navigation method based on AI assistance

Abstract

The invention discloses a mobile robot positioning and navigation method based on AI assistance. The method comprises the following steps: step 1, acquiring an offline map of a predetermined area and a CSI data set of a plurality of point locations in the area; step 2, carrying out clustering denoising on the amplitude of each CSI data in the CSI data set, and then carrying out PCA dimension reduction; step 3, adopting a one-dimensional convolutional neural network based on a residual network to obtain position estimation; step 4, based on the position estimation, performing resampling by adopting an adaptive Monte Carlo positioning algorithm to obtain starting point positioning data with relatively high precision; and 5, fusing odometer data and CSI position estimation, and realizing fusion positioning and navigation based on an extended Kalman filtering algorithm. According to the positioning and navigation method, the positioning precision and robustness can be improved, and low-cost and high-precision positioning is realized.

Description

TECHNICAL FIELD

[0001] The application relates to an AI-assisted mobile robot positioning and navigation method. BACKGROUND

[0002] Traditional indoor positioning methods usually rely on expensive hardware sensors (such as lidar, camera or inertial measurement unit) to provide accurate position information. However, these sensors not only have high cost, but also may be affected by factors such as light and occlusion in complex environments, making it difficult to ensure accuracy.

[0003] In recent years, with the rapid development of artificial intelligence technology, achieving high-precision mobile robot positioning and navigation in indoor environments has become a research topic of concern. Traditional positioning methods usually rely on expensive sensors such as lidar, camera or inertial measurement unit to obtain position information. These sensors can endow robots with the ability to perceive the surrounding environment, navigate space and interact, but each has significant limitations. For example, lidar can accurately measure the distance and shape of objects, but its high cost limits its widespread application; inertial measurement unit has lower cost, but accumulates errors over time and needs to be corrected by external systems; camera has relatively low cost and is easy to deploy on a large scale, but its performance is poor in poor lighting conditions. SUMMARY

[0004] The purpose of the present application is to provide an AI-assisted mobile robot positioning and navigation method that can improve the accuracy and robustness of positioning and achieve low-cost high-precision positioning.

[0005] To achieve the above purpose, the solution of the present application is:

[0006] An AI-assisted mobile robot positioning and navigation method, comprising the following steps:

[0007] Step 1: Obtain an offline map of a predetermined area and a CSI dataset of several points in the area;

[0008] Step 2: Cluster and denoise the amplitude of each CSI data in the CSI dataset, and then perform PCA dimensionality reduction;

[0009] Step 3: Obtain position estimation using a one-dimensional convolutional neural network based on residual network;

[0010] Step 4: Based on the position estimation, use an adaptive Monte Carlo localization algorithm to resample to obtain starting point positioning data;

[0011] Step 5: Fuse odometer data and position estimation to achieve fusion positioning and navigation based on extended Kalman filter algorithm.

[0012] In step 1, the robot carries the WiFi CSI acquisition device to move in the predetermined area, acquires CSI data of each point, and forms a CSI data set.

[0013] In step 2, the specific process of clustering and denoising the amplitude of each CSI data in the CSI data set is as follows.

[0014] The amplitude values of all CSI data in the CSI data set are extracted, and the redundant and extreme amplitude values are removed.

[0015] Based on the processed amplitude values, the Euclidean distance between two points is calculated.

[0016] The Euclidean distance is taken as the neighborhood radius, and all points are traversed. If the number of points in the neighborhood radius of a certain point is greater than or equal to the minimum point threshold, it is taken as a cluster, and all points in the cluster are marked, otherwise the point is marked as a noise point.

[0017] In step 2, the specific process of PCA dimension reduction is as follows.

[0018] The matrix composed of the CSI data after clustering and denoising is centered, and then the covariance matrix is calculated.

[0019] The eigenvalue and eigenvector of the covariance matrix are obtained by eigenvalue decomposition.

[0020] The eigenvalues are sorted by size, and the eigenvectors corresponding to the first k largest eigenvalues are used to construct a projection matrix.

[0021] In step 3, the one-dimensional convolutional neural network based on the residual network sequentially includes a first convolutional layer, a maximum pooling layer, a first residual block, a second residual block, a second convolutional layer, an application Dropout layer, a fully connected layer, and an output layer from input to output. The first convolutional layer uses 64 convolutional kernels, each with a size of 7x7 and a step of 1. The size of the maximum pooling layer is 2x2. The second convolutional layer uses 32 convolutional kernels, each with a size of 7x1.

[0022] The first residual block and the second residual block have the same structure and each include two convolutional layers. The output of each convolutional layer is also connected to batch normalization and ReLU activation function. Each convolutional layer uses 128 convolutional kernels, each with a size of 3x3.

[0023] The loss function of the one-dimensional convolutional neural network based on the residual network is as follows.

[0024]

[0025] wherein, is a true value, is a predicted value, n is a sample number;

[0026] According to the error calculated by the loss function, the weight and bias of the neural network are adjusted by using the following weight update formula,

[0027]

[0028] wherein, is the current weight, is the learning rate, is the partial derivative of the loss function with respect to the weight.

[0029] The specific process of the above step 5 is,

[0030] Step 51, according to the odometer data, the state transition equation of the extended Kalman filter algorithm is used to predict the position information of the robot at the next time;

[0031] Step 52, according to the observation data of the global position information, the state estimation of the extended Kalman filter algorithm is updated to obtain the Kalman gain;

[0032] Step 53, according to the Kalman gain, the predicted state obtained in step 51 is updated;

[0033] Step 54, according to the accurate position and direction estimation obtained, the navigation instruction is obtained to control the movement of the robot.

[0034] After using the above scheme, the present application effectively processes Wi-Fi channel state information (Channel State Information, CSI) and odometer (Odometry) data by collecting the deep learning model in artificial intelligence technology, and realizes low-cost and high-precision positioning. The present application introduces a one-dimensional convolutional neural network (1D-CNN) to automatically extract and learn the features of the CSI data, and combines the extended Kalman filter algorithm (EKF) for positioning error correction. By taking the CSI pre-positioning coordinates as the observation variable and combining them with the real-time collected odometer data of the robot, the AI model can intelligently analyze and estimate the robot trajectory, greatly improving the positioning accuracy and robustness. Compared with the traditional positioning method relying on expensive sensors, the AI auxiliary system significantly reduces the hardware cost, and through the self-adaptive learning ability of deep learning, improves the positioning performance in complex indoor environments. The present application also compares a plurality of filtering algorithms through experiments, verifies the effectiveness and feasibility of the AI auxiliary algorithm, and shows that the system has strong adaptability and accuracy in dynamic and complex environments. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1is a data collection point distribution map;

[0036] Figure 2 is a CSI preprocessing comparison chart;

[0037] Wherein, (a) is before preprocessing, (b) is after preprocessing;

[0038] Figure 3 is a flowchart of the present application;

[0039] Figure 4 is a CSI-AMCL resampled particle distribution comparison chart;

[0040] Wherein, (a) is resampling without CSI, (b) is resampling distribution after adding;

[0041] Figure 5 is a path map after long-distance driving. DETAILED DESCRIPTION

[0042] The technical solutions and beneficial effects of the present application will be described in detail below in combination with the drawings.

[0043] The present application provides an AI-assisted mobile robot positioning and navigation method, comprising the following steps:

[0044] Step one: CSI data collection

[0045] In the data collection process, first use the CSI collection device to collect the point position of the offline map, which is used for subsequent offline fingerprint map making. Taking the total area of data collection as 100 square meters as an example, which includes 100 reference points (RP) and 20 test points (TP). The data collection plan layout is shown in Figure 1 .

[0046] Figure 1 The fingerprint reference point and test point map of shows the distribution of collected RP and TP about CSI. The reason for selecting to collect RP and TP as reference set is that the reference set can be further divided into training set and test set for deep learning. In the process of offline fingerprint map generation, the collected CSI data will be stored in a notebook computer, and the CSI feature information of each position will be extracted through the subsequent processing algorithm.

[0047] After the offline map is collected, the robot carries the WiFi CSI collection device to move in the offline map area. The collected CSI data is preprocessed and matched with the offline map to obtain the most probable offline map RP point, and the required CSI positioning coordinates are obtained. At the same time, the odometer sensor on the robot is responsible for recording the speed and direction data of its current position in real time. The odometer data is used to supplement the global positioning information provided by the Wi-Fi CSI, and combined with these information to provide accurate motion model data input for the subsequent extended Kalman filter (EKF) algorithm.

[0048] Step two: CSI data preprocessing based on clustering algorithm

[0049] First, this embodiment needs to extract the amplitude in the CSI, and then remove the redundant same amplitude and extreme amplitude. The extracted CSI amplitude value is first calculated by the Euclidean distance between two points as in formula (1), and the obtained Euclidean distance becomes the neighborhood radius (epsilon) of each point. After traversing the data, each point is processed. For each unvisited point, if the number of points in its neighborhood is greater than or equal to the minimum point threshold minPts, it is expanded into a cluster, and all points in the cluster are marked. If the number of points in the neighborhood is less than minPts, the point is marked as a noise point. Finally, the function returns the denoised data subset.

[0050] (1)

[0051] As shown in Figure 2 , taking the reference point RP42 with coordinates (4, 2) as an example, the clustering denoising can clearly see that the amplitude value of the CSI data has a clear expression. After clustering and denoising of each RP, a clear and understandable amplitude value can be obtained, which has a profound effect on subsequent fingerprint positioning. Because the collected CSI data has 253 subcarriers, i.e. 253-dimensional data, such a large data is obviously not conducive to machine learning training and data processing, therefore, in this embodiment, principal component analysis (PCA) is performed on the data to reduce the dimension. First, the CSI data matrix after clustering and denoising is centered, that is, the mean value of each feature is subtracted. Assuming that there are , where n is the number of samples and p is the number of features. Subtract the mean value from each column (feature):

[0052] (2)

[0053] In formula (2), is the mean vector of each column. Calculate the covariance matrix of the centered data, and the formula for calculating the covariance matrix is:

[0054] (3)

[0055] In formula (3), is the covariance matrix between features. The eigenvalue decomposition of the covariance matrix is obtained, and the eigenvalue and eigenvector are obtained:

[0056] (4)

[0057] In formula (4), is the eigenvector matrix of the covariance matrix, is the corresponding eigenvalue diagonal matrix. The eigenvector represents the basis vector of the new space, and the eigenvalue represents the variance of the data along these basis vectors. According to the size of the eigenvalue, the eigenvectors corresponding to the top k largest eigenvalues are selected to form a new projection matrix , where k is the target dimension after dimensionality reduction, and the original data is projected onto the selected principal components to obtain the data after dimensionality reduction. Through this linear transformation, PCA reduces the data dimension while maintaining as much original data variance as possible, making data processing and subsequent neural network training more efficient. In this embodiment, the original CSI data is reduced from 253 dimensions to a 50x51 matrix through PCA, significantly improving training and positioning efficiency.

[0058] Step three: CSI positioning algorithm based on residual network 1D-CNN

[0059] This embodiment constructs a one-dimensional convolutional neural network (1D CNN) based on the residual network (Residual Network, ResNet) style. The core feature of this model is the use of residual blocks. The main advantage of residual blocks is to help solve the gradient vanishing problem in deep networks. By introducing direct connections that skip some layers (called jump connections or shortcut connections), the gradient can flow directly through these connections without being hindered by intermediate layers. Each residual block includes two layers of convolution, each followed by batch normalization (Batch Normalization) and ReLU activation function, and then the input is added before the last activation function through the shortcut connection. This model design is a variant of CNN, which enhances the support of deep networks by using residual blocks, especially suitable for handling complex or deep network architecture problems. This design helps to accurately learn and predict the coordinates of each position in the indoor environment, thereby improving the accuracy of positioning. Batch normalization is a key technology in AI models that can speed up network training and stabilize model performance. The following formula explains how to standardize after each layer of convolution:

[0060] (5)

[0061] (6)

[0062] In formula (5)-(6), is the standardized value (mean 0, variance 1) of the output of the kth neuron, is the output of the kth neuron, and are the mean and standard deviation of the small batch data, respectively, is a very small constant, usually 10 -8 ; is the output of the batch normalization layer, is the scaling factor (trainable parameter), is the offset (trainable parameter); batch normalization helps to maintain the stability of the input of each layer in the network, thereby improving the training efficiency of the model. It can be explained that the AI-assisted model solves the instability problem caused by the change of input data distribution through this mechanism. In order to reflect the self-learning ability of AI, the formulas of loss function and backpropagation algorithm are increased, and the loss function usually uses mean square error (MSE) to measure the difference between the predicted result and the true value:

[0063] (7)

[0064] In formula (7), is the true value, is the predicted value, and n is the number of samples. Through the backpropagation algorithm, the AI model adjusts the weights and biases of the network according to the error calculated by the loss function. The weight update formula of backpropagation is

[0065] (8)

[0066] In formula (8), is the current weight, is the learning rate, is the partial derivative of the loss function with respect to the weight. The AI model optimizes the network weights through repeated iterations, thereby continuously reducing the error and improving the accuracy of positioning. In order to prevent model overfitting, Dropout regularization technology is used in the AI-assisted model. Dropout randomly discards part of the neurons to make the model more generalizable. Its expression is:

[0067] (9)

[0068] In formula (9), is the output of the discarded neuron, and p is the dropout probability. This mechanism can be explained as AI prevents model overfitting to training data by randomly discarding part of the neurons, thereby improving the prediction ability of unknown data.

[0069] As shown in Table 1, the input layer of the model is the preprocessed CSI signal, which is 51x1, representing the single-channel amplitude data of 51 subcarriers. First, the input data is passed through a convolutional layer, which uses 64 convolutional kernels, each of size 7x7 with a stride of 1, to extract local features from the input data. The output dimension of the convolutional operation can be calculated by the formula:

[0070] (10)

[0071] (11)

[0072] where H and W are the height and width of the input, and are the height and width of the convolutional kernel, is the stride, and the output size after convolution is 51x64. Next, a 2x2 max-pooling layer is used to downsample the feature map, reducing the data dimension, and the output size becomes 25x64. Subsequently, the network enters two residual blocks, each composed of two convolutional layers, each using 128 3x3 convolutional kernels. The residual blocks use skip connections to add the input directly to the output, alleviating the problem of gradient vanishing, and the formula for the residual connection is:

[0073] (12)

[0074] where represents the nonlinear transformation after the convolutional layer, batch normalization, and ReLU activation function. Through this skip connection, the model can better learn deep features, and the output size of the residual block remains 25x128. Next, a convolutional layer with 32 7x1 convolutional kernels is used to further extract deep features, and the output size after convolution is 1x19x32. After the convolutional operation, a Dropout layer is applied with a dropout rate of 0.3 to randomly discard 30% of the neurons, preventing the model from overfitting. The formula for the Dropout layer is:

[0075] (13)

[0076] where is the output of the neuron, is the dropout probability. Then, the model enters a fully connected layer, which takes the flattened one-dimensional vector of the convolutional features as input and outputs a 256-dimensional feature representation. The fully connected layer maps the input to the output through a linear transformation, and the formula is:

[0077] (14)

[0078] In formula (14), is a weight matrix, is an input, is a bias term, is an output. Finally, the model gives a 25-dimensional positioning result through the output layer, representing the final coordinates or classification output. The output layer usually uses SoftMax or linear activation function, depending on the application requirements. Through this structure design, 1D-CNN can effectively extract spatial features from Wi-Fi CSI data, solve indoor positioning problems, and have significant advantages in accuracy.

[0079] Table 1

[0080]

[0081] In the experiment, the 1D-CNN network model showed significant advantages in positioning accuracy. The experimental results showed that the network could achieve about 60% accuracy within a positioning error range of 5 meters, and almost all errors within 15 meters were within the acceptable range. Compared with traditional support vector machine (SVM), k-nearest neighbor algorithm (KNN), deep neural network (DNN) and weighted k-nearest neighbor algorithm (WKNN), the 1D-CNN model had the smallest positioning error and the highest positioning accuracy. Therefore, the model showed extremely superior performance in indoor positioning scenarios.

[0082] Step four: high-precision robot initialization positioning algorithm based on AMCL-CSI

[0083] In this step, the adaptive Monte Carlo localization (AMCL) is combined with CSI pre-positioning to perform high-precision robot initialization positioning. Adaptive Monte Carlo localization uses particle filtering to realize robot positioning in the environment, using odometry data and laser radar or other sensors to estimate the probability distribution of robot position. On this basis, combined with the CSI pre-positioning coordinates, the accuracy and speed of initialization positioning can be improved.

[0084] In adaptive Monte Carlo localization algorithms, the traditional initialization step involves uniformly distributing particles across the entire map, which leads to a long convergence time. By combining adaptive Monte Carlo localization with CSI localization, the initial position information provided by CSI is used to initialize the particles within the range estimated by CSI, reducing the particle search space and improving localization efficiency. The particle distribution in adaptive Monte Carlo localization gradually converges, and the robot estimates the final localization result through a weighted average of the particles. The global position information from CSI provides the basis for the initial particle distribution, while adaptive Monte Carlo localization further refines the position estimation through sensor fusion, achieving the goal of precise localization.

[0085] (15)

[0086] In equation (15), It is a particle Location coordinates, It's the orientation of the particles. This is an estimate based on CSI pre-positioning error, representing the uncertainty in position estimation. In this embodiment, CSI pre-positioning is used as the observation for particle weight update, and the weight update formula is:

[0087] (16)

[0088] In equation (16), It is the weight of i particles. The Euclidean distance between the position of the i-th particle and the predetermined position of the CSI is defined as:

[0089] (17)

[0090] In equation (17), These are the coordinates of the i-th particle. The robot's position coordinates are obtained through CSI pre-positioning. The standard deviation parameter in the weight calculation determines the sensitivity of particle weights to distance changes. To ensure the sum of the weights is 1, the weights of all particles need to be normalized. The normalization formula is:

[0091] (18)

[0092] According to the normalized weights, the particles are resampled, and the state estimation of the robot is updated. By calculating the Euclidean distance of the particles using the CSI pre-localization information and updating the weights of the particles accordingly, the positioning accuracy and convergence speed of the adaptive Monte Carlo localization algorithm can be significantly improved. The particle weight update and normalization process in the formula ensure that the particle distribution after each update is closer to the true position of the robot, effectively integrating adaptive Monte Carlo localization with CSI pre-localization.

[0093] Figure 4 The contrastive diagram of CSI-AMCL resampled particle distribution is shown. As can be seen from (a)(b), after adding CSI, the positioning accuracy has been significantly improved.

[0094] Step five: EKF-based fusion positioning and navigation

[0095] In this step, the odometer data and Wi-Fi CSI pre-localization coordinates are fused based on extended Kalman filter, so as to realize the accurate positioning and navigation of the mobile robot. As a powerful filter, extended Kalman filter can combine local motion estimation (from odometer) and global position information (from CSI pre-localization) for dynamic state prediction and update, ensuring the navigation accuracy of the robot in complex environments.

[0096] 1. State prediction

[0097] At each time, the odometer data (robot speed and angular velocity) is used to predict the position information of the robot at the next time, including position and direction. This step is realized through the state transition equation of extended Kalman filter:

[0098] (19)

[0099] In formula (19), is the control input (such as linear speed v and angular speed w in odometer) represents the position information of the robot;

[0100] (20)

[0101] In formula (20), and predicted robot position coordinates, predicted robot direction angle. In addition, the state covariance matrix will also be updated based on the process noise prediction:

[0102] (21)

[0103] where is the Jacobian matrix of the state transition matrix.

[0104] 2. Observation Update

[0105] Whenever the robot receives global position information provided by the Wi-Fi channel state information, the state estimation of the extended Kalman filter is updated with these observation data. The observation equation is:

[0106] (22)

[0107] In equation (22), x^k is the global position coordinate of the robot obtained through CSI positioning, is the observation noise. Through the observation data, the Kalman gain of the state estimation is updated , and the calculation formula is:

[0108] (23)

[0109] 3. State Update

[0110] Using the Kalman gain, the predicted state is updated to correct the estimation of the robot's position and orientation

[0111] (24)

[0112] In equation (24), x^k is the global position coordinate of the robot obtained through CSI positioning, represents the observation residual, i.e., the difference between the CSI positioning data and the predicted state.

[0113] 4. Covariance Update

[0114] The state covariance matrix is updated to reflect the uncertainty of the current state estimation:

[0115] (25)

[0116] Based on the fused accurate position and orientation estimation, the robot can generate navigation instructions to ensure its movement according to the predetermined path. At the same time, the extended Kalman filter will continuously predict and update during the movement of the robot to ensure the accuracy of positioning and navigation.

[0117] Figure 5 ​The path comparison experiment in the figure shows the effect of the two positioning methods using only odometer data and odometer combined with CSI. The experimental results clearly show the performance difference of each method, and highlight the advantage of CSI global correction. The black solid line and the circle in the experimental figure indicate the ideal reference path (Actual path), which is the path that the robot should follow. However, the positioning result relying only on the odometer shows a significant deviation, especially at the connection of the turning and straight line segment, the cumulative error is obvious, the path of the black solid line and the square mark (Odom) deviates greatly from the actual path. This reflects that the odometer data is prone to significant cumulative error in long-term use, resulting in a decrease in positioning accuracy. In contrast, the EKF-CSI method combined with CSI shows obvious advantages. The path indicated by the black solid line and the cross mark (EKF-CSI) is very close to the actual path, and the error is significantly reduced.

[0118] This embodiment evaluates the performance of five filtering algorithms, including: extended Kalman filter algorithm combined with CSI signal, unscented Kalman filter (UKF) algorithm, positioning algorithm relying only on odometer, Kalman filter algorithm and CSI fingerprint positioning algorithm. Through the comparison of the positioning result path, the extended Kalman filter algorithm combined with CSI signal has the highest positioning accuracy and the smallest path error, showing excellent positioning stability and accuracy. The method using only odometer is simple but has large cumulative error, which needs additional sensor data correction. The standard Kalman filter effect is slightly better than the single odometer data, but the error is still obvious. The CSI fingerprint positioning method shows large error and instability, which is not suitable for high-precision applications.

[0119] Further analysis of the positioning error of each algorithm is shown in Table 2. The Kalman filter combined with CSI has the smallest error, proving its significant effect in improving accuracy and error control. The method relying only on the odometer performs well in the short term, but is inferior to the extended Kalman filter and CSI combination. The CSI fingerprint method has large cumulative error and the worst accuracy and stability. In summary, the extended Kalman filter algorithm combined with CSI signal has obvious advantages in the positioning of mobile robots, can significantly improve the positioning accuracy and effectively control the error accumulation, and is suitable for high-precision navigation requirements in complex environments.

[0120] Table 2 Comparison of positioning error of algorithms

[0121] The application provides an AI-assisted indoor positioning and navigation method for mobile robots, which intelligently fuses CSI signals and odometry data by combining extended Kalman filtering and CSI-AMCL strategy without using high-cost sensors, and fully utilizes the advantages of AI in feature extraction and data fusion. This AI-based solution significantly improves the accuracy and stability of positioning, especially when dealing with the cumulative errors caused by the influence of CSI signals on environmental changes and the long-term use of odometry. By automatically extracting features from CSI data through a deep learning model and intelligently fusing and dynamically adjusting data with AI, the system can accurately predict the robot's trajectory in complex environments and update the navigation path based on real-time data. Experiments show that this algorithm is superior to traditional unscented Kalman filtering, standard Kalman filtering, and simple Wi-Fi fingerprint positioning methods in terms of reducing error accumulation and improving path tracking accuracy. The experimental results further prove that, compared with the odometry calculation method alone, the positioning accuracy of the AI-assisted filtering algorithm is improved by 50% in the path navigation process, and compared with the Wi-Fi fingerprint positioning algorithm, the accuracy is improved by 66.67%. In comparison with various filtering algorithms, the accuracy is improved by 55.56% compared with the standard Kalman filtering algorithm, and by 77.33% compared with the unscented Kalman filtering algorithm. This AI-assisted positioning and navigation method not only has significant advantages in accuracy, but also can achieve efficient and stable positioning in complex dynamic environments, and has high application value and promotion prospects.

[0122] Those skilled in the art will appreciate that embodiments of the application can be supplied as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of complete hardware embodiments, complete software embodiments, or embodiments combining software and hardware aspects. Moreover, the application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) having computer-usable program code embodied thereon. The solutions in the embodiments of the application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0123] The application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a means for implementing the functions described in the flowcharts and / or block diagrams.Figure 1 one or more processes and / or functions specified in the block or blocks. Figure 1

[0124] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the Figure 1 one or more processes and / or functions specified in the block or blocks. Figure 1 one or more processes and / or functions specified in the block or blocks.

[0125] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 one or more processes and / or functions specified in the block or blocks. Figure 1 one or more processes and / or functions specified in the block or blocks.

[0126] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to encompass within their scope all such variations and modifications as are included within the scope of the invention.

[0127] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims, the application can be practiced otherwise than as specifically described herein.​

Claims

1. An AI-assisted mobile robot positioning and navigation method, characterized in that The method comprises the following steps: Step 1, obtaining an offline map of a predetermined area and a CSI data set of a plurality of points in the area; Step 2, clustering and denoising the amplitude of each CSI data in the CSI data set, and then performing PCA dimension reduction; Step 3, obtaining position estimation by using a one-dimensional convolutional neural network based on a residual network; Step 4, based on the position estimation, resampling is performed by using an adaptive Monte Carlo localization algorithm to obtain starting point positioning data; Step 5, fusing the odometer data and the position estimation, and realizing fusion positioning and navigation based on an extended Kalman filtering algorithm.

2. The AI-assisted mobile robot positioning and navigation method of claim 1, wherein: In the step 1, the robot carries a WiFi CSI acquisition device to move in the predetermined area to obtain the CSI data of each potential point to form the CSI data set.

3. The AI-assisted mobile robot positioning and navigation method of claim 1, wherein: In the step 2, the specific process of clustering and denoising the amplitude of each CSI data in the CSI data set is as follows: extracting the amplitude values of all CSI data in the CSI data set, and eliminating the redundant same amplitude values and extreme amplitude values; based on the processed amplitude values, calculating the Euclidean distance between two points; taking the Euclidean distance as the neighborhood radius, traversing all points, if the number of points in the neighborhood radius of a point is greater than or equal to the minimum point threshold, the point is taken as a cluster, and all points in the cluster are marked, otherwise the point is marked as a noise point.

4. The AI-assisted mobile robot positioning and navigation method of claim 1, wherein: In the step 2, the specific process of PCA dimension reduction is as follows: performing a centering operation on the matrix composed of the CSI data after clustering and denoising, and then calculating a covariance matrix; performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues and eigenvectors; sorting the eigenvalues according to the size, and constructing a projection matrix from the eigenvectors corresponding to the first k largest eigenvalues.

5. The AI-assisted mobile robot positioning and navigation method of claim 1, wherein: In the step 3, the one-dimensional convolutional neural network based on the residual network comprises, from input to output, a first convolutional layer, a maximum pooling layer, a first residual block, a second residual block, a second convolutional layer, an application Dropout layer, a full connection layer, and an output layer, wherein the first convolutional layer adopts 64 convolutional kernels, each convolutional kernel has a size of 7x7 and a step of 1; the maximum pooling layer has a size of 2x2; the second convolutional layer adopts 32 convolutional kernels, each convolutional kernel has a size of 7x1.

6. The AI-assisted mobile robot positioning and navigation method of claim 5, wherein: The first residual block and the second residual block have the same structure and each contain two convolutional layers, the output of each convolutional layer is further connected to batch normalization and ReLU activation function, and each convolutional layer adopts 128 convolutional kernels, each convolutional kernel has a size of 3x3.

7. The AI-assisted mobile robot positioning and navigation method of claim 5, wherein: The loss function of the one-dimensional convolutional neural network based on the residual network is as follows: ， wherein, is the true value, is the predicted value, n is the number of samples; according to the error calculated by the loss function, the weight and bias of the neural network are adjusted by using the following weight update formula, ， where, is the current weight, is the learning rate, is the partial derivative of the loss function with respect to the weight.

8. The AI-assisted mobile robot positioning and navigation method of claim 1, wherein: the specific process of the step 5 is as follows: Step 51, predicting the position information of the robot at the next time according to the state transition equation of the extended Kalman filtering algorithm based on the odometer data; Step 52, updating the state estimation of the extended Kalman filtering algorithm according to the observation data of the global position information to obtain the Kalman gain; Step 53, updating the predicted state obtained in the step 51 according to the Kalman gain; Step 54, according to the obtained accurate position and direction estimation, obtain navigation instructions, control the movement of the robot.

Images