KR102243917B1 - Device and method for generating geomagnetic sensor based location estimation model using artificial neural networks - Google Patents

Abstract

An apparatus for generating a position estimation model is disclosed. The location estimation model generation apparatus includes a map generator for generating a magnetic field map including magnetic field values corresponding to respective coordinates in an indoor space, a data generator for generating training data, and an artificial neural network using the training data. , ANN) to generate a location estimation model.

Description

An apparatus and method for generating a magnetic field-based position estimation model using an artificial neural network {DEVICE AND METHOD FOR GENERATING GEOMAGNETIC SENSOR BASED LOCATION ESTIMATION MODEL USING ARTIFICIAL NEURAL NETWORKS}

The present invention relates to an indoor location based service, and more particularly, to an apparatus and method for generating a location estimation model capable of estimating a user's location based on a change in a magnetic field value due to a user's movement.

Demand for indoor location-based services is increasing explosively, but the existing radio-based indoor positioning technology has not been able to meet the required performance requirements. In particular, as a radio wave-based location recognition technology such as Wi-Fi and BLE (Bluetooth Low Energy), various techniques such as trilateration, fingerprint recognition, and TDoA (Time Difference of Arrival) have been proposed. Due to temporal instability and problems such as attenuation, reflection, and diffraction caused by indoor obstacles, errors in location recognition are reported to be 2 to 5 meters in laboratory environments and more than 10 meters in large indoor environments such as airports and shopping malls. It is insufficient to satisfy the situation.

In order to solve such a problem, attempts have been made to improve performance by combining Pedestrian Dead Reckoning (PDR) technology using an IMU (Inertial Measurement Unit) in a smartphone, but it still provides satisfactory performance. I can't. In addition, technologies for improving location recognition performance using various sensors such as ultrasonic waves, cameras, LEDs, and lasers have been developed, but there is a problem in that expensive equipment or infrastructure needs to be built.

Accordingly, in the present invention, by pre-learning the change of the indoor magnetic field using a recurrent neural network (RNN) technology, which is one of the machine learning methods of artificial intelligence, the change pattern of the indoor magnetic field is recognized only with a portable terminal such as a smartphone. We propose a method of estimating the indoor location. According to the present invention, it is possible to implement a system at low cost without hardware equipment such as an access point (AP) or beacon used in the existing radio wave-based location recognition while greatly improving the accuracy of indoor location recognition. Accordingly, it is expected that the performance and cost of indoor location recognition can be dramatically improved.

Republic of Korea Patent Publication No. 2016-0080842 (published on Jul. 08, 2016) Republic of Korea Patent Publication No. 2011-0025025 (published on March 9, 2011)

A technical problem to be achieved by the present invention is to provide an apparatus and method for generating a location estimation model capable of estimating a user's location by using a change in an indoor magnetic field value according to a user's movement.

The apparatus for generating a position estimation model according to an embodiment of the present invention includes a map generator that generates a magnetic field map including a magnetic field value corresponding to each coordinate in an indoor space, a data generator that generates learning data using the magnetic field map, and And a learning unit for generating a position estimation model by learning an artificial neural network (ANN) using the training data.

In the method for generating a location estimation model according to an embodiment of the present invention, the map generator of the location estimation model generation device generates a magnetic field map from magnetic field values corresponding to each coordinate in an indoor space using a linear interpolation method, and the location estimation model is generated. And generating, by a data generation unit of the device, training data using the magnetic field map, and generating, by a learning unit of the position estimation model generation unit, a recurrent neural network using the training data to generate a position estimation model.

A location estimation apparatus according to an embodiment of the present invention is a location estimation model apparatus in which the location estimation model generated by the location estimation model generation method is stored, and a magnetic field that measures a magnetic field value that changes according to a user's movement in the indoor space A position estimating unit for estimating the user's position by inputting sequential magnetic field values measured by the sensor unit, the magnetic field sensor unit into the position estimating model, and a position for outputting the user’s position estimated by the position estimating unit Includes an output section.

In the case of the apparatus and method for generating a location estimation model according to an embodiment of the present invention, there is an effect of estimating the location of a user with high accuracy even in an environment where expensive equipment or infrastructure is not installed.

In order to more fully understand the drawings cited in the detailed description of the present invention, a detailed description of each drawing is provided.
1 shows the magnetic field distribution in the Northeast Asia region.
2 is a diagram for visualizing the magnitude of a magnetic field vector measured at an experiment site.
3 shows the basic structure of a recurrent neural network (RNN) used in the present invention.
4 is a diagram illustrating a generated magnetic field map.
5 shows a process of generating training data and training a neural network.
6 is a graph showing an error according to the number of executions for each number of hidden nodes.
7 shows the change of the learning error rate according to the size of the mini-batch.
8 shows a result of location recognition using evaluation data.
9 shows a performance result of location recognition according to the number of hidden nodes.
10 is a functional block diagram of an apparatus for generating a position estimation model according to an embodiment of the present invention.
11 is a functional block diagram of a location estimation apparatus according to an embodiment of the present invention.
12 is a flowchart illustrating a method of generating a position estimation model performed by the apparatus for generating a position estimation model illustrated in FIG. 10.
13 is a flowchart illustrating a method of estimating a position performed by the position estimating apparatus illustrated in FIG. 11.
14 is an exemplary hardware block diagram of the apparatus for generating a position estimation model illustrated in FIG. 10 or the apparatus for estimating a position illustrated in FIG. 11.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in the present specification are exemplified only for the purpose of describing the embodiments according to the concept of the present invention, and the embodiments according to the concept of the present invention are It may be implemented in various forms and is not limited to the embodiments described herein.

Since the embodiments according to the concept of the present invention can apply various changes and have various forms, the embodiments will be illustrated in the drawings and described in detail in the present specification. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosed forms, and includes all changes, equivalents, or substitutes included in the spirit and scope of the present invention.

Terms such as first or second may be used to describe various constituent elements, but the constituent elements should not be limited by the terms. The above terms are only for the purpose of distinguishing one component from other components, for example, without departing from the scope of the rights according to the concept of the present invention, the first component may be referred to as the second component and similarly the second component. The component may also be referred to as a first component.

When a component is referred to as being "connected" or "connected" to another component, it is understood that it is directly connected to or may be connected to the other component, but other components may exist in the middle. It should be. On the other hand, when a component is referred to as being "directly connected" or "directly connected" to another component, it should be understood that there is no other component in the middle. Other expressions describing the relationship between components, such as "between" and "directly between" or "adjacent to" and "directly adjacent to" should be interpreted as well.

The terms used in the present specification are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, terms such as "comprise" or "have" are intended to designate the presence of features, numbers, steps, actions, components, parts, or combinations thereof described herein, but one or more other features. It is to be understood that the presence or addition of elements or numbers, steps, actions, components, parts, or combinations thereof, does not preclude in advance the possibility of the presence or addition.

Unless otherwise defined, all terms used herein including technical or scientific terms have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs. Terms as defined in a commonly used dictionary should be construed as having a meaning consistent with the meaning of the related technology, and should not be interpreted as an ideal or excessively formal meaning unless explicitly defined in the present specification. Does not.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Earth's magnetic field refers to the flow of a huge magnetic field from the Earth's south magnetic pole to the north magnetic pole. According to the 2015 Global Magnetic Field Model by the National Oceanic and Atmospheric Administration (NOAA), the Earth's magnetic field has a different distribution by region. Referring to FIG. 1 showing the magnetic field distribution in the Northeast Asia region, Korea is located in a region of 50 μT (micro tesla).

The values of the Earth's magnetic field, which have different distributions by region, show a larger difference indoors. This is because materials such as H-beams are used in the structure of indoor buildings, and various architectural structures such as steel entrance doors, elevator structures, and electric motors affect the indoor magnetic field value. It was confirmed that the magnetic field values were distributed in the range of 10 μT to 120 μT at the actual test site (Korea University Engineering Hall lobby).

Earth's magnetic field values are expressed as vectors in three-dimensional space. Referring to FIG. 2, which visualizes the magnitude of the magnetic field vector measured at the experiment site, the movement of a pedestrian indoors causes a change in the measured magnetic field value, and can be visualized as a movement on a magnetic field map contour line. In addition, when attempting to recognize an indoor location using a magnetic field, the following characteristics must be considered.

1) In an indoor space, there may be several places with the same magnetic field vector value. Therefore, it is difficult to determine the current position using only a single magnetic field vector value.

2) Pedestrian movement has a continuous movement trajectory. This means that the difference between the position of the pedestrian at time t-1 and the position of the pedestrian at time t is small and has a continuous value. Therefore, the current position is affected by the position of the previous time.

3) Like the pedestrian movement trajectory indoors, the change in the magnetic field value is also continuous. That is, the magnetic field values at adjacent locations do not have a large difference, and therefore, the change of the magnetic field value according to the movement of a pedestrian has a continuous characteristic like a movement on a contour line.

Taken together, the trend of the change of the magnetic field vector value detected according to the movement of the pedestrian can be displayed as a curved waveform, which forms a unique waveform pattern as the length of the continuous magnetic field vector value, that is, the magnetic field vector value sequence becomes longer. Therefore, the current position that can be estimated has a characteristic that converges to one place.

In the present invention, a recurrent neural network (RNN) technique, which is one of deep learning technologies, is used to estimate an accurate indoor location by inputting a change trend of an indoor magnetic field value. Among various artificial neural networks (ANNs), a recurrent neural network (RNN) has a feedback loop from an output to an input. Therefore, the output result learned by the recurrent neural network (RNN) has a characteristic that is determined not only by the current input value but also by all past input values, so that the input and output data have a continuous relationship with each other on the time axis. It can be used suitably.

3 shows the basic structure of a recurrent neural network (RNN) used in the present invention.

을 입력 층(input layer)에서 은닉 층(hidden layer)으로의 링크에 적용되는 가중치,

를 은닉 층에서 은닉 층으로의 링크에 적용되는 가중치,

을 은닉 층에서 출력 층(output layer)으로의 링크에 적용되는 가중치, 2차원 공간상의 점인

를 시간

시점에서의 최종 출력,

를 중간층의 활성화 함수,

을 출력 층의 활성화 함수라 하면, 순환 신경망 모델의 시간

시점에서의 출력

는 수학식 1과 같다.

The weight applied to the link from the input layer to the hidden layer,

The weight applied to the link from the hidden layer to the hidden layer,

Is the weight applied to the link from the hidden layer to the output layer, which is a point in two-dimensional space.

Time to

Final output at the time point,

The activation function of the middle layer,

Is the activation function of the output layer, the time of the recurrent neural network

Output at the time point

Is equal to Equation 1.

In the present invention, based on a magnetic field map, an artificial neural network optimized for indoor positioning based on a magnetic field is constructed by learning in advance a magnetic field change pattern according to a pedestrian's moving path. The magnetic field map means that a magnetic field value and actual coordinates measured after measuring a magnetic field at a reference point set at regular intervals in an indoor space using a smartphone or a magnetic field sensor capable of measuring a magnetic field are configured in one table. At this time, as the distance between each measurement reference point becomes finer, the accuracy of location recognition increases, but there is a problem that the time and effort required to generate the magnetic field map increase. Therefore, in the present invention, noting that the magnetic field increases or decreases linearly with distance, the magnetic field map with high resolution is generated by interpolating the magnetic field values of the coordinates that are not actually measured through linear interpolation. can do. 4 is a diagram illustrating a generated magnetic field map.

In order to recognize indoor location using deep learning, it is necessary to optimize through learning of a neural network in advance. In order to generate training data for neural network training, the user's movement is virtually generated based on the magnetic field map. The generated virtual path is composed of coordinates and magnetic field values mapped to the coordinates. Since the amount of training data has a great influence on the indoor location recognition performance, the training data must be sufficiently generated so that the entire map can be sufficiently learned. By automatically configuring the process of generating the training data, the time required to generate the training data can be greatly reduced. Neural network through the feed-forward process and back-propagation process by inputting the magnetic field vector value of the generated training data as the input value of the neural network and the coordinates of the generated training data as the output target value of the neural network. An optimized neural network can be generated by learning a magnetic field map for. The process of generating training data and learning a neural network is illustrated in FIG. 5.

The process for learning a neural network is largely divided into a feed-forward process and a back-propagation process. In the forward propagation process, first, the values of the magnetic field vectors of the three X, Y, and Z axes are input to the input layer to calculate a linear function with the weight of the hidden layer. Thereafter, the calculation result is subjected to an activation function and the range of the calculation result is adjusted. In the present invention, a ReLU (Rectified Linear Unit) function is used as an activation function, but the scope of the present invention is not limited to the type of activation function. In a neural network that is not sufficiently trained, the weight of the hidden layer is not optimized, so the result output to the output layer shows a difference from the original target result. In this way, the difference between the result output to the output layer and the original target result is defined as a learning error. In order to reduce the learning error, a backpropagation process is performed. In the backpropagation process, first, an error signal in the output layer is calculated using an error function. In the present invention, the MSE (Mean Squared Error) function is used as the error function, but the scope of the present invention is not limited to the type of the error function. Next, the weight of the hidden layer is modified through a back-propagation weight update formula. In the present invention, a gradient descent optimizer was used. By repeatedly performing the above-described process, an optimized neural network that converges to the lowest error rate can be constructed.

The recurrent neural network has the characteristic of circulating and reusing the output result of the hidden layer computed at time t-1 at the next computation time t and t+1, which forms a correlation in the output result. In "Tensorflow", a deep learning framework used in the present invention, this is called a state, and it is processed as an input to the next unit time operation. Due to this characteristic of the recurrent neural network, it can be suitably used in indoor location recognition with a continuous path.

In the learning process of the artificial neural network, the number of hidden nodes in the hidden layer, the size of the mini-batch, the setting of the total number of executions (epoch), and the learning rate are factors that directly affect the learning result. It is important to find the optimal value according to the characteristics of the data. In the present invention, the number of hidden nodes is increased to 2, 5, 10, and 20, and the total number of executions converged to the target error rate is compared, and the result is shown in FIG. 6. Referring to FIG. 6, it can be seen that as the number of hidden nodes increases, the total number of executions converging to the target error rate decreases.

If the artificial neural network training is performed by configuring the entire data set of training data in one batch, the first learned data is forgotten as the number of training iterations of the neural network increases, and a problem that the gradient rapidly increases (Gradient Exploding) may occur. have. In addition, due to the structural characteristics of the "Tensorflow" application, which is a framework used in the present invention, if all neural networks for the entire batch size are configured in advance and an experiment is performed, a memory shortage problem may occur in the learning process. To solve this problem, in the present invention, a batch normalization process of learning each small unit by reconstructing the entire data set into a small unit batch size using mini-batch was performed. As the size of the mini-batch increases, the rate of convergence to the lowest error rate is faster, but the convergence error rate gradually increases. The change in the learning error rate according to the size of the mini-batch is shown in FIG. 7.

Depending on the size and resolution of the magnetic field map, the optimized value of the neural network is different, and the optimal performance in indoor location recognition can be obtained by configuring the optimized neural network for the real environment by adjusting the error rate through substitution of several parameter values.

자기장 값(또는 평가 데이터)은 신경망의 입력 값으로 입력되며, 신경망 순전파 과정을 통해 현재 좌표가 출력되고, 신경망의 특성을 이용하여 다음 단위 시간에서 연속적인 위치 인식 결과가 출력될 수 있다. 실제 위치 인식에서는 현재 위치를 알 수 없기 때문에 오차 함수 계산이 불가능하므로 신경망 최적화를 위한 역전파 과정을 진행할 수는 없다.In order to evaluate a trained neural network, that is, a position estimation model, position recognition may be performed by inputting evaluation data into an optimized neural network. Measured by the magnetic field sensor

The magnetic field value (or evaluation data) is input as an input value of the neural network, the current coordinates are output through a forward propagation process of the neural network, and a continuous position recognition result may be output at the next unit time using the characteristics of the neural network. In real position recognition, since the current position is not known, the error function cannot be calculated, so the backpropagation process for optimizing the neural network cannot be performed.

The number of hidden nodes was set to 20, the learning rate was set to 0.001, and the batch size was set to 200. As a result of performing a simulation using the generated learning data, the result was a minimum of 0m, a maximum of 2.091m, and an average of 1.256m. The location recognition error is shown. The result of location recognition using the evaluation data is shown in FIG. 8.

In addition to the single simulation result, the number of hidden nodes was gradually increased to 2, 4, 10, 50, 200, 500, and 1000. As a result of the simulation, as the number of hidden nodes increased, the performance of location recognition tended to improve. If more than 200 hidden nodes were used, it could be confirmed that the location recognition error slightly increased. Tables 1 and 9 show performance results of location recognition according to the number of hidden nodes when five different training data and evaluation data are used. In FIG. 9, a box thickened on each element is a range in which 80% of all position recognition samples are gathered, and other bars are a range in which 99% of samples are gathered.

Number of hidden nodes Minimum error (m) Mean error (m) Maximum error (m) 2 0 5.815 10.587 4 1.077 5.159 11.044 10 0.777 2.419 3.337 20 0.781 1.498 1.814 50 0.974 1.594 1.937 200 0.08 2.497 3.879 500 0.058 2.251 3.356 1000 0.085 2.204 3.656

10 is a functional block diagram of an apparatus for generating a position estimation model according to an embodiment of the present invention.

The location estimation model generation apparatus 100, which may be referred to as a location estimation model generation server, generates a magnetic field map including information on a magnetic field distribution in a predetermined indoor space, and estimates a location using the generated magnetic field map and learning data. By learning the model, an optimized position estimation model can be generated. The location estimation model generation apparatus 100 includes a map generation unit 110, a data generation unit 130, and a learning unit 150. According to an embodiment, the apparatus 100 for generating a location estimation model may further include at least one of a model evaluation unit 170 and a storage unit 190.

The map generator 110 may generate a magnetic field map including magnetic field distribution information according to coordinates of an indoor space in which position estimation is to be performed. The magnetic field map may mean that coordinates of each reference point in an indoor space in which reference points are set at regular intervals and a magnetic field value corresponding thereto are configured as one table.

The magnetic field value at each of the reference points may be measured using a magnetic field sensor, such as a smart phone, and the measured magnetic field value is the position estimation model generating apparatus 100 through a predetermined input interface or wired/wireless communication network. It may be transmitted to the map generator 110.

In addition, the map generator 110 may increase the resolution of the magnetic field map generated by using a predetermined algorithm, for example, linear interpolation. That is, by using the linear interpolation method, a magnetic field map with high resolution may be generated by interpolating a magnetic field value at a point where a magnetic field value is not directly measured.

The data generator 130 may generate training data to be used for training and/or evaluation of the position estimation model. For example, the data generator 130 may generate a virtual path on an indoor space in which position estimation is to be performed. The virtual path may include coordinates of each point included in the virtual path and a magnetic field value mapped to each point. A set of coordinates of each continuous point included in the virtual path and a magnetic field value mapped to each point may be referred to as learning data.

In this case, when a position estimation model, that is, a recurrent neural network, is trained by configuring the entire training data set into one batch, the first learned data is forgotten as the number of iterations of the recurrent neural network training increases, and the gradient exploding rapidly increases. ) Problems may occur. Accordingly, the data generator 130 may divide the entire training data into a plurality of mini-batch having a predetermined batch size. In this case, a recurrent neural network may be trained using training data divided into a plurality of mini-batch.

The learning unit 150 learns an artificial neural network (ANN), in particular a recurrent neural network (RNN), using the training data generated by the data generation unit 130, and thus an optimized recurrent neural network, That is, a location estimation model can be generated. Specifically, the learning unit 150 inputs a magnetic field value included in the training data generated by the data generation unit 130 through a forward propagation process into the input layer of the recurrent neural network, and output coordinates corresponding to the input magnetic field value are obtained. You can get it. In addition, the learning unit 150 may modify the weight of the hidden layer constituting the recurrent neural network by using an error between the output coordinate and the coordinate included in the training data, that is, a target value through a backpropagation process. The learning unit 150 may generate an optimized recurrent neural network that converges to a minimum error rate or a target error rate by repeatedly performing a learning process using training data having continuous magnetic field values and continuous coordinate values. The optimized recurrent neural network may also be referred to as a position estimation model.

The model evaluation unit 170 may evaluate an error rate of the position estimation model learned by the learning unit 150 by using a part of the training data generated by the data generation unit 130, that is, the evaluation data. In real position recognition, the error function cannot be calculated because the current position is not known, so the backpropagation process for optimizing the neural network cannot be performed. Accordingly, the model evaluation unit 170 may input a continuous magnetic field value included in the evaluation data into the position estimation model to obtain a coordinate value corresponding thereto. By comparing the error between the output coordinate value and the coordinate value included in the evaluation data, the error rate of the position estimation model may be evaluated. According to an embodiment, the evaluation data may be separately generated by the data generator 130. In this case, the process of generating the evaluation data may be the same as the process of generating the training data.

The storage unit 190 may store data used or generated in a process of generating a location estimation model, a recurrent neural network algorithm, an optimized location estimation model, and the like.

The location estimation model generation apparatus 100 illustrated in FIG. 10 may be implemented as a computing device such as a server, a personal computer (PC), a tablet PC, and a notebook. The computing device will be described in detail with reference to FIG. 14.

11 is a functional block diagram of a location estimation apparatus according to an embodiment of the present invention.

The location estimation apparatus 300 may be implemented as a computing device in which a location estimation model generated by the location estimation model generation apparatus 100 is stored, and the computing device will be described in detail with reference to FIG. 14. The location estimation apparatus 300 installs the location estimation model directly from the location estimation model generation apparatus 100 through a wired/wireless communication network or via at least one relay server, for example, an app store server (also referred to as an application store server). The location estimation model can be installed in the location estimation apparatus 300 by downloading in the form and installing the installation file. Accordingly, it is obvious that the location estimation apparatus 300 and the location estimation model generation apparatus 100 may include a communication interface for performing a communication function.

The position estimation apparatus 300 may include a magnetic field sensor unit 310, a position estimation unit 330, and a position output unit 350.

The magnetic field sensor unit 310 may measure a magnetic field value that changes as a user carrying the position estimating device 300 moves in an indoor space at a predetermined time interval. Accordingly, sequential magnetic field values may be output from the magnetic field sensor unit 310.

The position estimating unit 330 may estimate the user's position by using magnetic field values output from the magnetic field sensor unit 310 as an input value of the position estimation model. The location estimation model used by the location estimation unit 330 may mean an optimized artificial neural network generated by the location estimation model generation apparatus 100 shown in FIG. 10. The magnetic field values at several locations in an indoor space may be the same. However, since the moving path corresponding to the continuously changing magnetic field value converges into one, the output value of the position estimation unit 330 for the initial magnetic field value, that is, the output coordinate does not exactly match the actual user's position, but the continuous magnetic field After the value is input, the output value of the position estimating unit 330 is distributed within a relatively small error range from the actual user's position.

The position output unit 350 may provide the user's position estimated by the position estimating unit 330 to the user through a predetermined display device. For example, the location output unit 350 may display the estimated user's location or the estimated user's moving path from at least a part of a map of an indoor space to perform location estimation.

The location estimation apparatus shown in FIG. 11 includes a mobile phone, a smart phone, a notebook, a net-book, an e-reader, a tablet PC, and a PDA. A portable or portable computing device such as a (personal digital assistant), navigation, portable multimedia player (PMP), MP3 player, or MP4 player may be implemented.

In addition, each of the configurations of the location estimation model generating apparatus 100 shown in FIG. 10 and the location estimation apparatus 300 shown in FIG. 11 is shown to be functional and logically separate, and each configuration must be separate. The average expert in the technical field of the present invention can easily infer that it does not mean that it is classified as a physical device or written as a separate code. In addition, each of the configurations of the location estimation model generation apparatus 100 and the location estimation apparatus 300 may mean a functional and structural combination of hardware for performing the technical idea of the present invention and software for driving the hardware. , It may mean a logical unit of a predetermined code and a hardware resource for executing the predetermined code, and does not necessarily mean a physically connected code or a single type of hardware.

12 is a flowchart illustrating a method of generating a position estimation model performed by the apparatus for generating a position estimation model illustrated in FIG. 10. In describing the method of generating a position estimation model, a detailed description of a description overlapping with that described above will be omitted.

First, the map generation unit 110 of the location estimation model generation apparatus 100 may generate a magnetic field map including magnetic field distribution information for each coordinate of an indoor space to perform location estimation (S110). According to an embodiment, the map generator 110 may generate a magnetic field map with high resolution using a predetermined algorithm, for example, a linear interpolation method.

The data generation unit 130 of the position estimation model generation apparatus 100 may generate training data for use in training a recurrent neural network (S130). For example, the data generator 130 may generate a virtual path on an indoor space in which position estimation is to be performed. The virtual path may include coordinates of each point included in the virtual path and a magnetic field value mapped to each point. According to an embodiment, the data generator 130 may divide the training data into a plurality of mini-batch having a predetermined batch size. In this case, a recurrent neural network may be trained using training data divided into a plurality of mini-batch.

The learning unit 150 of the position estimation model generation apparatus 100 learns an artificial neural network, in particular a recurrent neural network, using the training data generated by the data generation unit 130 to generate an optimized artificial neural network, that is, a position estimation model. It can be generated (S150). Specifically, the learning unit 150 may correct the weight of the hidden layer included in the recurrent neural network by using the error between the coordinates output through the forward propagation process and the target coordinate in the backpropagation process. In addition, the learning unit 150 may generate an optimized recurrent neural network that converges to a minimum error or a target error by repeatedly performing a learning process using learning data having sequential magnetic field values and sequential coordinate values.

According to an embodiment, the method for generating a location estimation model may further include a location estimation model evaluation step S170 by the model evaluation unit 170. That is, the model evaluation unit 170 may evaluate an error rate of the position estimation model learned by the learning unit 150 by using the evaluation data.

13 is a flowchart illustrating a method of estimating a location performed by the apparatus for estimating a location illustrated in FIG. 11. In describing the position estimation method, a detailed description of a description overlapping with the previously described description will be omitted.

First, the magnetic field sensor unit 310 of the position estimating device 300 may measure a magnetic field value that changes according to a movement of a user holding the position estimating device 300 in an indoor space at a predetermined time interval (S310). ). Accordingly, sequential magnetic field values may be output from the magnetic field sensor unit 310.

The position estimating unit 330 of the position estimating apparatus 300 may estimate the user's position by using sequential magnetic field values output from the magnetic field sensor unit 310 as an input value of the position estimation model (S320). Since the user's moving path corresponding to the continuously changing magnetic field value in the indoor space converges to one, the output value of the position estimating unit 330 for the initial magnetic field value, that is, the output coordinate does not exactly match the actual user's position. The output value of the position estimating unit 330 after continuous magnetic field values are input distributes the actual user's position within a relatively small error range.

The position output unit 350 of the position estimating device 300 may provide the user's position estimated by the position estimating unit 330 to the user through a predetermined display device (S330). For example, the location output unit 350 may display the estimated location of the user or the estimated moving path of the user from at least a part of a map of an indoor space where location estimation is performed.

In addition, since the position estimation model installed in the position estimation apparatus 300 is a position estimation model generated by the position estimation model generation apparatus 100, the position estimation method includes the position estimation apparatus 300 using the position estimation model generation apparatus ( 100) Alternatively, the method may further include downloading the location estimation model in the form of an installation file via a separate server and installing the downloaded installation file to install the location estimation model in the location estimation apparatus 300. .

14 is an exemplary hardware block diagram of the apparatus for generating a position estimation model illustrated in FIG. 10 or the apparatus for estimating a position illustrated in FIG. 11.

The location estimation model generation apparatus 100 of FIG. 10 and/or the computing device 500, which is an example implementation of the location estimation apparatus 300 of FIG. 11, includes a communication interface 501, an input interface 503, an output interface 505, and It includes a memory 507, a storage device 509, at least one processor 511, and a system bus/control bus 513.

The communication interface 501 is configured to communicate with a device external to the computing device 500. The communication interface 501 is configured to perform wired or wireless communication, and may include, for example, a MAC chip capable of being connected to a wireless LAN or a wired LAN.

The input interface 503 may receive a user input for controlling the computing device 500 and/or a magnetic field value for position estimation. The input interface 103 may include a keyboard, a touch pad, a mouse, a magnetic field sensor, and the like.

The output interface 505 may output an operation result of the computing device 500, for example, a magnetic field map generation result, a learning result of a recurrent neural network, an evaluation result of a recurrent neural network, or a position estimation result. The output interface 505 may include a display, a speaker, or the like.

The memory 507 includes volatile memory and/or non-volatile memory. The memory 507 temporarily or non-temporarily stores various types of data and programs. The volatile memory temporarily stores various data and program segments, and the non-volatile memory may non-temporarily store various setup data and boot programs required for setup or booting.

The storage device 509 is a mass storage medium that stores various types of data and stores various programs. The storage device 509 includes at least a program for performing a method for generating a position estimation model or a method for estimating a position according to the present invention. The storage device 509 includes data to be used in a program of a method for generating a position estimation model or a method for estimating a position. These data can be accessed and managed by specific programs. The data used for the location estimation model generation and location estimation and this specific program are preferably configured as a database. A hard disk drive (HDD) or a solid state drive (SSD) may be an example of the storage device 509.

The processor 511 may load a program code of a program stored in a nonvolatile memory or a storage device 509 and execute the program code. The processor 511 includes an execution unit capable of executing an instruction of a program code, and trains a program for generating a position estimation model or estimating a position, and various data used for generating a position estimation model or estimating a position. You can run a program for doing so.

The system bus/control bus 513 is configured to transmit and receive control signals or data signals between each hardware block. The system bus/control bus 513 may be a parallel bus or a serial bus.

The present invention has been described with reference to the embodiments shown in the drawings, but these are only exemplary, and those of ordinary skill in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical idea of the attached registration claims.

100: position estimation model generation device
110: map generator
130: data generation unit
150: Learning Department
170: model evaluation unit
190: storage
300: position estimation device
310: magnetic field sensor unit
330: position estimation unit
350: position output unit
500: computing device

Claims (11)

A map generator for generating a magnetic field map including a 3D magnetic field vector value corresponding to each coordinate in the indoor space;
Learning including the coordinates of each point included in the virtual paths and a 3D magnetic field vector value corresponding to the coordinates of each point using the magnetic field map and generating all virtual paths that a pedestrian can move in the indoor space A data generator for generating data; And
By learning a recurrent neural network (RNN) using the training data, generating a position estimation model that estimates the current position of the user by receiving a continuous 3D magnetic field vector value that changes according to the user's movement. Position estimation model generating apparatus including a learning unit. The method of claim 1,
The magnetic field map is characterized in that it is generated using a linear interpolation (linear interpolation),
Position estimation model generation device. The method of claim 1,
The training data is characterized in that it is divided into a plurality of mini-batch having a predetermined batch size (batch size),
Position estimation model generation device. The method of claim 2,
The learning unit repeats learning until the output result of the recurrent neural network reaches a target error rate,
Position estimation model generation device. The method of claim 1,
The position estimation model generation device,
Further comprising a model evaluation unit for evaluating the error rate of the position estimation model using the evaluation data generated by the data generating unit,
Position estimation model generation device. Generating a magnetic field map by collecting a 3D magnetic field vector corresponding to each coordinate in an indoor space by using a linear interpolation method by a map generator of the position estimation model generating apparatus;
The data generation unit of the location estimation model generation device generates all virtual paths that a pedestrian can move in the indoor space, and corresponds to the coordinates of each point included in the virtual paths and the coordinates of each point using the magnetic field map. Generating training data including a three-dimensional magnetic field vector value; And
The learning unit of the location estimation model generating device learns an artificial neural network (ANN) using the training data, receives a continuous 3D magnetic field vector value that changes according to the user's movement, and receives the current location of the user. A method for generating a position estimation model comprising the step of generating a position estimation model to estimate. The method of claim 6,
The training data is characterized in that it is divided into a plurality of mini-batch having a predetermined batch size (batch size),
How to create a location estimation model. The method of claim 6,
Generating the position estimation model,
Repeating the learning until the output result of the artificial neural network reaches a target error rate,
How to create a location estimation model. The method of claim 6,
The location estimation model generation method,
Further comprising the step of evaluating an error rate of the position estimation model by using the evaluation data generated by the data generation unit, the model evaluation unit of the position estimation model generation device,
How to create a location estimation model. In the position estimation apparatus in which the position estimation model generated by the method according to claim 6 is stored,
A magnetic field sensor unit measuring a continuous three-dimensional magnetic field vector value that changes according to the movement of the user in the indoor space;
A position estimating unit for estimating a current position of the user by inputting a continuous 3D magnetic field vector value measured by the magnetic field sensor unit into the position estimating model; And
Position estimating device comprising a position output unit for outputting the current position of the user estimated by the position estimating unit. The method of claim 10,
The location estimation device is a smartphone (smart phone), a mobile phone (mobile phone) or a tablet PC (tablet personal computer),
Position estimation device.

Images