RBCN-Net: A Data-Driven Inertial Navigation Algorithm for Pedestrians

Abstract

Pedestrian inertial navigation technology plays an important role in indoor positioning technology. However, low-cost inertial sensors in smart devices are affected by bias and noise, resulting in rapidly increasing and accumulating errors when integrating double acceleration to obtain displacement. The data-driven class of pedestrian inertial navigation algorithms can reduce sensor bias and noise in IMU data by learning motion-related features through deep neural networks. Inspired by the RoNIN algorithm, this paper proposes a data-driven class algorithm, RBCN-Net. Firstly, the algorithm adds NAM and CBAM attention modules to the residual network ResNet18 to enhance the learning ability of the network for channel and spatial features. Adding the BiLSTM module can enhance the network’s ability to learn over long distances. Secondly, we construct a dataset VOIMU containing IMU data and ground truth trajectories based on visual inertial odometry (total distance of 18.53 km and total time of 5.65 h). Finally, the present algorithm is compared with CNN, LSTM, ResNet18 and ResNet50 networks in VOIMU dataset for experiments. The experimental results show that the RMSE values of RBCN-Net are reduced by 6.906, 2.726, 1.495 and 0.677, respectively, compared with the above networks, proving that the algorithm effectively improves the accuracy of pedestrian navigation.

1. Introduction

A successful outdoor positioning system based on the global positioning system (GPS) has been developed. However, satellite signals are weakened indoors, such as underground and inside large buildings, making GPS useless for positioning. Indoor navigation and positioning services can be made available to users through indoor location-based services, which can boost pedestrian travel efficiency and make it easier for managers to prioritize pedestrian traffic.

The demand for indoor positioning is increasing as personalized networks become more popular. Indoor positioning technologies rely on various technologies, such as Bluetooth [1,2], WiFi [3], machine vision [4], ultra-wide band (UWB) [5] and inertial navigation [6]. The first four technologies are highly precise and relatively mature. However, each of these technologies has its own challenges to consider. Bluetooth and WiFi can be affected by electromagnetic interference and indoor obstacles. Machine vision needs to take into account user privacy concerns. UBW requires advanced facility deployment. Finally, inertial navigation can be challenged by drift during extended operation. Despite these limitations, no single positioning technology can provide a stable high-accuracy positioning service. Multisensor fusion technology is currently under development to address this challenge. In the complex indoor environment, high-precision inertial navigation can provide a reliable positioning service for pedestrians for a short period of time when other positioning technologies are temporarily disabled by electromagnetic interference. Additionally, it can provide correction information for other positioning technologies because inertial navigation is not disturbed by external factors and does not rely on any external signals. Therefore, high-precision pedestrian inertial navigation has a high research value.

The development of microelectro mechanical technology has led to the inclusion of inertial measurement units (IMUs) based on these systems in smart devices such as mobile phones, watches and wristbands. The use of IMU sensors in these devices for inertial navigation is very convenient and does not require the purchase of additional expensive equipment. However, compared to platform-based inertial guidance, IMU sensors in smart devices have several limitations due to their lower cost and other factors. Specifically, they generally have lower accuracy and higher noise levels. Conventional techniques frequently use double-integration methods to solve for position coordinates based on acceleration data, in which positioning errors caused by sensor measurement errors can quickly increase and accumulate, in addition to the direct solution process itself generating some errors. How to effectively eliminate the cumulative error of positioning has been a hot research topic in the field of pedestrian inertial navigation positioning. Better localization accuracy can be attained using the Kalman filter and zero velocity update (ZUPT) algorithm together [7], but these methods require pedestrians to fix IMUs on their feet, which is not compatible with the usage scenarios of most smart devices. In addition, these methods require high measurement accuracy of IMU, resulting in the requirement of more expensive sensors [8], which has kept such algorithms in the laboratory stage.

Data-driven methods have been recognized in recent years as the most effective means to suppress error drift by transforming the inertial tracking problem into a data sequence learning problem. The algorithms simultaneously collect IMU data and ground truth trajectory data over a short period to regress motion parameters (e.g., velocity vector and heading) and are proven to outperform traditional heuristics in terms of navigation accuracy and robustness. Among them, RIDI [9] made a breakthrough in coordinate system normalization by using SVM to regress the walking velocity vector and achieve the first accurate estimation of pedestrian velocity in a real indoor environment. Influenced by the generalization ability of deep neural networks, RoNIN [10] proposed a new concept of motion model-based RIDI and constructed multiple neural network models to successfully predict high-precision trajectories. To reduce the cumulative error caused by continuous double integration, IONet [11] partitioned the IMU data into independent windows and used long short term memory (LSTM) [12] algorithm to regress the rate of change of pedestrian speed and heading for each window. TLio [13] merged relative state measurements of neural network outputs into a stochastic cloning Kalman filter to further reduce pedestrian attitude, velocity and sensor bias. To solve the problem of inaccurate smartphone pose estimation, IDOL [14] proposed a two-stage neural network framework that uses a recurrent neural network (RNN) to estimate device pose in the first stage and a combination of RNN and extended Kalman filter to estimate device position in the second stage. Wang [15] combined neural networks and ZUPT to recognize pedestrian walking patterns using multihead convolutional neural networks, and adaptively adjusted the step detection threshold based on the walking pattern recognition results.

However, the existing data-driven methods still suffer from the problem of low navigation accuracy due to the simplicity of their network models. Taking inspiration from RoNIN, we have designed a new deep neural network model, called RBCN-Net (see Figure 1), to improve the ability of the network to regress pedestrian motion features. We achieve this by combining ResNet18 [16], bidirectional long short term memory (BiLSTM) [17], a convolutional block attention module (CBAM) [18], and a normalization-based attention module (NAM) [19]. By adding the attention mechanism, we address the problem of the network’s inability to distinguish the importance of features. To demonstrate the effectiveness of RBCN-Net, we conduct comparative experiments by constructing multiple network models using the VOIMU dataset to verify its ability to improve navigation accuracy.

Figure 1. Overview of the method.

The main contributions of this work are summarized as follows:

(1)

To enhance the network’s ability to learn channel and spatial features, the NAM is added to each Basicblock of ResNet18, and the CBAM is added after the maximum pooling layer of ResNet18. Meanwhile, in order to enhance the network’s capability in long-range learning modeling, we add a BiLSTM network after the 8th Basicblock-NAM structure.

(2)

A dataset called VOIMU which contains 80 sets of IMU data of various walking routes with a total path length of 18.53 km and a time of 5.65 h. The ground truth trajectories are based on ocular inertial odometry. We will share all the datasets and codes for further research.

The rest of the paper is organized as follows: Section 2 presents the network model; Section 3 presents the training data collection method, the data preprocessing steps, and the network model implementation details; Section 4 presents the experimental results and analysis; finally, Section 5 concludes the paper.

2. Description of the Algorithm

To solve the network performance degradation problem caused by the difficulty of convergence of training gradient dispersion as the depth of the network model increases in convolutional neural network (CNN) [20], ResNet [16] appends constant mapping to CNN to change the learning objective to the residual $H\left(x\right)-x$ of the output and input features, making the output features contain more information than the input features. However, ResNet is unable to distinguish the importance between feature channels after feature extraction. As an efficient and lightweight attention mechanism, a normalization-based attention module (NAM) [19] does not require additional fully connected [21] and convolutional layers and instead highlights salient features using the variance metric between input feature channels.

2.1. Improved ResNet-NAM

Indoor positioning scenarios are characterized by a small amount of sensor data and high real-time requirements. Therefore, we choose ResNet18, which is a smaller network with the advantages of fast training speed and low risk of overfitting. ResNet18-1D version includes 8 Basicblock structures, and each Basicblock structure includes two 3×1 convolutional modules with the same number of channels. To distinguish the feature importance, this algorithm adds the NAM module after the 2nd convolution module of all the Basicblock structures. The Basicblock-NAM structure is shown in Figure 2.

Figure 2. Basicblock-NAM structure.

The NAM calculates the variances between channels of the input feature vector f using the scale factor $\gamma$ in the normalized BN. The variances are multiplied with the weight ${\omega }_i$ corresponding to each scale factor ${\gamma }_i$, and then the weights factor $M_{NAM}$ is obtained by the sigmoid activation function. The number of $M_{NAM}$ is the same as the number of feature channels, and the numerical magnitude reflects the importance of different feature channels. Finally, the output features are obtained by multiplying $M_{NAM}$ with the input feature vector element by element, and the input and output feature sizes remain constant. The weights are calculated as follows:

$$M_{NAM}\left(f\right)=sigmoid\left({\omega }_{\gamma }\left(BN\left(f\right)\right)\right)$$

(1)

In Equation (1), $\omega$ is the proportion of the scaling factor ${\gamma }_i$ among all scaling factors, f is the input feature vector of the NAM, and BN( ) is the inter-channel variance calculated as follows:

$$BN\left(f\right)=\gamma \frac{f-{\mu }_f}{\sqrt{{\sigma }_f^2+ϵ}}+\beta$$

(2)

In Equation (2), ${\mu }_f$ and ${\sigma }_f$ are the mean and standard deviation of the small batch f, $ϵ$ and $\beta$ are trainable affine transformation parameters (scale and displacement). The improved ResNet18-NAM input and output features are related as:

$$y=F(x,\left\{W_i\right\})+x$$

(3)

In Equation (3), x and y are the input and output features of ResNet18-NAM, and $F(x,\left\{W_i\right\})$ is the residual mapping function. If the dimensions of the input features x and $F(x,\left\{W_i\right\})$ are the same, the values of the corresponding channels are added directly; otherwise, linear mapping is required to match them.

2.2. BiLSTM Structure

The inertial data collected by IMU devices follow a time-series pattern due to the highly repetitive nature of pedestrian pace. Santiago Cortés et al. [22] considered only CNN networks and ignored the time-series nature of the data. To fully learn the temporal correlation of inertial data features, enhance the long-range learning performance of deep neural networks and achieve more nonlinear transformations, a bidirectional long short-term memory (BiLSTM) network is added after the ResNet18-NAM in our algorithm. BiLSTM is a variant of LSTM [12] network, consisting of forward LSTM and backward LSTM stacking, which not only has the long-range sequence learning capability of LSTM model but also can better explore the backward and forward dependencies of inertial data. To obtain the BiLSTM output $H_t$, the hidden state $\overrightarrow{h_t}$ of the forward LSTM output and the hidden state $\overleftarrow{h_t}$ of the backward LSTM output are connected. $H_t$ is calculated as follows:

$$H_t=W_{\overrightarrow{hf}}\overrightarrow{h_t}+W_{\overleftarrow{hf}}\overleftarrow{h_t}+b_h$$

(4)

In Equation (4), $W_{\overrightarrow{hf}}$ is the connection weight matrix from the forward LSTM to the output layer; $W_{\overleftarrow{hf}}$ is the connection weight matrix from the backward LSTM to the output layer; $b_h$ is the bias of the output layer.

3. Pedestrian Inertial Navigation Algorithm Based on RBCN-Net

3.1. Framework of the Study

The algorithm is a data-driven supervised learning method [23] for pedestrian inertial navigation [24]. It regresses the pedestrian velocity vector in the 2D plane using ground truth trajectories and IMU data, which is based on the pedestrian motion modeling concept proposed by RoNIN. Three key steps are included: (1) collecting the required training data; (2) converting the collected IMU data to a unified coordinate framework; (3) constructing a deep neural network to learn the relationship between data features and motion features for regression of the velocity vector.

3.2. Data Acquisition

In this paper, ground truth trajectories are captured by visual–inertial SLAM [25], which uses a Google Tango phone with motion-tracking technology [26] to capture the 3D poses of pedestrians. With sufficient light and clear vision, the position measurement error after 10 min of pedestrians walking is less than 0.3 m, which can provide low-drift pseudo-ground truth trajectory. For IMU data acquisition, we implement an IMU data acquisition application based on the Huawei Mate 20 phone to record data from the phone’s built-in acceleration, gyroscope and gravity sensors. When collecting data, the pedestrian is allowed to manipulate the phone naturally, trying to imitate the daily use habits of the phone, including texting, calling, shaking naturally in a handshake and putting it in the pocket of clothes. The Google Tango is strapped to the chest area because the pedestrian’s chest area sways much less than the extremities when walking. The device is worn in the way shown in Figure 3. In this paper, a total of five healthy young people were selected as testers and the actual campus environment was used as the experimental environment. Each person walked naturally for 3 to 10 min according to the data collection requirements, and a total of 80 sets of experimental data of different walking routes were obtained, which were divided into training set, test set, and validation set according to the ratio of 6:3:1. The total path length of the dataset VOIMU was 18.53 km, and it took 5.65 h.

Figure 3. Diagram of the equipment of the two cell phones: Google Tango is strapped to the chest area and the IMU phone (Huawei Mate 20) is used normally, including entering text, shaking it handheld, making phone calls and placing in the pocket of clothes.

3.3. IMU Coordinate System Normalization

Pedestrians use their phones in a variety of ways, making the phone’s built-in IMU sensor coordinate system unstable. In extreme cases, the coordinate system of each frame of collected IMU data has different tilt angles relative to the ground, resulting in inconsistent motion representation and making the feature learning of the network very difficult. A stable IMU data frame will make the neural network regression task easier. Therefore, we use the coordinate system normalization method proposed by RIDI [9] to convert all acceleration and gyroscope ${[a,w]}_i$ to a coordinate system aligned with the direction of gravity.

In this paper, a Gaussian smoothing [27] with a standard deviation of 0.2 is applied to the bit-pose data provided by Google Tango to suppress the high frequency. we assume that the experimental walking ground is very flat, i.e., vertical displacement is zero, and the gesture data are converted to the velocity components $(v_x,v_y)$ in the two-dimensional plane. A normal human gait cycle lasts approximately one second [28], and the sampling frequency of the IMU data acquisition application designed for this experiment is 200 Hz, so when the sliding window mechanism is used to slice ${[a,w]}_i^n$, the size of the sliding window is 200 frames of data, and the time length of the divided IMU sequence ${[a,w]}_{i=1\sim 200}^n$ is approximately one gait cycle. This is more conducive to the subsequent network training and learning. The label for the ${[a,w]}_{i=1\sim 200}^n$ is the corresponding pedestrian real velocity vector $(v_x,v_y)$ collected by Google Tango.

3.4. RBCN-Net

To improve the mapping capability of IMU sequence ${[a,w]}_{i=1\sim 200}^n$ to labels $(v_x,v_y)$, further reduce the interference of non-critical information, and make the network focus on more effective information, we modify the ResNet18-NAM 1D framework by adding the CBAM [18] after the maximum pooling layer. The CBAM is formed by serially connecting two attention submodules, CAM and SAM, to enhance the network’s ability to learn channel and spatial features, respectively. The network with the addition of CBAM still has limited capability in long-range learning modeling, so the BiLSTM network is added after the 8th Basicblock-NAM structure in this paper. BiLSTM can realize the network’s feature learning for long-range sequences as well as fully learn the temporal correlation of IMU sequences. The framework of the proposed RBCN-Net is shown in Figure 4.

Figure 4. RBCN-Net Architectures.

The network’s input IMU sequence is ${[a,w]}_{i=1\sim 200}^n$ whose dimension is 6 × 200.The input sequence is first subjected to preliminary feature extraction and dimensionality reduction using a 7 × 1 convolutional kernel and a 3 × 1 maximum pooling layer to obtain a 64 × 50 feature vector $f_1$.

$$f_1=MaxPool\left(f^{7\times 1}\left({[a,w]}_{i=1\sim 200}^n\right)\right)$$

(5)

Then $f_1$ is fed into the CBAM. CAM compresses $f_1$ based on maximum pooling and average pooling for length and width, respectively, to produce two new feature sequences, which are fed into a two-layer perceptron to produce two new feature vectors, which are activated by the Sigmoid function to produce the module’s final output weight coefficients, $M_{cam}$.

$$M_{cam}\left(f_1\right)=\sigma (MLP\left(AvgPool\left(f_1\right)\right)+MLP\left(MaxPool\left(f_1\right)\right))$$

(6)

Finally, the output feature vector $f_{cam}$ of CAM is obtained by multiplying the weight $M_{cam}$ with $f_1$ element by element.

$$f_{cam}=M_{cam}\left(f_1\right)⨀f_1$$

(7)

The SAM takes $f_{cam}$ as input and first compresses it using maximum pooling and average pooling. After compression, the two new feature sequences obtained are spliced by corresponding channels and further dimensional compression is performed using a 7 × 1 convolution, and the final compressed features are activated by the sigmoid function to obtain the spatial attention weights $M_{sam}$.

$$M_{sam}\left(f_{cam}\right)=\sigma \left(f^{7\times 1}[AvgPool\left(f_{cam}\right);MaxPool\left(f_{cam}\right)]\right)$$

(8)

The weights $M_{sam}$ are multiplied element by element with $f_{cam}$ to obtain the output of the SAM as well as the output feature vector $f_{CBAM}$ of CBAM. The dimensionality of $f_{CBAM}$ is the same as that of $f_1$.

$$f_{CBAM}=M_{sam}\left(f_{cam}\right)⨀f_{cam}$$

(9)

After passing 8 Basicblock-NAM, the $f_{CBAM}$ obtains 512 × 7 feature vectors, and then input to the BiLSTM layer, which has two hidden layers with input and output dimensions of seven. The fully connected layer connects the 512 × 14 feature vectors output from the BiLSTM layer into the final feature vector with the same dimension as the velocity label, and the physical meaning represents the pedestrian’s velocity vector in the navigation coordinate system. In the fully connected layer, the dropout layer parameter of 0.5 is used to prevent overfitting. The mathematical expression of the proposed RBCN-Net is shown in Equation (10).

$$\tilde{v}({\tilde{v}}_x,{\tilde{v}}_y)=f_{\theta }\left({[a,w]}_{i=1\sim 200}^n\right)$$

(10)

4. Experimental Results and Analysis

4.1. Evaluation Indicators

The velocity vector $\tilde{v}({\tilde{v}}_x,{\tilde{v}}_y)$ predicted by the RBCN-Net is converted to the estimated coordinates ${\tilde{p}}_i({\tilde{x}}_i,{\tilde{y}}_i)$, and the resulting trajectory is the estimated trajectory. As described in the previous section, real trajectories are generated from Google Tango data, and data frame i corresponds to real coordinates as $p_i(x_i,y_i)$. The evaluation benchmarks include root mean square error (RMSE) and average positioning accuracy (RTE). The RMSE provides a visual representation of the global agreement between the estimated and true trajectories [29], and the calculation formula is shown in Equation (11).

$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n[{(x_i-{\tilde{x}}_i)}^2+{(y_i-{\tilde{y}}_i)}^2]}$$

(11)

In Equation (11), n represents the total number of data frames. The RTE is calculated by adding the difference between ${\tilde{p}}_i({\tilde{x}}_i,{\tilde{y}}_i)$ and $p_i(x_i,y_i)$ at the same time and dividing it by the true track length. The calculation formula is shown in Equation (12).

$$RTE=\frac{{\sum }_{i=1}^n\sqrt{[{(x_i-{\tilde{x}}_i)}^2+{(y_i-{\tilde{y}}_i)}^2]}}{n*trajectory\_length}$$

(12)

4.2. Experimental Environment and Hyperparameter Configuration

All studies in this paper were conducted on a laptop with a GPU configuration of GTX 1650. Because the training and testing results of the neural network model are strongly influenced by the parameters, so the learning rate is set to 0.0001 and the loss function of the model is compared every 20 epochs during the training process. The learning rate is reduced by a factor of 0.1 if the validation loss is not reduced. The batch size is set to 128 and the epochs are set to 1200. The training is completed in about 6 hours. The mean square error (MSE) loss function is used. It should be noted that the network will not converge if the initial learning rate for model training is set too large, causing the loss function value to oscillate.

4.3. Analysis of Validation Set Results

The evaluation metrics obtained for the validation set on the final trained RBCN-Net are shown in Table 1. The total trajectory length of the validation set is 1968 m, the average RMSE is 1.912, and the RTE values are all around 1%. It indicates that the model has excellent generalization performance on the validation set, and the error rate is not high. Meanwhile, the error of the validation set itself is relatively stable, with little difference between the same indicator parameters, because after normalization of the IMU coordinate system, the way the pedestrian manipulates the phone during data collection has little impact on the accuracy of navigation. The RMSE of NO.3 is higher than the other validation sets because the magnitude of the RMSE value is inversely proportional to the length of the trajectory.

Table 1. Evaluation metrics of the validation set on the model.

Validation Set Number	Trajectory Length/m	RMSE/m	RTE
NO.1	189	1.617	1.057%
NO.2	156	1.405	1.005%
NO.3	402	2.549	0.84%
NO.4	362	2.766	1.034%
NO.5	112	2.336	2.346%
NO.6	164	0.934	0.691%
NO.7	245	1.726	1.329%
NO.8	338	1.918	1.514%

4.4. Analysis of Test Set Results

In VOIMU, the test set contains twenty-four sets of experiments; the three groups with the best results and the three groups with the worst results are excluded, and nine experimental groups were randomly selected from the remaining test set for index evaluation. Table 2 displays the RBCN-Net evaluation indexes for the nine randomly selected experiments. The trajectory lengths of the nine groups of experiments totaled 1749 m, and the average RMSE and RTE were 2.07 and 1.45%, respectively. It indicates that the errors between the estimated trajectories and the real trajectories are small, all around the meter level, and the trajectory shapes match well. The RMSE and RTE of NO.2 are only 0.763 and 0.561%, which is a difficult positioning accuracy to be achieved by traditional inertial navigation methods [30].

Table 2. Evaluation indexes of part of the test set.

Test Set Number	Trajectory Length/m	RMSE/m	RTE
NO.1	221	1.031	0.603%
NO.2	176	0.761	0.561%
NO.3	206	1.551	0.937%
NO.4	292	4.035	1.439%
NO.5	128	2.304	2.281%
NO.6	67	1.766	2.96%
NO.7	324	3.446	1.167%
NO.8	131	2.015	1.963%
NO.9	204	1.728	1.057%

One more set of experiments was randomly selected among the nine sets for detailed data elaboration. The trajectory length of this experiment is 221.65 m and the time taken is 212.2 s. The raw magnet data of this experiment is shown in Figure 5. With each peak change, the magnetometer data’s magnitude varies in a pattern like a wave, and this corresponds to a pedestrian inflection event. In the case of short travel time, the pedestrian attitude obtained by solving the IMU data using the conventional attitude differential equation is closer to the actual attitude [31]. However, as the time increases, the solved heading angle error gradually becomes larger, which eventually makes the solved trajectory have a large heading angle deviation from the actual walking trajectory. In this paper, we use the RBCN-Net to fit the IMU data with the labels $(v_x,v_y)$ without considering the actual attitude of the phones with IMU. The set of experimental IMU data predicted by the RBCN-Net for the velocity vector $\tilde{v}({\tilde{v}}_x,{\tilde{v}}_y)$ versus the label $(v_x,v_y)$ is shown in Figure 6. The complete trajectory comparison is shown in Figure 7.

Figure 5. The raw magnet data of this experiment. (a) X-axis magnetometer data, (b) Y-axis magnetometer data, (c) Z-axis magnetometer data.

Figure 6. Velocity vector comparison. (a) ${\tilde{v}}_x$ versus $v_x$, (b) ${\tilde{v}}_y$ versus $v_y$.

Figure 7. Comparison of the predicted trajectory and the real trajectorty.

In Figure 7, the starting coordinates are both (0, 0), and the estimated and true end coordinates are ${\tilde{p}}_{end}$(−91.32, 48.81) and $p_{end}$(−93.09, 49.93), respectively, with an error of 2.24 m between the end coordinates. According to Newton’s first law of motion, the dual integration of the acceleration data is the trajectory length in the ideal case. However, in practice, sensor errors in IMU are unavoidable, and using the direct integration algorithm magnifies the errors exponentially. The estimated and true trajectory lengths of this experiment are 219.08 m and 221.65 m, respectively, with only 2.57 m difference. It indicates that the RBCN-Net completes the learning of dynamic and static error features in the acceleration data. There are some deviations between trajectories in region 1 of Figure 7, however, the trajectories in region 2 almost overlap with small errors, and the deviations are corrected rather than cumulative. The magnitude of the distance error between ${\tilde{p}}_i({\tilde{x}}_i,{\tilde{y}}_i)$ and $p_i(x_i,y_i)$ at the same time t is shown in Figure 8, where the distance error oscillates between $[0,2.5]$, indicating that the RBCN-Net effectively eliminates the inertial navigation system’s cumulative error. This experiment has an RMSE of 1.031 and an RTE of 0.603%.

Figure 8. Distance error of coordinate points under simultaneous inscription.

4.5. Comparison of Base Network Models

The effects of different networks on pedestrian inertial navigation systems are investigated to validate the superiority of the proposed RBCN-Net. The four base models, CNN, LSTM, ResNet18 and ResNet50, are intended to be used in comparison experiments with the proposed RBCN-Net on the dataset VOIMU presented in this paper. The measured evaluation metrics are shown in Table 3. The RBCN-Net suggested in this research outperformed.

Table 3. Position evaluation of five competing networks.

Network Model	Average RMSE/m	Average RTE
CNN	8.985	7.804%
LSTM	4.805	3.46%
ResNet18	3.574	2.567%
ResNet50	2.756	1.894%
RBCN-Net	2.079	1.481%

The other four networks with the best assessment metrics, an average RMSE of 2.079 and an average RTE value of 1.481%. ResNet50 and ResNet18 are second to RBCN-Net because the residual structure can effectively address the issue of network degradation. The LSTM network comes next. Although LSTM helps RNN’s long-term dependence issue to some level, it can still be challenging for longer-sequence data. The experimental outcomes are the poorest because CNN is not very well suited for learning time series.

A randomly selected set of experiments in the test set was used for visual trajectory comparison, which had a total trajectory length of 176.42 m, took 160.2 s, and contained long straight lines and seven right-angle turns. The trajectory comparison is shown in Figure 9. The RBCN-Net predicted trajectory has the greatest overlap with the true trajectory. In the zoomed-in area, the pedestrian makes a second turn, and only this model predicts the route close to the true trajectory.

Figure 9. Comparison of five competing networks trajectory and the real trajectory.

The specific evaluation indicators are shown in Table 4. RBCN-Net obtained an RMSE of 0.761, RTE value of 0.56%, and the endpoint coordinates differed by only 1.14 m in this experiment. RBCN-Net’s navigation accuracy is much better than other networks’. The effect is only second to the RBCN-Net is ResNet50, whose endpoint coordinates differ from the real coordinates by 9.52 m, and the RMSE value is 2.313, which is somewhat different from the true trajectory. ResNet18 and LSTM do not differ much from each other in this experiment.

Table 4. Specific evaluation indicators.

Network Model	Trajectory Length/m	RMSE/m	RTE	End Coordinates
CNN	149.69	7.5	6.01%	(2.92,−58.26)
LSTM	172.68	3.295	2.86%	(3.25,−72.83)
ResNet18	174.54	3.228	2.77%	(3.35,−74.32)
ResNet50	170.69	2.313	1.54%	(−1.55,−74.53)
RBCN-Net	180.54	0.761	0.56%	(−7.34,−80.27)
Ground truth	176.42			(−6.12,−80.17)

4.6. Analysis of Ablation Experiments

To evaluate the effectiveness of CBAM, ResNet-NAM and BiLSTM for feature extraction, ablation experiments are conducted, and networks are designed to remove each module separately. The evaluation metrics of each model are measured on the previous test set, and the comparison results are shown in Table 5. Incorporating the NAM attention mechanism into the ResNet18 BasicBlock structure reduces the RMSE and RTE values of the ResNet-NAM model by 0.2 and 0.15, respectively. Adding CBAM after the maximum pooling layer reduces the RMSE and RTE values by 0.27 and 0.2, respectively. The above results confirm the effectiveness of the interactive assignment of feature weights by CBAM and NAM. In the proposed algorithm, the CBAM outperforms the NAM in terms of improving navigation accuracy. The addition of the BiLSTM layer to ResNet18 reduces the RMSE value by 0.93 and the RTE value by 0.68. This indicates that the BiLSTM is more important for the learning ability of time-related features than the attention mechanism. The addition of NAM and CBAM attention mechanisms to ResNet-LSTM reduces the RMSE value of 0.24, RTE value of 0.2, RMSE value of 0.39 and RTE value of 0.35, respectively. The RBCN-Net augments the ResNet-LSTM with both attention mechanisms, and the RMSE value is as low as 2.079 and the RTE value is as low as 1.481%. RBCN-Net reduces the RMSE value of 0.6 and the RTE value of 0.44 compared to ResNet-LSTM. RBCN-Net is more accurate for inertial data mapping compared with other models.

Table 5. Comparison of results of ablation experiments.

Network Model	Average RMSE/m	Average RTE
ResNet18	3.574	2.567%
ResNet18-NAM	3.364	2.41%
CBAM-ResNet	3.301	2.379%
ResNet18-LSTM	2.645	1.887%
ResNet18-NAM-LSTM	2.402	1.701%
CBAM-ResNet-LSTM	2.248	1.537%
RBCN-Net	2.079	1.481%

5. Conclusions and Future Work

The data-driven class of inertial pedestrian navigation techniques faces the issue of low navigation accuracy brought on by straightforward network models. To solve this problem, the novel deep neural network RBCN-Net is proposed. On the dataset VOIMU used in this paper, networks including CNN, LSTM, ResNet18, ResNet50 and others are built and evaluated against RBCN-Net. The navigation accuracy is greatly increased with RBCN-Net, which also roughly matches the real trajectory.

The shortcoming of this paper is that the obtained trajectories are trajectories in two-dimensional planes, and are not adapted to indoor maps. Some unforeseen pedestrian behaviors, such as collisions, emergency turns, etc., cannot be handled by it. As a result, we will continue to develop our work to provide precise indoor localization on any mobile device, regardless of measurement units, users, or settings.