KR101637195B1

KR101637195B1 - Method for classifying human mobility state using particle filter and probabilistic distribution parameter dependent on speed

Info

Publication number: KR101637195B1
Application number: KR1020150082874A
Authority: KR
Inventors: 송하윤; 백지현
Original assignee: 홍익대학교 산학협력단
Priority date: 2015-06-11
Filing date: 2015-06-11
Publication date: 2016-07-07

Abstract

Disclosed is a method of determining a mobility state of a specific target, based on the average of particles obtained by performing a particle update process of updating the value of the particles defined as independent random variables. The method comprises the steps of: calculating one or more parameters for a first type probability distribution, based on speed values belonging to a first time window having a size according to predetermined rules; and calculating a value on cumulative probability, at which the specific target has a specific speed, using the calculated one or more parameters. At this time, the value of the particle is updated by using the cumulative probability.

Description

[0001] The present invention relates to a velocity-dependent probability distribution parameter and a method for classifying a human motion state using a particle filter,

The present invention relates to a method and apparatus for classifying a human motion state using a particle filter used in an information processing field, and more particularly, to a method using a velocity-dependent probability distribution parameter.

Due to the proliferation of smartphones, a large amount of GPS data has been generated as a result of easily obtaining smartphone user's GPS data using various applications. In addition, even if the GPS data does not depend on the GPS data, a large amount of position data can be generated by providing a technique of determining the position of a person or an object using various techniques.

If the moving state of a person can be known only by the position of the person obtained using a smartphone or a portable GPS collector and the time information at each position, the result can be used for various applications. For example, when applied to motion-related applications, when a user jogs, the calorie consumption can be accurately calculated by checking the current user's state in real time without calculating the calories consumed by the exercise at the average speed. Also, for example, if a user repeatedly stops at a specific position while moving, it can be estimated from the position data only that there is a pedestrian crossing at that position. If you set the type of state to a total of two things: stay (stable, stay, stationary) and mobile (mobile), you can express everyone's status in both.

The positioning data collected using the position data collection device may contain errors for various reasons. Therefore, the position data containing the errors must be filtered or classified. That is, the classification step of the mobility states must be performed prior to the use of the location data. This kind of classification can not be done deterministically, but it can be done stochastically.

In the present invention, the simplest classification having only two states of 'mobile' and 'stable, stay, stationary' is used to classify a human's movement state. This sort of classification will increase the accuracy required by application programs for position data. In the next step, the classification of speed value will lead to a classification of transportation.

Human movement status discrimination was used as an intermediate process in studying the effect of weather on human behavior patterns.

However,

Horanont T, Phithakkitnukoon S, Leong TW, Sekmoto Y, Shibasaki R (2013) Weather Effects on the Patterns of Peaple's Everyday Activities: PLos ONE 8 (12): e81153.doi: 10.1371 / journal.pone.0081153

In this study, we used m sets of location information of latitude, longitude, and time, which are continuous time, to distinguish between movement and suspension. When the distance between the ordered pairs of all the different position information of this set is less than a certain limit, it is judged as a stop, and it is judged as the movement.

In the present invention, a technique used in a so-called " particle filter " is used to perform the classification described above.

Particle filters are also called sequential Monte Carlo methods and are based on Bayesian statistics. Particle filters are used for parameter estimation and state estimation. The basic idea of a particle filter is to create and use a large number of independent random variables. The independent random variables at this time are called particles. Particles are initialized from the state space and initialized. Also, the values of the particles are updated with input of weights, which are observations newly input. The theory of particle filters is also presented in Paper 2 below.

However,

Chen, Zhe "Bayesian filtering: From Kalman filters to particle filters. and beyond "Technical report, McMaster University, 2003.

Particle filter algorithms include Sampling Importance Resampling (SIR) and Sequential Importance Sampling (SIS). The biggest difference between the two is the absence of a resampling process.

The operation sequence of the SIS algorithm among the particle filters is as follows. First, initialize the generated particles (X) appropriately. Secondly, the new measured value Z is input and the likelihood probability P (Z | X) is updated. Third, the value of the particles X is updated using the weight W obtained by using the likelihood probability. Finally, set the conditions and repeat the second and third steps.

Currently, Particle Filter can be used to estimate the robot's position in the robot field.

However,

S Thrun "Particle Filters in Robotics" Proceeding UAI'02 Proceedings of the Eighteenth Conference on Uncertainty in Artificial Intelligence, Pages 511-518, 2002

In addition, it is shown in the following paper 4 that a particle filter can be used to estimate a human's position using a wi-fi signal in computer engineering.

However,

"A Particle Filter for Location Estimation of WiFi Users", Journal of the Korean Institute of Information Scientists and Engineers, Vol. 39, No. 2012.

In the present invention, as a method for analyzing position data, a new technique for combining parameters relating to position data to a particle filter is provided. Particularly, when calculating the weight used for updating the particles in the particle filter, if the parameter is set to a fixed constant, an analysis result that does not reflect the actual situation may appear. In order to solve this problem, the present invention provides a technique of setting the parameter to a value varying with time and using the parameter.

In the position data analysis method according to one aspect of the present invention, a weight of a particle used for updating particles in a particle filter is varied in time by a parameter constituting a cumulative distribution function of a specific probability distribution function Technology. Specifically, the parameter is calculated and changed by using the statistical values of the velocities included in the window moving with time. The statistical value includes an average (expected value) and a standard deviation of the velocities.

In the present invention, the movement state of a person will be grasped by using a Particle Filter based on position data obtained by using an arbitrary position determining device such as GPS. In the present invention, among the particle filter algorithms, an SIS algorithm without a resampling process can be used.

In one aspect of the invention, the velocity value can be used directly in the particle filter, which may result in some inaccurate results. That is, particles associated with a slow velocity may be considered stationary, which may be a different result than the fact. And a toggle between a 'move' state and a 'stop' state may be caused by a position error. Therefore, in the present invention, it is possible to further use a method of using the history value of the speed.

In the present invention, as a method of analyzing the position data, a velocity is calculated using time, latitude and longitude included in GPS data, which is movement information of a person, and then the particle filter Particle Filter) to determine the state of human movement.

The human moving state can be given as a probability value that is a stationary state and a probability value that is a moving state. In the present invention, it is possible to determine whether the movement state is 'stop' or 'movement' by using the probabilistic values obtained through the particle filter for determining the movement probability of a human. The results obtained by one embodiment of the present invention can be shown on a map.

According to an aspect of the present invention, it is possible to provide a discrimination method for discriminating a moving state of a specific object by using a particle filter having particles defined as N independent random variables. The method includes calculating a current velocity for the particular object and calculating a value for a cumulative probability that the particular object has the current velocity; Repeating the particle update process for updating each of the particles by a predetermined number of times according to a predetermined rule; And calculating an average value of the values of the N particles that have been updated, and determining whether the specified object is in a stopped state or a moving state based on the average value. In this case, a weight used for updating the value of the particle is calculated using the cumulative probability.

The particle updating process may include calculating a weight by subtracting a value of the cumulative probability from the value of each particle; And updating the value of each particle by subtracting the weight from the value of each particle or adding the weight to the value of each particle.

The current velocity may be obtained by multiplying a predetermined weight by one or more velocity values detected before the first time and the first time with respect to the specific object and adding them to each other to obtain a time weighted- Or a velocity value detected at the first time with respect to the specific object or an average value of one or more velocity values detected before the first time and the first time with respect to the specific object have.

At this time, the more the time at the first time, the larger the value of the weight to be multiplied.

In this case, the velocity value detected at the first time may be an uncorrected velocity value, and the velocity value detected before the first time may be a corrected or uncorrected velocity value.

According to another aspect of the present invention, there is provided a method of generating a particle filter, the method comprising: initializing a particle filter having N independent random variables as particles, Repeating the particle update process a number of times according to a predetermined rule; And calculating an average value of the N particles having been updated and determining whether the specific object is in a stopped state or a moving state using the average value. At this time, the current velocity provided to the specific object at the first time is used as a new observation value to be input to the particle filter.

At this time, the particle update process may include: obtaining the current velocity; Updating a likelyhood probability that the specific object has the current velocity based on a current value of the N particles; Obtaining a weight using the updated likelihood probability; And updating the particle using the weight.

In this case, the step of updating the particle may include updating the particle to a value of 0 when the particle has a negative value.

At this time, the likelihood probability can be obtained by using the cumulative distribution function of the exponential distribution.

At this time, the current speed may be a speed value detected at the first time with respect to the specific object. Or the current speed may be an average value of the one or more speed values detected before the first time and the first time with respect to the specific object or an average value of the one or more speed values detected before the first time and the first time (T) at the first time, which is obtained by multiplying the velocity values by a predetermined weight and adding them to each other.

According to another aspect of the present invention, there is provided a movement state determination device including a processing unit configured to determine a movement state of a specific object using a particle filter having particles defined as N independent random variables. Here, the processing unit may calculate a current speed for the specific object, and calculate a value relating to a cumulative probability that the specific object can have the current speed; Repeating the particle update process for updating each of the particles by a predetermined number of times according to a predetermined rule; And a step of calculating an average value of the values of the N particles having been updated and determining whether the specified object is in a stop state or a moving state based on the average value. In this case, a weight used to update the particle value is calculated using the cumulative probability.

According to another aspect of the present invention, there is provided a method of determining mobility, comprising: calculating a velocity using a computer by calculating distances between two points using information including collected time, latitude and longitude information; Generating N particles; Obtaining a weight (W) using a velocity Z as a measurement value; Updating a probability value of the particles; And determining whether the particle is in a stationary state or a moving state by using an average of the probability values of the particles.

According to another aspect of the present invention, there is provided a mobility determining apparatus comprising: a velocity measuring unit; And a processing unit. Here, the processing unit may include calculating a speed by calculating distances between two points using information including time, latitude and longitude information collected through the processing unit; Generating N particles; Obtaining a weight (W) using a velocity Z as a measurement value; Updating a probability value of the particles; And determining whether the particle is in a stationary state or a moving state by using an average of probability values of the particles.

According to another aspect of the present invention, a computer-readable medium provided with: a velocity measurement unit; Calculating a speed by calculating a distance between two points using information including time, latitude, and hardness information collected through the processing unit, by the processing unit of the computer apparatus including the processing unit; Generating N particles; Obtaining a weight (W) using a velocity Z as a measurement value; Updating a probability value of the particles; And determining whether the particle is in a stationary state or a moving state by using an average of the probability values of the particles.

In this case, the step of calculating the velocity may include a step of obtaining the current velocity and the immediately preceding time data by applying it to a haversine formula.

At this time, an exponential distribution can be used when the weight W is obtained.

At this time, when the weight is obtained, P_z, which is a probability when the velocity obtained through the cumulative distribution function of the exponential distribution is Z, can be used.

According to the present invention, the movement state of the position data can be accurately analyzed by using the particle filter.

FIG. 1 is a flowchart of a moving state determining method according to a first embodiment of the present invention.
FIG. 2 is a graph showing a value according to the number of position information according to the second embodiment of the present invention.
3A is a diagram showing position information of a vehicle movement on a map according to a second embodiment of the present invention.
FIG. 3B is a velocity-probability graph for the position information of the vehicle according to the second embodiment of the present invention. FIG.
4A is a view showing location information on a map at a highway rest area according to a second embodiment of the present invention.
FIG. 4B is a velocity-probability graph for the position information staying in a highway rest area according to the second embodiment of the present invention. FIG.
5A is a view showing position information of a vehicle movement on a map according to a second embodiment of the present invention.
FIG. 5B is a velocity-probability graph for the position information of the vehicle movement according to the second embodiment of the present invention. FIG.
FIG. 6 shows the still state and the moving state on the map using the kml viewer.
FIG. 7 is a diagram for classifying data of a section that moves smoothly without traffic congestion into a movement state and a stop state.
FIGS. 8 and 9 are graphs showing a specific time stop, a velocity at the time of movement, and a probability value obtained from the particle filter in the entire data of FIG.
10 is a flowchart illustrating a method for determining mobility using a computer according to an embodiment of the present invention.
11A is a diagram showing one observation data on the vehicle moving speed on a map.
FIG. 11B is a map showing information on the average value of the particles obtained from the map according to an embodiment of the present invention, along with the observation data on the vehicle moving speed according to FIG. 11A.
12A is a map showing another observation data on the vehicle moving speed.
FIG. 12B is a map showing information on the average value of the particles obtained from the map according to the fifth embodiment of the present invention, together with one observation data on the vehicle traveling speed of FIG. 12A.
Figures 13A and 13B show the results obtained from an experiment to vary the window size.
FIG. 14 is a table summarizing the experimental results using a plurality of position data for window sizes 5 to 10. FIG.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, the present invention is not limited to the embodiments described herein, but may be implemented in various other forms. The terminology used herein is for the purpose of understanding the embodiments and is not intended to limit the scope of the present invention. Also, the singular forms as used below include plural forms unless the phrases expressly have the opposite meaning.

&Lt; Embodiment 1 >

FIG. 1 is a flowchart of a moving state determining method according to a first embodiment of the present invention.

In this embodiment, the moving state of a specific object can be determined using a particle filter having N independent random variables (Particle [k]) as particles. Here, k is an integer of 0 to N-1. Each of the particles has a value of 0 or more and 1 or less. The value of each particle may be a probability value.

In the moving state determining method, at step S110, one or more velocity values detected before the first time and the first time may be provided to the specific object.

Then, in step S120, the time weighted-rate (T) at the first time may be calculated using one or more velocity values detected before the provided first time and the first time.

Then, in step S130, a value Pt / C relating to the cumulative probability that the specific object may have the time weighted-velocity T can be calculated.

Then, in step S140, the particle update process for updating the value of each of the particles may be repeated a number of times according to a predetermined rule. At this time, the particle update process may include: (1) calculating a weight W by subtracting a value (Pt / C) for the cumulative probability from the value of each particle; and (2) And updating the value of each particle by subtracting the weight W from the value of the particle (S142).

Then, in step S150, the average value X of the N particles that have been updated can be calculated.

Then, in step S160, the average value X may be used to determine whether the specific object is stopped or moving.

&Lt; Embodiment 2 >

The position data error results in a velocity value error and therefore results in a motion state error. Therefore, in the second embodiment of the present invention, past velocity values can be used. Past velocity values have a weighted value for each timestamp. The more recent the velocity value, the greater the effect can be obtained.

In one embodiment of the present invention, the number of velocity value samples included in the window to create a 'velocity value window' may be referred to as a 'window size'. Thus, the window size may have any natural number value. In this embodiment, the window size is set to 5. That is, the following equation (1) is for this embodiment, and it can be confirmed that five time stamps are used.

In the present embodiment, so-called 'time weighted speeds (T)' generated using the following Equation (1) are used.

[Equation 1]

T represents a weighted average of time weighted speeds. The symbol a shown in Equation (1) represents a rate for determining a weight for a velocity value at each time stamp. alpha can have a value of 0 or more and 1 or less. In Equation (1), a total of five velocity values are used. The T value can be derived using one current speed value and four past speed values. In the algorithm used in this embodiment, the particle filter can use the above-described T as the weight of the particles. In the present invention, 'time weighted rate (T)' may be referred to as 'time weighted-rate (T)'.

If the current time in Equation 1 is t, then the current velocity is _Vt and the past four velocities are _Vt-1 , _Vt-2 , _Vt-3 , and _Vt-4 . V _t-1 is closer to current than V _t-4 . If α is set to 0.0, then only the current velocity is used in equation (1). If? is set to 1.0, then only the past velocity value V _t-4 is used in equation (1).

Algorithm 1 shown below is an algorithm for classifying a moving state using a particle filter used in the present embodiment.

[Algorithm 1]

Algorithm 1 is based on the basic algorithm of particle filter. It is composed of particle creation and initialization (line 6), weight update and particle update (line 9 & line 10), iteration (while ~ end while) ). In algorithm 1, the number of particles as the input value and the velocity value of the four position information of the previous time adjacent to the current position data are received. Also, the moving state is derived from the resultant value. It is called once every time the location information is input. it is determined that the X value returned from the line 15 is a 'moving' state when the value of X is equal to or greater than a predetermined value (or less), and the state is determined to be in a 'stopped' state when the value is less than (or exceeds) the predetermined value.

More specifically, at line 16 of algorithm 1, it is possible to determine which of two or more 'motion states' belongs to which motion state. At this time, the motion state can be divided into a plurality of states including a stop state and a movement state according to one or more transportation means. The moving state according to the one or more means of transportation may be a concept including a moving state by walking, a moving state by a bicycle, a moving state by a vehicle, and a moving state by a high-speed train. The determination of which state among the plurality of states can be performed by comparing an average value of the N particles that have been updated to one or more predetermined threshold values.

The operation of Algorithm 1 is as follows:

Step 1) The current velocity (V _t ) is obtained (line 2) using the position information at the current time using the harbor sign formula.

Step 2) The time weighting rate T is calculated using Equation 1 (line 4).

Step 3) After generating N particles ( Particle [N] ), initialize (line 6).

Step 4) The weight W is obtained by using the following Equation 2 and Equation 3 (line 9).

Step 5) Update the values of the N particles (i.e., probability values) (line 10).

Step 6) Repeat step 4 and step 5 above according to the conditions (line 7).

Step 7) The average of the particles (i.e., the average value of the probability values) is obtained (line 14), and the final movement state is determined (line 15).

&Quot; (2) "

&Quot; (3) "

W =

The reason for using the while statement in Algorithm 1 (line 7) to update the N particles many times is that the probability of the particles is expected to be stable without changing much. Therefore, it can be repeated a constant number of times. Since the value of each particle is a probability value, it can be adjusted so that it does not deviate from the value of 0 to 1 interval, for example, when the value is negative in the updating process, it is changed to 0.

The four speed values, which were used to determine the current speed value calculated by using the harbor sign formula and the moving state of the position information of the previous time, a total of 5 speed values, can be applied to the time weighting speed formula according to the time sequence . Therefore, in this embodiment, the window size of the time weighted speed is 5. At this time, because the current time information and the five time information values of the previous time are used together, the state determination can be performed using only the speed of the current position information without applying the time weighting speed until five pieces of information are collected . Therefore, when the position information is inputted, the information inputted from the first time to the fourth time can judge the state using the current speed. Also, in actual experiments, five speed values can be separately stored as the latest speed values each time the position information of the next time is input.

We used the results of the study that the distribution of human movement speed follows the exponential distribution in order to express T, which is a kind of speed value, in probability. The results of this study can be found in the following paper 5. It is also possible to modify the present invention by using this distribution if other distributions can be used as the distribution of the human moving speed.

However,

HY Song and JS Lee, "Finding probability distributions of human speeds," in AMBIENT 2014, The Fourth International Conference on Ambient Computing, Applications, Services and Technologies, 2014, pp. 51.55.

The T value can also be regarded as a kind of movement speed since it is a value obtained by using the speed values. Therefore, the T value can be applied to the exponential distribution. Therefore, when calculating the weight, we can use P _t , which is the probability that the velocity obtained through the exponential distribution is T. In order to have a larger value of the probability P _{t as the} velocity increases, the P _t value can be obtained by using the cumulative distribution function of the exponential distribution. The method of obtaining P _t is shown in Equation (2).

The equation used to calculate the weight in Algorithm 1 is shown in Equation 3 above.

First, in Equation (2), T is substituted into the cumulative distribution function of the exponential probability distribution to obtain P _t . In this case, λ is the reference value of the result in the paper 3, and it can be determined as a constant value of 0.15949. λ = 0.15949 means that both the mean and the standard deviation in the velocity distribution are about 6.27 m / s.

Next, the weight can be obtained by substituting P _t into the equation (3). In Equation 3 C is a P _t is too large, the weight is constant to be adjusted P _t is a value lowered to prevent the coming out of a negative number.

Finally, the final moving state determination can determine whether the moving state is a static state or not using the average (X) of the probability values of the particles. This is a simple conditional statement that can be compared to a specific constant of probability. This constant value is a value between 0 and 1, and can be determined by a value that is deemed appropriate through experiments.

The inventor has actually conducted the experiment using the above Algorithm 1.

Experiments were carried out by using location information, speed calculation using position information, applying velocity value to time weighted velocity, input velocity value obtained from time weighted velocity into particle filter algorithm, And graphical representation.

- Location information collection: The location information used in this experiment was collected using the location information application installed on the smartphone. In addition, time, latitude, and hardness values were extracted from this location information and used as the input value of the experiment.

- Calculation of speed using position information and harbor sign formula: You can calculate the distance between two points and calculate the speed using two position information. At this time, the time, latitude, and longitude information inputted at the first time may be impossible to calculate because there is only one point of information. Therefore, the position information entered first was held for judgment and the second information was input. After that, the speed and state of the second position are obtained, and the same speed and state are entered in the first position. This is because the recorded time difference between the first and second positions is less than a few seconds, so the probability of a state change within a short time may be low. Therefore, the speed of the input location information after excluding the first location information is obtained by using the Harbor sign formula.

At this time, in order to obtain the time weighting rate (T) using Equation (1), it was necessary to find an appropriate? Having the highest accuracy. For this, accuracy was obtained by increasing the value of α by 0.1 from 0 to 1. A total of 3,371,577 positional information were moved, and the accuracy was changed according to the constant α. In this case, when α is 0.3, the best accuracy is obtained because 2,231,090 position information results are correct. As the α increases from 0.0, the accuracy increases, and it is found that it falls off from 0.3. Also, when the constant α has a value of 1.0, the accuracy is significantly lowered. This means that the current state of the current location information is determined using only V _t-4 , and the current state and the velocity value of the fourth previous time are not related to each other. Therefore, the α value was increased by 0.01 unit between 0.2 and 0.4, and the α value at the highest accuracy was found.

FIG. 2 is a graph showing a value according to the number of position information according to the second embodiment of the present invention. In this case, the X-axis represents the number of positional information estimated to correspond to the movement state classification, and the Y-axis represents the alpha value used in the time-weighted velocity expression. It can be seen that the highest accuracy is obtained when α is 0.3.

At this time, the accuracy of the experimental results can be confirmed by the speed of the position information. Basically, if the speed is zero, it is a stop, otherwise it is a move. Therefore, it is possible to indicate the accuracy of the determination of the movement state of the entire information by counting the number of pieces of information in which the speed is 0 and the information in which the speed is greater than 0 is determined as the movement.

FIG. 3A is a map showing information on the average value of the particles obtained from the map according to one embodiment of the present invention together with one observation data on the vehicle moving speed. In FIG. 3A, a latitude and longitude information is used to indicate the moving state and the stop state on the map, and the location mark is displayed. At this time, the stopped state represents the asterisk place mark 100, and the moving state represents the circle place mark 200. [ Each place mark indicates location information made of latitude and longitude recorded by time. At this time, the window size is 5 and the constant α value is 0.3.

FIG. 3A shows the data obtained from 14:00:00 to 14:00:02 on January 31, 2014.

FIG. 3B is a velocity-probability graph for the position information of the vehicle according to the second embodiment of the present invention. FIG. The average of the probability of particles of position information at each time and the velocity at each point are shown. The X axis of the graph represents time, the left Y axis represents the velocity (m / s), and the right Y axis represents the probability value. In the graph, the stop and move states can be distinguished by using different notations for the respective position information determined as the moving state and the stop state in FIG. 3A. Accordingly, marks indicating the speed are marked differently depending on the state. The mark denotes the velocity value of the position information determined to be in the moving state, and the * mark denotes the average of the probability of the particles of the position information determined to be moving. The mark is an average of the probability of particles of the position information determined to be stationary, and the inverted triangle mark represents the velocity of the position information determined to be stationary.

In FIG. 3B, it can be seen that all probability values are mainly distributed between 0.4 and 1.0. The reason for this is that the initial weight of particles is initialized to 0.5, and if the velocity is high, the weight is increased. In the process of updating the weight to the probability value of the particle, if the particle is smaller than 0, It is because it passes. As a result, the average value may not drop much from 0.5, even though it has a large speed.

Figure 4a is a map showing information on the average value of the particles obtained therefrom in accordance with one embodiment of the present invention, along with one observation data of the observed traveling speed when the vehicle stays in the highway rest area;

The window size, alpha value, experiment time, and placemark information of FIG. 4A may be as shown in FIG. 3A. FIG. 4A shows the data obtained from the time period from 21:45:00 to 21:47:00 on November 07, 2014.

FIG. 4B is a velocity-probability graph for the position information staying in a highway rest area according to the second embodiment of the present invention. FIG. The X axis, Y axis, and mark information of the graph in FIG. 4B may be as shown in FIG. 3B. As can be seen from the graph, the velocity values at this time are all close to zero, and all of them can be regarded as accurate results determined to be stationary.

5A is a map showing information on the average value of the particles obtained therefrom in accordance with an embodiment of the present invention, along with other observation data on the vehicle traveling speed. The window size, alpha value, experiment time, and placemark information of FIG. 5A may be as shown in FIG. 3A.

FIG. 5A shows the data obtained from the time period between 11:46:00 and 11:48:00 on November 07, 2014.

FIG. 5B is a velocity-probability graph for the position information of the vehicle movement according to the second embodiment of the present invention. FIG. The X axis, Y axis, and mark information of the graph in FIG. 5B may be as shown in FIG. 3B. As can be seen from the graph, the speed is decreasing in the forward and backward time and the speed is increasing in the backward time based on the position information determined to be stopped. Type.

In the present invention, the reason for performing the state determination using the position information of the time for determining the current state and the four velocity values of the previous time, total five velocity values is that, This is because there is almost no difference from the experiment using 5 samples. Also, the speed of time that affects the current speed should not be too far away. The reason why the probabilities of using the more information of the number of positions and the information of the five pieces of information are not substantially different is that it is necessary to lower the weights at the distant time speeds. Therefore, in the process of obtaining the time weighted speed values, It becomes a very small value as it multiplies, and it hardly affects the speed calculation value used to determine the final state.

&Lt; Third Embodiment >

In the third embodiment of the present invention, GPS data obtained using an application program called Sportstracker installed on a smartphone is extracted and used as an input value of the experiment. The experiment used data for one day. The GPS data may be obtained by other methods.

Algorithm 2 using the particle filter used in this embodiment is described below.

[Algorithm 2]

Algorithm 2 uses the basic algorithm of particle filter, which consists of particle creation and initialization, weight update and particle update, iteration, and result extraction. In Algorithm 2, the current velocity value is denoted by Z, not Vt. Algorithm 2 differs from Algorithm 1 in that it uses the current velocity value Z itself.

Initialization allows you to assign a probability value between 0 and 1 so that the particles have different values from each other.

The time, latitude, and longitude information entered at the beginning is calculated by calculating the distance between two points. Since there is only information of one point, speed calculation is impossible. Therefore, the first input information can hold the judgment, obtain the speed and state of the second position after receiving the second information, and apply the same speed and state to the first position information. This is because the time difference between the first input and the second input is less than a few seconds, and the probability of occurrence of the state change within a short time is low. Also, the speed after the second GPS information can be obtained by using the haversine formula.

According to the algorithm 2, first, N particles are generated (line 03), a weight W is calculated using a velocity Z as a measurement value (line 07), and a probability value of particles is updated (line 09) .

We used the results of the study in the paper 6 that the distribution of the human movement speed follows the exponential distribution in order to express the velocity Z as probability when calculating the weight (W).

However,

Lee, Jun - Seok, Song, Ha - Yun "Estimation of the distribution of human movement speed", Spring Conference of 2014, Vol. 2014.

That is, we use P_z, which does not use velocity Z when weights are calculated, but probability of Z obtained through exponential distribution. To obtain the higher probability P_z as the velocity increases, the P_z value can be obtained using the cumulative distribution function of the exponential distribution.

f (Particle [k], Z) is a simple formula of constant product and arithmetic operations using probability values obtained from particle values and velocities.

According to the method of determining the moving state of a human according to the third embodiment of the present invention, the final state determination can determine whether the current state is a stop state or a moving state using an average of the probability values of the particles of the corresponding data (line 13).

FIG. 6 shows the still state and the moving state on the map using the kml viewer. The asterisk place mark 100 is used for the stop state, and the circle place mark 200 is used for the movement state. The place mark indicates one piece of data consisting of latitude and longitude recorded by the time obtained from the Sportstracker.

FIG. 6 is a flowchart illustrating a procedure of moving from a first place (e.g., Seoul Gwanak-gu) to a third place (e.g., Gyeonggi-nam) after staying in a second place Data.

FIG. 7 is a diagram for classifying data of a section that moves smoothly without traffic congestion into a movement state and a stop state. In FIG. 7, it can be seen that there is a large moving state and a small stop state.

Figs. 8 and 9 are graphs showing the specific time stop, the velocity at the time of movement, and the probability values obtained from the particle filter in the entire data of Fig. Data for 2 minutes each were displayed. FIG. 8 shows data obtained from 12:55:00 to 12:57:00 of January 31, 2014, and FIG. 9 shows data obtained from 14:00:00 to 14:00 of January 31, 2014 Data obtained from the time period up to 2 minutes 0 seconds.

Referring to FIG. 8, it can be seen that the position data determined to be stationary are maintained for a predetermined time. This is also true of the moving state. Generally, it is thought that the result is a correct result because the movement state of a person changes from a stop to a movement in a short time, and the state is not maintained for a certain time and is rarely changed to a stop. Since the particle filter reflects the value obtained by using the velocity and the exponential distribution, the relationship between the velocity and the probability shows that the velocity of the stationary state is lower than that of the moving state. However, when the probability value is high, Can be seen. The velocity of the stationary state is distributed below 5m / s, and the velocity of the moving state is represented by various values. Also, when the state changes from the stationary state to the moving state, the velocity of the moving state shows an increasing distribution, and the velocity decreases when the state changes from the moving state to the stationary state.

8 and 9, it can be seen that probability values are mainly distributed between 0.4 and 1.0. The probability of the moving state is much in the vicinity of 0.4, and the probability of the stationary state is relatively close to 1.0.

Referring to FIG. 9, it can be seen that there is a moving state between stopping states in the vicinity of 14: 1: 30 seconds. It is judged to be the result of sudden increase in speed. It can also be seen that there is one stationary state between moving states.

When the latitude and longitude of the GPS data are recorded at the wrong position, the velocity value is incorrect, so when calculating the weight, the incorrect velocity value should have a small influence on the weight. Therefore, when X _t is obtained for error judgment, it is possible to obtain a more accurate result by using the five peripheral position data as well as the velocity value, moving it like a window, and checking the surrounding speed. The embodiment in which such an idea is reflected is Embodiment 2 described above.

10 is a flowchart illustrating a method of determining mobility using a computer according to a third embodiment of the present invention.

In step S11, a speed Z is calculated by calculating distances between two points using information including collected time, latitude and longitude information.

In step S12, N particles are generated.

In step S13, the weight W is obtained using the velocity Z, which is a measurement value.

In step S14, the probability value of the particles is updated.

In step S15, an average of the probability values of the particles is used to determine whether the particle is in a stationary state or in a moving state.

The fourth embodiment is a modified embodiment of the second embodiment described above. The algorithm using the particle filter used in this embodiment is described in the following algorithm 3.

[Algorithm 3]

The operation of Algorithm 3 is as follows:

Step 1) The current speed Vt is obtained by using the position information at the current time using the harbor sign formula (line 4).

Step 2) The time weighting rate T is calculated using Equation 1 (line 6).

Step 3) After generating N particles (Particle [N]), initialize (line 8).

Step 4) The weights W are obtained by using Equations 2 and 3 (line 11).

Step 5) Update the particle value (i.e., probability value) (line 12-16).

Step 6) Repeat steps 4 and 5 above (line 10) to update the values of the N particles.

Step 7) Repeat the second for statement (line 10) 5 times (line 9). The number of times may be set to a different value.

Step 8) Find the average of the particles (i.e., the average value of the probability values) (line 19) and determine the final movement state (line 20).

In Algorithm 3, the constant 5 is used to repeat N particle updates. The reason for this constant iteration is to stabilize the probability values. On the other hand, common SIS algorithms only update particle values once.

In the step of generating and initializing N particles of Algorithm 3, the average value of the particles may be 0.5.

The condition of the 12th line in Algorithm 3 is two cases when the value of the corresponding particle is less than or equal to 0.5, the weight is greater than 0.5, and the particle value is greater than or equal to 0.5 and the weight is less than 0.5. This is a way to make the particle's value close to 0 or 1. If the weight of a particle with a value less than 0.5 is large, it is subtracted from the weight, and if it is small, the weight of the particle can be added up to 1. If the weight of a particle having a value larger than 0.5 is small, the weight is subtracted from 0, and if the weight is large, the weight can be added to 1. The above contents are shown in Table 1 below.

Condition Particle <= 0.5 Particle> 0.5 Weight <= 0.5 line 15 (+) line 13 (-) Weight> 0.5 line 13 (-) line 15 (+)

After the particle is updated, it is repeated, and finally, the moving average X is obtained by using the average of the probability of the particles. For example, if the average value of these probabilities is less than the experimental value 0.75, it can be judged as the movement, and the rest can be judged as the stop.

The remaining steps except the above can be the same as the algorithm 1.

11A is a diagram showing one observation data on the vehicle moving speed on a map.

FIG. 11B is a map showing information on the average value of the particles obtained from the map according to an embodiment of the present invention, along with the observation data on the vehicle moving speed according to FIG. 11A.

One portion of the contiguous data in FIG. 11B shows low speed and high stop probability, while the other portion shows high speed and high travel probability. This shows a typical stopping and moving situation of a vehicle. In the vicinity of 01:20:13 of FIG. 11B, five result values are classified as stop, and the resultant value corresponds to the asterisk place mark in the lower left portion of FIG. 11A. The part indicates a situation where the vehicle makes a U-turn, has a relatively low speed value, and has a large time difference. It can be seen that the vehicle accelerates around 01:20:53.

The fifth embodiment is a modified embodiment of the third embodiment described above. The algorithm using the particle filter used in this embodiment is described in the following Algorithm 4.

[Algorithm 4]

Algorithm 4 is a modification of Algorithm 2. The input value of Algorithm 2 is the number of particles (N) and the current velocity (Z), but the input value of Algorithm 4 differs in the number of particles (N) and average velocity (Z _avg ). For example, the average speed means the average of the speed values of the current position data and the previous four time values. That is, it means the average of the five speed values in total. This is used to compensate for inaccurate moving conditions when an incorrect velocity value is obtained due to an error in the position data. Therefore, when determining the movement state of the current position data, not only the current velocity but also the velocity value of the previous data can be reflected. In this case, the data before the five velocity values are collected is not input to the second input value of the particle filter algorithm at the average velocity, but the velocity of the current position data can be input.

In Algorithm 2, the calculated weight is always subtracted from the previous particle to update the particle. In Algorithm 4, the particle is updated by subtracting or adding according to the condition.

The condition in the ninth line of Algorithm 4 is that when the value of the corresponding particle is less than or equal to 0.5, the weight is greater than 0.5, and the value of the particle is greater than or equal to 0.5 and the weight is less than or equal to 0.5 . This is a way to make the particle's value close to 0 or 1. If the weight of a particle with a value less than 0.5 is large, the weight is subtracted to make it close to 0, and if the weight is small, the weight of the particle can be added up to 1. If the weight of a particle with a value greater than 0.5 is small, the weight is subtracted to make it close to 0, and if the weight is large, the weight is added to 1. The above contents are shown in Table 1.

After the particle is updated, it is repeated, and finally, the moving average X is obtained by using the average of the probability of the particles. If the average value of these probabilities is smaller than the experimental value of 0.75, it can be judged as the movement, and the rest is judged as the stop.

The remaining steps except the above can be the same as algorithm 2.

12A is a map showing another observation data on the vehicle moving speed.

FIG. 12B is a map showing information on the average value of the particles obtained from the map according to the fifth embodiment of the present invention, together with one observation data on the vehicle traveling speed of FIG. 12A.

12B shows a section of the date corresponding to FIG. 12A as a velocity-probability graph. The mark information used in FIGS. 12A and 12B may be the same as FIG.

9 and 12B are compared with each other, FIG. 9 uses only the current speed, and FIG. 12B uses the average speed. In Fig. 12B, all the data are judged to be moved, whereas the data judged to be in the stop state exists in Fig. In FIG. 12B, it is understood that the problem that the state continuously changes within a short time has been solved. In addition, in FIG. 12B, the first three data at the beginning are determined to be the movement, which is considered to be the influence of the speed of the previous time since the average speed is used.

The sixth embodiment is a modified embodiment of the second embodiment described above. The algorithm using the particle filter used in this embodiment is described in the following Algorithm 5.

Algorithm 5 further includes a parameter value calculation step that considers the human movement velocity distribution. The exponential distribution is a representative distribution related to the speed of human movement. For a full set of human movement speeds, the parameter for the exponential distribution representing the human movement velocity may be considered to be a constant, but this parameter may vary for some velocity values. If a set of consecutive low velocity values is given, the low velocity values tend to be determined as a stationary state with the parameter of the constant. Therefore, this parameter value must be recalculated for all inputs of the new velocity value in one window. The tendency of the velocity values can then be reflected as a probability distribution for the particle filter. For all observed human movement velocity values, the parameter l for the human velocity distribution has a value of 0.15949, which means that both the expected value and the standard deviation for the human velocity are all 6.27 m / s. However, human walking speed is usually less than 5 m / s, so the parameter λ should be recalculated in this case. Therefore, the standard deviation σ of velocity values in one window can be calculated, and the reciprocal of σ can be used as λ. At this time, if all velocity values in one window are 0, λ should have the maximum value for the actual calculation. Then, the position parameter γ must be calculated to resolve the difference between the expected value μ and the standard deviation σ (of the velocity values within the window). The position parameter? Can be calculated by the following equation (4).

&Quot; (4) "

[Algorithm 5]

The operation of algorithm 5 is as follows:

Step 1) Get new location information (line 1)

Step 2) The current velocity (Vt) is obtained (line 2) using the new position information using the harbor sign formula.

Step 3) The time weighting rate T is calculated using Equation 1 (line 3).

Step 4) Calculate lambda using lambda window (line 4).

Step 5) After generating N particles (Particle [N]), initialize (line 5).

Step 6) A weight W is obtained by using the following equations (5) and (6) (line 8).

Step 7) Update the value of the particle (that is, the probability value) (line 9-13).

Step 8) Repeat Step 5 and Step 6 above to update the values of the N particles (line 7).

Step 9) Repeat the second for statement (line 7) 5 times (line 6). The number of times may be set to a different value.

Step 10) The average of the particles (i.e., the average value of the probability values) is obtained (line 16), and the final movement state is determined (line 17).

&Quot; (5) "

&Quot; (6) "

W =

In Algorithm 5, the constant 5 is used to repeat N particle updates. The reason for this constant iteration is to stabilize the probability value. On the other hand, common SIS algorithms only update particle values once.

The condition in the ninth line of Algorithm 5 is two cases when the value of the corresponding particle is less than or equal to 0.5 and the weight is greater than 0.5, or when the particle value is greater than 0.5 and the weight is less than or equal to 0.5. This is a way to make the particle's value close to 0 or 1. If the weight of a particle with a value less than 0.5 is large, it is subtracted from the weight, and if it is small, the weight of the particle can be added up to 1. If the weight of a particle having a value larger than 0.5 is small, the weight is subtracted from 0, and if the weight is large, the weight can be added to 1. The above description is shown in Table 1 above.

The representation shown on line 10 of algorithm 5 represents particles with negative weights and the representation on line 12 represents particles with positive weights.

If the weights are large enough, small particle values are smaller, but larger particle values are larger. The resulting weight is subtracted to make the smaller value smaller, but the weight is added to make the larger value larger.

On the other hand, for small weights, the larger the particle value becomes smaller and the smaller the particle value becomes larger. As a result, the weights are added to make the small particle value larger, but the weight is subtracted to make the larger particle value smaller.

In Algorithm 5, the particle value and weight reference value for determining whether the particle value is small or large is 0.5.

In the embodiment using Algorithm 5, two time windows are required which are different from each other. One window is for the time weighted-rate (TWS) and the other window is for those parameters of the exponential distribution. Experiments were conducted to determine the optimal window size. As a result, the window size should be as small as possible, but large enough to stabilize the time weighted-rate and λ. In the above experiment, we changed the window size from 1 to 20.

Figures 13A and 13B show the results obtained from an experiment to vary the window size. The x-axis represents the window size and the y-axis represents the time weighted-velocity and the value of λ. Symbol & cir & denotes a time weighted-speed. Figures 13A and 13B show a simple reduction in [lambda], decreasing to 0.15949. Additionally, a window size greater than 5 can be used to stabilize the time-weighted-rate. 13B shows a stabilized time weighted-velocity with window size 6.

FIG. 14 summarizes the experimental results using 4,786,484 positional data for window sizes 5 to 10. In Table 2,? Of Equation (1) is presented according to a given combination of window sizes. Then the window size for the calculation of lambda is set to 8 and the window size for the time weighted-velocity can be set to 6, where a is 0.72.

Although the present invention has been described using the embodiment for acquiring the positional information by using the GPS device for convenience, the positional information obtained by using the positioning apparatus or the system other than the GPS is described in various embodiments You can understand what is applicable to the example.

It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the essential characteristics thereof. The contents of each claim in the claims may be combined with other claims without departing from the scope of the claims.

Claims

A method of determining a motion state of a specific object based on an average value of particles obtained by performing a particle update process for updating values of particles defined as independent random variables,
Calculating one or more parameters for a first type probability distribution based on velocity values belonging to a first time window having a magnitude according to a predetermined rule; And
Calculating a value with respect to a cumulative probability that the specific object will have a certain rate, using the calculated one or more parameters;
/ RTI >
And updates the value of the particle by using the cumulative probability.
How to determine exercise status.

The method according to claim 1,
The particle update process includes:
Calculating a weight by subtracting the value of the cumulative probability from the value of each particle; And
Updating the value of each particle by subtracting the weight from the value of each particle or by adding the weight to the value of each particle;
/ RTI >
How to determine exercise status.

The method according to claim 1,
The specific speed is,
(T) calculated by multiplying one or more velocity values belonging to a second time window having a size according to a predetermined rule by a predetermined weight,
A specific velocity value detected for the specific object, or
An average value of one or more velocity values detected for the specific object
And determining a motion state of the subject.

4. The method according to claim 3, wherein when the time weighted-velocity is calculated, the value of the weighted value multiplied by the velocity value obtained more recently becomes larger.

The method according to claim 1,
Wherein the motion state is divided into a plurality of states including a stop state and a movement state,
Wherein the determination of which state among the plurality of states is made by comparing an average value of the values of the particles with a predetermined threshold value or more,
How to determine exercise status.

The method according to claim 1,
The first type probability distribution is an exponential distribution,
Wherein the one or more parameters comprise:
An inverse number (?) Of a standard deviation (?) Of velocity values belonging to the first time window; And
(?) Of velocity values belonging to the first time window and a position parameter (?) Calculated by the standard deviation,
/ RTI >
How to determine exercise status.

7. The method of claim 6, wherein the cumulative probability is

/ RTI > is calculated on the basis of the value of the motion state.

The method according to claim 1,
The particle update process includes:
Obtaining the specific speed;
Updating a likelyhood probability even if the specific object has the specific velocity based on a current value of the particles;
Obtaining a weight using the updated likelihood probability; And
Updating the particles using the weight;
/ RTI >
How to determine exercise status.

9. The method of claim 8, wherein the updating of the particles comprises updating the particles with a value of zero if the particles are negative.

6. The method of claim 5,
Wherein the motion state includes a plurality of states including a stop state and a movement state according to one or more means of transportation,
Wherein the moving state according to the one or more means of transportation includes a moving state by walking and a moving state by the vehicle,
How to determine exercise status.

And a processing unit configured to determine a motion state of a specific object based on an average value of the particles obtained by performing a particle update process for updating values of particles defined as independent random variables,
Wherein,
Calculating one or more parameters for a first type probability distribution based on velocity values belonging to a first time window having a magnitude according to a predetermined rule; And
Calculating a value with respect to a cumulative probability that the specific object will have a certain rate, using the calculated one or more parameters;
Respectively,
And updates the value of the particle by using the cumulative probability.
Motion state discrimination device.

A computer-readable recording medium storing a program for causing a computer to execute a motion state determination method for determining a motion state of a specific object based on an average value of particles obtained by performing a particle update process for updating values of particles defined as independent random variables As a medium,
The program may cause the computer to:
Calculating one or more parameters for a first type probability distribution based on velocity values belonging to a first time window having a magnitude according to a predetermined rule; And
Calculating a value with respect to a cumulative probability that the specific object will have a certain rate, using the calculated one or more parameters;
Quot ;, and "
And updates the value of the particle by using the cumulative probability.
Computer-readable medium.