Pedestrian positioning method based on cam, sarb and lstm network

Abstract

The application relates to the technical field of neural networks, in particular to a pedestrian positioning method based on a CAM, a SARB and an LSTM network, which comprises the following steps: collecting a footstep vibration signal; collecting a pedestrian video image through a camera and extracting a pedestrian track; labeling the footstep vibration signal; expanding data samples by adopting a sliding window segmentation method and labeling each data segment with pedestrian track data labels extracted from a video; constructing a neural network model based on a CAM module, an input module, a SARB module and an LSTM neural network model, training the neural network model by using the expanded data set, and calculating a pedestrian track; and constructing a navigation error evaluation benchmark. The application processes sensor data by constructing a deep neural network model, reduces manual participation, improves positioning accuracy, establishes a grade classification rule of navigation error, and measures positioning accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks, in particular to a pedestrian positioning method based on CAM, SARB and LSTM networks. BACKGROUND

[0002] In recent years, with the rapid development of intelligent buildings, the demand for indoor personnel position information is gradually increasing. In large shopping malls, hospitals, airports and other places, it is necessary to provide indoor navigation services and crowd information. In addition, indoor positioning technology also helps to optimize building energy management systems and improve building energy efficiency. However, since the satellite signal received indoors is weak, outdoor positioning technology cannot effectively position indoors, so indoor positioning technology is still a research hotspot.

[0003] Step vibration positioning technology is a technology that uses human gait vibration signals for indoor positioning; this technology mainly uses the characteristics of human gait vibration signals propagating indoors, by arranging a certain number of vibration sensors indoors, after collecting human gait vibration signals, using signal processing technology for analysis and processing, to realize the positioning of people in the indoor location. The traditional method is a positioning algorithm based on wavelet decomposition and TDOA, which requires a large amount of manual participation to select appropriate signal features to extract step events, and the positioning accuracy is not ideal.

[0004] Deep neural networks are a kind of multi-layer supervised learning neural networks, which show good results in processing time series data. For example, a target trajectory recognition method based on residual network and attention mechanism, which uses LSTM to learn time-related features in sequence data, and residual convolutional network to learn local and spatial features; however, using residual convolutional network and LSTM to extract features through splicing to fuse different types of features, there are feature incompatibility, information redundancy, and performance degradation; on the other hand, it also increases the computational complexity, resulting in large memory occupation. SUMMARY

[0005] In view of the shortcomings of the existing method, the present application processes sensor data by constructing a deep neural network model, reduces manual participation, improves positioning accuracy, establishes a hierarchical division rule of navigation error, and measures positioning accuracy.

[0006] The technical scheme adopted by the present application is: a pedestrian positioning method based on CAM, SARB and LSTM networks includes the following steps:

[0007] Step one, collect step vibration signals;

[0008] Further, the step vibration signal is collected by a geophone to collect a voltage signal and amplify it, and then through analog-to-digital conversion, data with a preset sampling frequency is obtained.

[0009] Further, the footstep vibration signal is preprocessed by low-pass filtering and linear normalization.

[0010] Step two, collect pedestrian video images through the camera, and extract pedestrian trajectory;

[0011] Further, step two specifically includes:

[0012] Step 21, lay a checkerboard carpet on the ground, take pictures of the checkerboard from different angles and heights, obtain the camera intrinsic parameters and the extrinsic parameters of the camera installation position, and establish a coordinate transformation model from the image coordinate system to the world coordinate system;

[0013] Step 22, select the midpoint between the heel key points of the two feet as the pedestrian foot landing point, and extract the pedestrian foot landing point from the video frame;

[0014] Step 23, sequentially connect the foot landing points of each frame to obtain an initial trajectory, and then perform coordinate conversion through the coordinate conversion model to obtain the spatial trajectory coordinate information of the pedestrian with the ground plane as the two-dimensional plane XY axis and the vertical ground as the Z axis.

[0015] Step three, label the footstep vibration signal;

[0016] Further, step three specifically includes:

[0017] Step 31, traverse each frame timestamp of the video, find the timestamp of the vibration signal, and mark the real coordinate position;

[0018] Step 32, for the vibration signal of the pedestrian position coordinate with unmatched timestamp, use uniform interpolation to mark.

[0019] Step 33, use the instantaneous speed of each frame of the vibration signal to represent the pedestrian motion trajectory;

[0020] Further, the formula of the instantaneous speed is:

[0021]

[0022]

[0023] wherein, V x and V y are the speeds in x and y directions respectively, ts is the timestamp set, x[i] and y[i] are the coordinate position sets of the i-th frame of the trajectory respectively.

[0024] Step four, expand the data samples after labeling by using the sliding window segmentation method, and label each data segment with the pedestrian trajectory data extracted from the video as a label;

[0025] Step 5: Construct a neural network model based on the CAM module, input module, SARB module, and LSTM; train the neural network model using the expanded dataset; and calculate the pedestrian trajectory.

[0026] Furthermore, step five specifically includes:

[0027] Step 51: The CAM module performs max pooling and average pooling operations on the width and height of the vibration signal, respectively; and feeds it into a two-layer MLP network; then generates weights through the Sigmoid activation function, multiplies the input data and weights to obtain the first feature sequence; then uses a 7×7 convolution kernel and max pooling layer to extract features and reduce the dimensionality of the first feature sequence to obtain the second feature sequence.

[0028] Step 52: The second feature sequence is processed by the input module and 4 sets of 2*SARB modules to obtain the third feature sequence;

[0029] Furthermore, the SARB module contains two 3×1 convolutional layers, one self-attention layer, and one residual connection.

[0030] Step 53: Input the third feature sequence into the bidirectional LSTM network to obtain the fourth feature sequence, and then output the predicted pedestrian velocity vector through the output layer and the fully connected layer.

[0031] Step 54: Calculate the loss function between the predicted pedestrian speed and the actual pedestrian speed, and stop iterating when the maximum iteration is satisfied or the loss function satisfies the preset threshold.

[0032] Furthermore, the formula for calculating pedestrian trajectories is as follows:

[0033]

[0034]

[0035] in, It is the pedestrian's initial position. pedestrian t i The position of time, t i Let i be the timestamp of the i-th frame. and Let x be the speed of the pedestrian in the x-axis and y-axis directions in the i-th frame.

[0036] Step 6: Establish a navigation error assessment benchmark;

[0037] Further, the navigation error evaluation benchmark is classified according to the value size of ATE and RTE; if the weighted average value of ATE and RTE is less than 1, the navigation accuracy is considered excellent; if the weighted average value of ATE and RTE is greater than 1 and less than 3, the navigation accuracy is considered to be qualified; and if the weighted average value of ATE and RTE is greater than 3, the navigation is considered to fail.

[0038] Advantages of the present application:

[0039] 1、The one-dimensional residual module of the CAM+SARB+LSTM deep learning neural network extracts a feature vector, performs feature learning on the footstep vibration signal, provides a one-dimensional residual module by introducing the idea of residual learning to solve the training difficulty and performance degradation problem of the deep residual network, and can extract rich features while overcoming the problem of incompatible features, greatly reducing the amount of calculation and improving the generalization ability of the model.

[0040] 2、Compared with the TDOA positioning algorithm, no additional sensor equipment needs to be added, which is convenient and easy to promote, and has the advantages of high precision and low cost.

[0041] 3、The deep learning neural network is applied to indoor positioning technology, an accurate mathematical model is established to predict the indoor position of a person, and the indoor positioning precision based on the footstep vibration signal is significantly improved, and the positioning time is shortened. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 It is a pedestrian positioning method flow chart based on CAM, SARB and LSTM network of the present application;

[0043] Figure 2 It is a footstep vibration signal positioning method block diagram based on a deep neural network;

[0044] Figure 3 It is a vibration signal collection schematic diagram of a person to be positioned in an indoor area;

[0045] Figure 4 It is a schematic diagram of the overall framework of the deep neural network model;

[0046] Figure 5 It is a vibration signal collection system structure diagram;

[0047] Figure 6 It is based on a self-attention residual module;

[0048] Figure 7 It is a schematic diagram based on a self-attention mechanism;

[0049] Figure 8 It is a schematic diagram of a channel attention mechanism;

[0050] Figure 9 is a comparison result figure of the partial real trajectory and the predicted trajectory;

[0051] Figure 10 is a positioning evaluation result of 28 paths in the test set. DETAILED DESCRIPTION

[0052] The application will be further described below in connection with the drawings and embodiments, which are simplified schematic diagrams and only schematically show the basic structure of the application, and thus only show the components related to the application.

[0053] As shown in Figure 1 ,a pedestrian positioning method based on CAM, SARB and LSTM network includes the following steps: 2 Step one, collection and preprocessing of foot vibration signals;

[0054] Collection of foot vibration signals: in this embodiment, 5 healthy young people are selected as subjects, the average age of the subjects is 22±5 years old, the height is 171±15 cm, and the weight is 70±16 kg. A 50 square meter laboratory is selected as the experimental environment.

[0055] The person to be positioned does not need to carry equipment and walks in the area in a normal posture. The millivolt-level weak voltage generated by the seismometer is transmitted to the MCC118 analog-to-digital conversion module through the AD620 operational amplifier to sample at a sampling frequency f s The sampling frequency is 3000 Hz; the collected data includes the vibration sensor data of six channels at each sampling time point and the corresponding sampling time sequence,

[0056] is an experimental collection schematic diagram; the data recorded by the indoor natural motion equipment of a plurality of persons to be positioned is collected, and the data is saved in the internal storage space of the Raspberry Pi. The vibration signal collection system structure diagram is shown in Figure 3 . Figure 5

[0057] Preprocessing of foot vibration signals: the original data collected by the Raspberry Pi is preprocessed, and the preprocessing includes low-pass filtering and linear normalization.

[0058] The vibration frequency caused by the steps of pedestrians is generally less than 100 Hz, so a Butterworth low-pass filter is used to remove high-frequency noise.

[0059] In order to improve the convergence speed and generalization ability of the model, linear normalization processing is performed on the denoised signal, and the formula is:

[0060]

[0061] Wherein, x represents original data, x_norm represents normalized data, min(x) represents the minimum value in data x, and max(x) represents the maximum value in data x.

[0062] Step two, indoor pedestrian trajectory extraction;

[0063] Camera calibration: install a camera at a suitable position in the room, lay a checkerboard carpet on the ground, take pictures of the checkerboard from different angles and heights, obtain the camera intrinsic parameters and the extrinsic parameters of the camera installation position, establish a coordinate transformation model from the image coordinate system to the world coordinate system, and transform the relationship into:

[0064]

[0065] Wherein, Z c represents the camera coordinates, [u, v] represents the image pixel coordinates, [X w , Y w , Z w ] represents the world coordinates of the pedestrian target, R represents the camera rotation matrix, T represents the camera translation matrix, and M c represents the camera intrinsic matrix.

[0066] The experimenter walks naturally according to the data acquisition requirements of the application, the camera records the walking video, obtains 560 groups of vibration signals and video files of different walking routes, as shown in Figure 3 .

[0067] Pedestrian foot point extraction: in this embodiment, a pedestrian pose detection technology is used to detect and extract pedestrian key points to construct a human skeleton graph, wherein the pedestrian pose detection technology can use the openpose library; the midpoint between the key points of the two feet is selected as the pedestrian foot point, and the pedestrian foot point extraction is performed on each frame of the recorded video to obtain the pedestrian foot point positioning result in the image coordinate system.

[0068] Pedestrian trajectory extraction: sequentially connect the foot points of each frame to obtain an initial trajectory, first use a Kalman filter to smooth the image trajectory to obtain an optimal trajectory, and then perform coordinate conversion calculation through the established coordinate conversion model to obtain the spatial trajectory coordinate information of the pedestrian with the ground plane as the two-dimensional plane XY axis and the vertical ground as the Z axis.

[0069] Step three, vibration signal data labeling;

[0070] Step 31, traverse each frame timestamp of the video, find the corresponding timestamp in the vibration signal data, and mark the real coordinate position corresponding to the timestamp in the vibration signal data;

[0071] Step 32, for the vibration signal that fails to directly match the video timestamp marked pedestrian position coordinates, uniform interpolation can be used with known coordinate positions to mark, complete the position labeling of each frame of the foot vibration signal.

[0072] Step 33, calculate the instantaneous speed of each frame of the vibration signal to describe the motion trajectory of the pedestrian, complete the pedestrian speed labeling of the vibration signal; by calculating the position difference of the current frame and the previous two hundred frames, and dividing by the time difference to get the average speed, the pedestrian speed in x and y directions is represented as:

[0073]

[0074]

[0075] where, V x and V y are the speeds in x and y directions, ts is the timestamp set, x and y are the sets of trajectory coordinate positions.

[0076] The data labeled after the completion of one experiment is merged and packaged into a foot vibration signal data set, and multiple vibration data sets are constructed into a foot vibration signal positioning database.

[0077] Step four, adopt sliding window segmentation method to expand the data samples completed labeling, and give each data segment the pedestrian trajectory data extracted from the video as a label;

[0078] The pedestrian trajectory data includes the pedestrian position coordinates and speed of the vibration point.

[0079] Assuming that the original data is X, the data length is L, the data is divided into sub-data windows with a length of L window , the step length of the sliding window is L interval , then n sub-data windows are obtained after segmentation, and the data of each sub-window is:

[0080]

[0081] where, represents the i-th sliding window, represents the j-th element of the data window; n represents the number of sub-windows after segmentation, and the calculation method is:

[0082]

[0083] The segmented sub-data window is labeled with speed, and multiple sub-data windows form a batch for training.

[0084] Step five, construct a CAM+SARB+LSTM neural network model, and train the neural network model using the expanded data set;

[0085] As Figure 4 Construction of network model: the network model includes channel attention module (CAM), input module (INPUT), self-attention residual module (SARB), LSTM network and output module.

[0086] The data sequence [a, g] n as an input data of the input module, the dimension of the input data sequence is 1500x6; the CAM module extracts features from the input data as shown in Figure 8 ; the data is input into the CAM module, and the input data is subjected to maximum pooling and average pooling operations based on width and height respectively to obtain two new feature sequences; the two sequences are input into a two-layer MLP network to obtain a feature vector; the feature vector is subjected to a Sigmoid activation function to generate a weight, and the input data and the weight are multiplied to obtain a feature sequence Feat1; a 7x7 convolution kernel and a maximum pooling layer are used to extract features and reduce dimensions of Feat1 to obtain a 64x50 feature sequence Feat2.

[0087] As Figure 4 , the data is input into a deep residual module composed of 4 groups of 2*SARB modules, and the SARB module structure is shown in Figure 6 and 7 ; the residual convolution and the self-attention mechanism are combined to form the SARB module, and the self-attention mechanism of the SARB module is used to capture the long-term dependence relationship in the sequence data, while the residual convolution is used to extract features and transform the sequence data; this combination can deeply mine the feature information in the data and has better generalization ability; the SARB module includes two 3x1 convolution layers, a self-attention layer and a residual connection; in the convolution layer, a one-dimensional convolution with a convolution kernel of 3 is used for convolution operation, and the convolution result is normalized and activated; in the self-attention layer, the weight of each position in the feature map is calculated, and the self-attention result is normalized; in the residual connection, a convolution layer and a normalization layer are used to convert the channel number of the input feature map to the same channel number as the output feature map of the convolution layer, so as to perform addition operation; finally, the outputs of the convolution layer, the self-attention layer and the residual connection are added to obtain the output of the self-attention residual block, and a feature sequence Feat3 is obtained.

[0088] The vibration signal data belongs to a time sequence, and bidirectional LSTM can better explore the effective information in the past and the future. The feature sequence Feat3 is input into an LSTM layer, the LSTM layer is a bidirectional LSTM, that is, it is composed of two LSTM cycle layers with opposite information transmission, the input and output dimensions of the LSTM layer are 63, and it has 2 hidden layers. At this time, the feature sequence Feat3 becomes a feature sequence Feat4 with 1024 channels and 126x1. Finally, the feature sequence Feat4 is output through a full connection layer to obtain a 2D output vector. The physical meaning of the 2D output vector represents the speed of the pedestrian in the navigation coordinate system, and the complete pedestrian walking trajectory is calculated.

[0089] The output layer is a full connection layer with a dimension of 2, and the output is a 2D prediction vector.

[0090] The loss function can reflect the difference between the predicted value and the label value of the model. The root mean square error function is used as the loss function, and the loss function is used for back propagation to update the parameters of the neural network.

[0091] The root mean square error function represents the square root of the sum of the distances between the target variable and the predicted value, and is the most commonly used regression loss function. In this embodiment, the root mean square error function is used as the loss function, and the CAM+SARB+LSTM neural network learns and predicts every 1500 frames of the preprocessed [a, g] n The sequence is 1 span, and the 1th span is calculated. The root mean square error between the output 2D vector and the label label(V x , V y ) measured by the motion capture system is obtained, and the loss function can be calculated by formula (7):

[0092]

[0093] Among them, is the average speed of the neural network regression in the x-axis direction at time i, is the average speed of the neural network regression in the y-axis direction at time i, and are the speeds of the i-th pedestrian in the x-axis and y-axis directions.

[0094] Model training: the maximum number of training steps is set to 2000 epochs, and in each training period, the prediction output of the network model and the corresponding loss function are calculated by forward propagation. The weights of the model are optimized by the back propagation algorithm and the Adam optimizer.

[0095] In the training process, the loss function of the network model is compared every 20 epochs, and when the loss function converges within a preset threshold sigma, the training of the network model is terminated, the loss function value is saved, and the performance of the model is evaluated.

[0096] After training, enter the test phase, input the test data set into the network model, output the pedestrian speed on the two-dimensional plane, and calculate the pedestrian trajectory, the formula is:

[0097]

[0098]

[0099] Wherein, is the position of the pedestrian at the initial time, is the position of the pedestrian at the t i time, t i is the timestamp of the i-th frame, and are the speeds of the i-th pedestrian in the x-axis and y-axis directions, respectively, j=2, 3…i.

[0100] The application can be built by various open source deep learning frameworks, and in the embodiment, a neural network framework is built by using PyTorch.

[0101] Step six, establish a navigation error evaluation benchmark;

[0102] Specifically includes:

[0103] Error evaluation: define the video tracking pedestrian trajectory to generate a real trajectory, and the real position of the i-th data frame corresponding to the person to be positioned is The speed vector generated by the network model is predicted, and the trajectory calculated is the trajectory to be evaluated, and the position of the i-th frame corresponding to the person to be positioned is The evaluation benchmark includes absolute trajectory error (ATE) and relative trajectory error (RTE); the absolute trajectory error calculates the difference between the predicted position and the real position, which can very intuitively reflect the global consistency of the trajectory to be evaluated and the real trajectory; the absolute trajectory error is defined as the total root mean square error between the real trajectory and the trajectory to be evaluated, and the absolute trajectory error is as shown in the formula:

[0104]

[0105] In the formula, n is the total number of frames of the data set, and represent the predicted position coordinates of the i-th data frame of the person to be positioned; x i and y i represent the real position coordinates of the i-th data frame of the person to be positioned.

[0106] Relative pose error is used to calculate the root mean square error over a fixed time interval. This involves calculating the change in position between the actual and predicted positions over a given time interval, and then averaging the relative errors over all time intervals to obtain the relative position error. This standard is suitable for estimating system drift. The relative trajectory error is shown in the following equation:

[0107]

[0108] In the formula, m is the number of fixed time intervals in the test set.

[0109] In this embodiment, the fixed time interval is defined as 30 seconds. For sequences shorter than 30 seconds, the position error of the last frame is scaled proportionally.

[0110] Navigation accuracy assessment: Navigation accuracy is graded based on the numerical values ​​of ATE and RTE; if the weighted average of ATE and RTE is less than 1, the navigation accuracy is considered excellent; if the weighted average of ATE and RTE is greater than 1 and less than 3, the navigation accuracy is considered passable; if the weighted average of ATE and RTE is greater than 3, the navigation is considered a failure.

[0111] like Figure 9 , 10 As shown, the navigation accuracy evaluation results of the present invention under 28 experiments show that the ATE and RTE values ​​of all 28 experiments are below 2.

[0112] In this embodiment, the accurate pedestrian trajectory information provided by the video tracking system is used as calibration information, and a neural network model is trained through a latent loss function to realize the feature learning of foot vibration signals by the neural network.

[0113] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A pedestrian localization method based on CAM, SARB, and LSTM networks, characterized in that, Includes the following steps: Step 1: Collect footstep vibration signals; Step 2: Capture pedestrian video images using a camera and extract pedestrian trajectories; Step 3: Label the footstep vibration signals; Step 4: Extend the labeled footstep vibration signals using the sliding window segmentation method, and label each data segment with the pedestrian trajectory data extracted from the video. Step 5: Construct a neural network model based on the CAM module, input module, SARB module, and LSTM. Train the neural network model using the dataset processed in Step 4 and calculate the pedestrian trajectory. Step five specifically includes: Step 51: The CAM module performs max pooling and average pooling operations on the width and height of the vibration signal, respectively; and feeds it into a two-layer MLP network; then generates weights through the Sigmoid activation function, multiplies the input data and weights to obtain the first feature sequence; then uses a 7×7 convolution kernel and max pooling layer to extract features and reduce the dimensionality of the first feature sequence to obtain the second feature sequence. Step 52: The second feature sequence is processed by the input module and 4 sets of 2*SARB modules to obtain the third feature sequence; The SARB module contains two 3×1 convolutional layers, one self-attention layer, and one residual connection; Step 53: Input the third feature sequence into the bidirectional LSTM network to obtain the fourth feature sequence, and then output the predicted pedestrian velocity vector through the output layer and the fully connected layer. Step 54: Calculate the loss function between the predicted pedestrian speed and the actual pedestrian speed, and stop iterating when the maximum iteration is satisfied or the loss function satisfies the preset threshold; Step 6: Establish a navigation error assessment benchmark.

2. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 1, characterized in that, Footstep vibration signals are obtained by collecting voltage signals through a seismic detector, amplifying them, and then converting them from analog to digital to obtain data at a preset sampling frequency.

3. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 1, characterized in that, Foot vibration signals are preprocessed using low-pass filtering and linear normalization.

4. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 1, characterized in that, Step two specifically includes: Step 21: Lay a checkerboard carpet on the ground, take pictures of the checkerboard from different angles and heights, obtain the camera's intrinsic and extrinsic parameters, and establish the transformation from the image coordinate system to the world coordinate system; Step 22: Select the midpoint connecting the key points of both heels as the pedestrian's foot landing point, and extract the pedestrian's foot landing point from the video frame; Step 23: Sequentially connect the landing points of each frame to obtain the initial trajectory, and then perform coordinate transformation to obtain the spatial trajectory coordinate information of the pedestrian with the ground plane as the two-dimensional XY plane and the Z axis perpendicular to the ground plane.

5. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 1, characterized in that, Step three specifically includes: Step 31: Traverse and find the timestamps of the vibration signals, and mark the actual coordinates. Step 32: For vibration signals of pedestrian location coordinates without matching timestamps, use uniform interpolation to mark them; Step 33: Use the instantaneous velocity of each frame of the vibration signal to represent the pedestrian's movement trajectory.

6. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 5, characterized in that, The formula for instantaneous velocity is: （3） （4） in, and The velocities in the x and y directions are respectively, ts[] is a set of timestamps, and x[ i ] and y[ i [These are the] the [numbers]. i A set of trajectory coordinates of a frame.

7. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 1, characterized in that: The formula for calculating pedestrian trajectory is: ； in, It is the pedestrian's initial position; It is the position of the pedestrian at ts[i], where ts[i] is the position of the first... i Frame timestamp; and For the first i The velocity of the pedestrian in the x and y directions of the frame; j =2, 3… i .

8. The pedestrian localization method based on CAM, SARB, and LSTM networks according to claim 1, characterized in that: The navigation error evaluation benchmark is based on the absolute trajectory error. ATE and relative trajectory error RTE The accuracy of navigation is graded based on the magnitude of the numerical value. like ATE and RTE If the weighted average is less than 1, the navigation accuracy is considered excellent. ATE and RTE If the weighted average value is greater than 1 and less than 3, the navigation accuracy is considered to be passing. ATE and RTE If the weighted average value is greater than 3, navigation is considered to have failed.

Images