KR20180049648A - Prediction method of object's next position using classification model and a device for the same - Google Patents

Abstract

The present invention provides a model which predicts in real time moving of an object and generates a moving pattern by using a convolution neural network (CNN) which is a deep learning classification model. A method for predicting a moving path comprises the following steps of: collecting geolocation information which is obtained by a specific object when moving, converting a geolocation information set into pass data of an array shape, and generating a plurality of preprocessed passes and a plurality of prediction locations corresponding to the plurality of preprocessed passes from the pass data; learning a prediction model by using the preprocessed passes and the prediction locations; and inputting first pass data of the array shape converted from a first geolocation information set obtained from the geolocation information according to another moving path of the specific object to the learned prediction model, and determining a subsequent prediction location subsequent to the new moving path.

Description

[0001] The present invention relates to a method for predicting a next movement path of an object using a classification model and a device for the same.

The present invention relates to computing technology, and more particularly, to a technique for predicting the next movement path of an object.

According to the development and popularization of smart phones, people who do not have a GPS-specific collecting device can collect location data anytime and anywhere using a smart phone. The development of smart devices such as smart watches also makes it easier to collect location data. In applications using a lot of location data, motion data is checked using the collected location data, recommendation of restaurants recommended near the current location, and moving data are used in other fields such as social science analysis and traffic volume analysis through movement pattern analysis . Therefore, if motion can be predicted from movement data as big data and movement pattern can be generated, it can be used in various other fields such as analysis of social phenomenon or traffic volume using mobile data.

Many existing studies such as the following articles [1], [2], and [3] have used a method of generating a movement pattern tree as a method of generating movement patterns from moving data. There is a method to generate the movement pattern tree and to use the generated tree to mine the movement pattern. There is a study to predict the next location for the new movement using Markov Chains as in [4].

However,

Efficient Spatio-Temporal Moving Pattern Mining (STMPM) using Moving Sequence Tree

Articles [2]

Continuous pattern mining using the FCPGrowth algorithm in trajectory data warehouses

Thesis [3]

Temporal moving pattern mining for location-based service

The paper [4]

Next Place Prediction using Mobility markov Chains

However, as the number of position data becomes larger and larger, there is a demand for a technique capable of generating movement patterns from movement data as big data. Deep run is one of the biggest emerging fields of data processing technology. Deep learning is deepening the neural network field of machine learning and it is a field to learn through data and to derive the result. It is not uncommon that there was research that attempted to create a movement pattern using this deep learning model. In [5], there was a study on the movement pattern generation using Deep Autoencoder, and the movement pattern was generated through this model. However, the movement pattern derived from the paper [5] suggests the usual moving path, and can not suggest the predicted movement in real time.

Thesis [5]

Extracting Typical User's Moving Patterns Using Deep Learning

Deep learning, on the other hand, has started in the field of neural networks among many fields of artificial intelligence. Deep running is the construction of a neural network with a deep structure by stacking several layers of this neural network. Deep run derives its results through parameter adjustment within the model based on the input data. Deep learning is currently used mainly in image classification, speech recognition, and so on.

  Deep learning models are divided into several categories according to their nature. First, a model learned through learning data is classified into a classification model that classifies input data and a generation model that generates new data based on learning data.

  There are several kinds of models in the classification model of deep learning. Class-Deep Belief Network, Feed Forward Neural Network, and Convolution Neural Network. In the above paper [5], the model used in the study of movement pattern extraction using deep learning is the generation model Deep Autoencoder.

CNN is a model that is widely used in many researches and practical use at present and shows good classification performance. It is a model mainly used for image classification.

CNN is a model that classifies characteristics based on learned data about un-learned data after learning data by adjusting parameters between layers through input data. To do this, CNN needs training and test data.

1 illustrates a structure of a CNN according to an exemplary embodiment of the present invention.

The structure of CNN is a combination of convolution layer, pooling layer and FNN as shown in FIG. CNN's learning is as follows. When the data is input, the size of the data is reduced through the convolution layer and the pooling layer of the repeated configuration. This reduced data is input to the Neural Network and the error function is used to adjust the values of the parameters W and b between layers. This trained CNN classifies the images using the adjusted parameters when a new image is input.

In order to solve the above-mentioned problems, the present invention intends to create a model for predicting the movement in real time and generating a movement pattern using the Convolution Neural Network (CNN), which is a deep learning classification model. In order to measure the accuracy of the next position prediction and movement pattern generation using the CNN, we propose an accuracy measurement method and measure the accuracy of the movement prediction and the pattern made by our motion prediction model. That is, we use a deep learning classification model called CNN for the movement prediction and provide a movement prediction model together with a criterion for measuring the accuracy of the experiment.

According to an aspect of the present invention, there is provided a movement path predicting method including: acquiring a set of geolocation information (ex: GPS data) by collecting geolocation information obtained when a specific object moves; (1, 0, 0, 0}) of the path data (ex: {1, 2, 3} - {2}, {1, 2, 0} - {3}, {1, 2, 3, 0} - {0} (31-45) and a plurality (50 in FIG. 12), inputting the plurality of preprocessed paths 31 to 45 and the plurality of predicted positions into a predictive model prepared in advance, and learning the predictive model And a first set of geolocation information (ex: GPS data) obtained from the geolocation information according to the new moving path of the specific object is stored in the first path data (ex: {1, 3}, and inputting the first pass data to the learned prediction model to determine a next predicted position following the new moving path.

Wherein the plurality of preprocessed paths are input to the input layer of the predictive model and the plurality of predictive positions are input to the output layer of the predictive model to learn the predictive model, And determines the information output from the output layer of the learned predictive model as the next predictive position when the input predictive model is input to the input layer of the learned predictive model.

In this case, the plurality of preprocessed paths have an array shape having the same first length, and in the converting step, a plurality of path data are provided, and the first length is the longest of the plurality of path data And is equal to or longer than the length of the path data.

Wherein the plurality of preprocessed paths all have an array type having the same first length and the first pass data is input to the array having the first length to input the first pass data to the learned predictive model Wherein one or more null data (ex: {1, 3} -> {1, 3, 0, 0}) is padded to the first pass data for the conversion can do.

At this time, in the converting step, a plurality of path data are provided, the plurality of path data each have an array shape, and at least two different path data among the plurality of path data have different lengths .

At this time, the plurality of preprocessed paths are each in the form of an array, and some or all of the plurality of preprocessed paths may be generated by padding null information to the information included in the path data.

At this time, the step of converting into the path data may include preparing information on the global map in which the sub-areas to which the labels are assigned are defined, and when arranging the plurality of pieces of the geolocation information according to the flow of time, Sequentially observing sub-regions in which the specific object is located, and generating labels of sub-regions corresponding to the observed results in an array in time order, wherein each of the elements adjacent to each other in the array representing the path data May represent different sub-regions.

In addition, in the converting step, a plurality of the path data are provided, and the number of the plurality of preprocessed paths (31 to 45) corresponds to a value obtained by adding the lengths of the respective arrays representing the plurality of path data to each other can do.

Generating second path data by padding the next predicted position to the end of the first pass data; inputting the second pass data to the predicted predicted model to generate a second subsequent And determining the predicted position.

Further, in the converting step, a plurality of path data are provided,

Wherein the plurality of preprocessed paths have an array shape having the same first length and the first length is equal to or longer than a length of the longest path data among the plurality of path data, The first length may be reset to the smallest value among the first natural numbers having a square of a natural number and larger than the first length.

According to the present invention, the predicted movement of the object can be presented in real time using the learned movement prediction model. In addition, a motion prediction model for generating a real-time motion pattern can be generated.

1 illustrates a structure of a CNN according to an exemplary embodiment of the present invention.
Fig. 2 is a visualization of Definition 1, which shows three-dimensional movement data.
3 shows two-dimensional moving data.
FIG. 4 illustrates an overall map including sub-regions according to an embodiment of the present invention.
FIG. 5 is a table showing an example of preprocessed movement data according to an embodiment of the present invention.
FIG. 6 shows an algorithm of a preprocessing process according to an embodiment of the present invention.
FIG. 7A shows an algorithm of a path prediction method for generating a predicted movement pattern according to an embodiment of the present invention.
FIG. 7B shows an example of a movement pattern prediction according to an embodiment of the present invention using a convolutional neural network.
FIG. 8 is a graph for predicting the actual movement and the movement path during the week of the experimenter 1 according to the embodiment of the present invention.
FIG. 9 shows the above-described experimental results according to an embodiment of the present invention.
FIG. 10 shows the above-described experimental results according to an embodiment of the present invention.
FIG. 11 is a graph for predicting the actual movement and the movement path during the week of the experimenter 3 according to the embodiment of the present invention.
12 is a diagram for explaining a model learning method according to a second embodiment of the present invention.
13 is a flowchart of a model learning method according to a second embodiment of the present invention.
FIG. 14 is a flowchart illustrating a motion path prediction method according to another embodiment of the present invention.

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.

≪ Embodiment 1 >

In the present invention, movement data of seven students were collected on average for one year in order to collect movement data. The device we used to collect was a smartphone and collected GPS data using the Sport Tracker application and categorized the GPS data into weekly and weekly movements. However, since the obtained GPS data can not be used immediately for model training, it is necessary to preprocess raw data and convert it into data suitable for model training.

  Before preprocessing the movement data, it is necessary to define the movement first. In the following paper [6], we define the following definition of movement 1 (Definition 1):

Definition 1. A trajectory or spatio-temporal sequence is a sequence of triples

T = <x ₀ , y ₀ , t ₀ >, ..., <x _n , y _n , t _n >

where ti (i = 0 ... n) denotes a timestamp such that 0 <i <n ti <ti + 1 and (xi, yi) are points in R2.

[6]

WhereNext: a Location Predictor on Trajectory Pattern Mining

FIG. 2 is a visualization of the definition 1, and shows three-dimensional movement data.

3 shows two-dimensional moving data.

FIG. 4 illustrates an overall map including sub-regions according to an embodiment of the present invention.

Referring to FIG. 2, it can be seen that the position changes with time.

That is, since it is a positional change with time, in order to train information about the positional change, the concept of the positional change should be included in the preprocessed data. However, the magnitude of the data representing such a movement change by the coordinates itself as shown in Fig. 3 becomes too large. Therefore, to reduce the size of data, each region is cut into polygons as shown in FIG. 4 and the cut regions are labeled. Therefore, movement is not a coordinate but a change in the labeling of each area.

FIG. 5 is a table showing an example of preprocessed movement data according to an embodiment of the present invention.

In the example shown in FIG. 5, the 'preprocessed movement data' includes six pieces of data. Each column in Fig. 5 shows each data. The 'preprocessed movement data' shown in FIG. 5 is obtained, for example, by preprocessing the movement data indicating the movement path in which the object sequentially passes through the sub-areas 1, 2, 3, 4, 5 and 6 in FIG. The movement data can be expressed as [1, 2, 3, 4, 5, 6].

Each of the six pieces of data included in the 'preprocessed movement data' shown in FIG. 5 has a movement history at that position and a next position as a label. If the position is not defined, it can be indicated as 'null' or '0'. In FIG. 5, it is indicated as '0'.

For example, when the object is in the first sub-region 1, the movement history is expressed as {1,0,0,0,0,0} (= {1, null, null, null, null, , The next position may be indicated by the actual next position '2'. For example, when the object is in the third position, sub region 3, the movement history is represented as {1,2,3,0,0,0} (= {1,2,3, null, null, null}) , The next position may be indicated by the actual next position '4'. For example, when the object is in the sub-area 6, which is the last position, the movement history is represented by {1,2,3,4,5,6}, and the next position is represented by '0' .

The number of data included in the 'preprocessed movement data' is the same as the number of elements of the movement data before preprocessing. For example, the number of 'preprocessed movement data' in FIG. 5 is 6, which is equal to 6, which is the number of elements of the movement data {1,2,3,4,5,6}.

This preprocessing procedure follows Algorithm 1 of FIG.

FIG. 6 shows an algorithm (algorithm 1) of a preprocessing process according to an embodiment of the present invention.

The algorithm 1 creates a preprocessed path (ProcessedPath) which is preprocessed motion data from a path (path data) that is motion data before preprocessing and adds it to the preprocessed path set (ProcessedPath_Set). At the same time, information on the next position of each preprocessed path is received from the path and added to the label. Finally, it returns the preprocessed path (ProcessedPath) and the label (Label) which are data for the map learning of the classification model.

The description of each line is as follows.

Line 1: ProcessedPath is defined as a variable of the N * N matrix, where N is the square root of the largest number of predictions or the longest number of trajectories. For example, N = 3 when having the length of the locus 9.

Line 2: It can be repeated up to Line 8 by the length of the path.

Line 3: Creates a ProcessedPath with a path equal to the length of the path up to Line 5.

Line 6: Add the created ProcessedPath to the ProcessedPathSet.

Line 7: Add the following path to the Label set.

Line 9: Returns ProcessedPathSet and Label.

The data used in the training of the model generally requires a train set and a validation set, and a test set for the experiment. A train set is a set of data that is used to train the model, and a validation set is used to determine the degree to which the model is properly trained as a set to prevent over-summing of model learning. Most of the data is used as train set, and some are used as validation set and test set. In the present invention, the data is composed of three data sets, train set, validation set, and test set.

In the preprocessing step, GPS raw data was processed as moving data for model training. Therefore, when the model is trained with the preprocessed movement data, a model for predicting the next position is generated for the new movement data. The learning of the model is performed by inputting data in the form of learning the existing classification model and adjusting the weight and bias value of each neural network layer which is a parameter of the model. The model of learning generation and learning rate selects learning rates and generations that produce the best results over multiple runs of learning.

After the model is trained, the model predicts the most suitable position based on the input moving data.

7A shows an algorithm (algorithm 2) for a path prediction method for generating a predicted movement pattern according to an embodiment of the present invention. As shown in FIG. 7A, the model predicts the next position for the input movement. Add this predicted position to the end of the first movement, input it back into the model, and predict the next position. If you repeat this behavior, you can predict the next position as you want.

The movement prediction thus made becomes a movement pattern. As a result, the next position is predicted and the movement pattern is generated for the current movement.

PredictPath, which is a function for finding a predicted path, receives Predicted Path, which is movement data for predicting a next location, and returns a Predicted_Path including a predicted Predicted Path. The value of n can be determined according to the length of the movement to be predicted.

Predicted_Label receives the next predicted position according to the current path Path using the Predict function which is a predicted function of CNN, which is a trained model. This predicted position is added to the original path and input to the CNN prediction function again.

It is possible to obtain a predicted movement path by repeating such a motion by a desired movement length desired.

FIG. 7B illustrates an example of a mechanism for the algorithm of FIG. 7A according to an embodiment of the present invention.

Examples of input and output of prediction performed by CNN are shown in FIG. 7B, and input / output values are actual values observed in the experiment. The trajectories of [487, 389, 359, 360] are used as inputs to the CNN and include the remaining vacancies filled with zeros for the predicted positions. If the label for the current position is 360 and the CNN prediction is used, then the most likely position as the next position may be 336. This prediction algorithm can be repeated as many times as desired through algorithm 2. [

In the present invention, a method for measuring the accuracy of the position prediction is devised in order to measure the accuracy of the predicted movement path. In order to measure the accuracy of the predicted position, the next moved position and the predicted position are compared with each other. There are three possible cases. First, when the actual moved position is the same as the predicted position, and secondly, when the actual moved position is predicted, the third position is predicted to be completely different. In the present invention, these three cases are scored differently and added to the total score for the position prediction.

  First, when the predicted position is the same as the actual position, one point is given as a difference point according to the distance when the predicted position is near the actual position, and when the predicted position is completely different from the actual position,

  In the case of the second case, the method of assigning a score is to add the distance between the actual position and the predicted position divided by the maximum distance to the nearby position. This can be expressed by the following equation.

&Lt; Formula 1 >

In this case, MAX _D (p _b , a _b ) means the maximum distance possible, that is, the maximum distance between the widest possible boundary portions at the predicted position of the polygon. p _b denotes the farthest boundary portion with respect to the predicted position of the polygon, and a _b denotes the nearest boundary portion with respect to the actual position of the polygon.

The percentile sum of the total predicted trajectory is shown in Equation 3. In this case, the total number of trajectories in the first case is c ₁ , and the total number of trajectories in the second case is c ₂ . c ₂ can be calculated through Equation (2). In this case, case 2 (p) may be the prediction positions classified in the second case. This may be the total number for adjacent prediction positions classified in the second case.

&Quot; (2) &quot;

&Quot; (3) &quot;

In this case, N _p means the total number of predicted positions.

In the present invention, since all the subjects were collected for the same school students, the model was trained for the movement during the day by selecting the area around the school. The area around the school was cut to 24x24 and each area was labeled from 1 to 576. In order to investigate the performance of next location prediction and movement pattern generation through CNN, we calculated how much movement pattern similar to the actual movement pattern occurred in the manner described in Chapter 4. We compared performance with other classification models to compare the performance of the models. Ubuntu 15.04 for OS and GPU for GTX 980 using Python and theano.

In one embodiment of the present invention, the model was trained using only one person's movement data per experiment in order to generate a movement pattern for each person. Each model was trained for each person's movement pattern so that one model predicted the next position of one person. The prediction length visualizes five consecutive prediction paths starting from the next one position prediction. By training the model, the results were obtained as shown in FIG. FIG. 8 is a graph for predicting the actual movement and the movement path during the week of the experimenter 1 according to the embodiment of the present invention.

 FIG. 9 and FIG. 10 show the above-described experimental results according to an embodiment of the present invention. FIG. 9 shows the prediction accuracy of the seven weekly shifts of a total of seven subjects. Fig. 10 shows the prediction accuracy of the movement of the weekends of seven experimenters. The reason why one person is used for the movement of the weekend is that there is no week shift of the experimenter 5 with respect to the experimental position set in the embodiment. In one embodiment, the data used in the training of the model and the number of data used in the experiment are indicated in the train set and the test set, respectively, for each experimenter. And we showed Misprediction error for each experiment. The misprediction error indicates the error rate of the model trained from the train set for the predicted data of the test set. It can be said that the error rate is the case where the next position is accurately predicted. Therefore, the lower the error rate, the higher the accuracy. Finally, the accuracy of prediction defined by the accuracy measurement method defined by the present invention is shown. The accuracy according to the error rate indicates the degree of accuracy with respect to the precise position. The accuracy measurement method according to the accuracy measuring method of the present invention gives addition points when the vicinity of the predicted position is predicted, .

 As shown in FIG. 9, the accuracy ratio of the prediction to the next one position is about 40 to 69%, and the prediction result of the weekend movement is 16 to 67% If you think about the fact that human movement is not always regular, it can be said that the accuracy is not low.

  If we look at the number and accuracy of experimental data from each experimenter, we can see that the more mobile data, the higher the accuracy. It can be seen that the more mobile data, the finer the learning of the model. When comparing FIG. 9 and FIG. 10, it can be considered that the accuracy of movement during weekdays is similar to that of weekends because the number of data is larger than that of weekends. As a result, Volunteer 1 and Volunteer 2 are more accurate than other experimenters. In other words, it is a general tendency that, when there are many movements, more similar movements are seen, and as the size of data increases, accuracy increases. However, the accuracy of Experiment 6 (Volunteer 6) and Experiment 7 (Volunteer 7) are very different. Although the values of Train set, Validation set, and Test set are similar, there is a large difference in accuracy, while Misprediction ratio is similar. This implies an arbitrary nature of human mobility. Experiment 6 has more arbitrary properties in mobility.

  Depending on the length of the predicted movement path, the degree of accuracy decreases gradually. The reason for this is that when the first movement route is predicted and the actual movement is different, the input of the next location prediction becomes erroneous. Therefore, if the initial prediction fails, the probability of predicting the next location will become lower. As the number of paths to be predicted increases, it is possible to judge only that the experimenter normally exhibits such a tendency to move rather than the meaning of the accuracy of the prediction.

FIG. 11 is a graph for predicting the actual movement and the movement path during the week of the experimenter 3 according to the embodiment of the present invention.

Figs. 8 and 11 are a visualization of the results of Experient 1 and Experiment 3 according to the actual movement and the length of the predicted movement path. 8, it can be seen that the tendency of movement is similar to the accuracy of movement. Therefore, the outline of the overall movement can be obtained even if the accuracy of the travel route becomes lower as the prediction of the travel route becomes longer. However, as shown in FIG. 11, it can be seen that, after the initial position prediction fails, the prediction becomes difficult due to continuous error in the next prediction.

  As a result, it can be said that the prediction of the movement path using the classification model shows both the accuracy of the prediction and the tendency of movement. Therefore, it has been shown that the movement path can be predicted by preprocessing and tracing a human movement path according to the research method of the embodiment of the present invention.

  However, the fact that the movement of a person does not always have a certain pattern makes it difficult to predict a movement of a person. Also, the disadvantage of human movement data is that there is not a lot of movement during the day, so when using the data on a daily basis, there is not enough data to be used for deep running. Errors due to changes in moving or traffic also make model training and prediction difficult. Nevertheless, it showed that this accuracy can be shown.

  We have experimented with human movement in the experiment, but in practical use it will be able to predict movements with even better accuracy for movements with comparable regularity, such as the movement of an automobile or the movement of a moving plane of a freight vehicle see. We have also experimented with the models provided by theano, but we will see better predictions if we do the same experiment using deep-learning libraries and models with better accuracy in the past.

In one embodiment, the movement pattern is predicted using the CNN, which is a classification model of the deep running model, for the movement of the object, and a movement pattern is generated. And we used GPS data, which is movement data, to train in the model by training only the movement information instead of inputting the existing map. Through experiments, it was shown that the movement pattern predicted in the present embodiment generated the movement pattern of the actual person, and the performance of the model according to the accuracy measurement method was measured. The research of the present invention shows that it is possible to generate a movement pattern using a deep running in a new way using a movement itself rather than a map in which movement is performed. I think that the prediction of the next location of this type of object at present where big data is discussed can be applied to many location based services. Especially, it can be applied more because it can predict and send the next location in real time.

&Lt; Embodiment 2 >

12 is a diagram for explaining a model learning method according to a second embodiment of the present invention. 12 (a) shows a whole map partitioned in a matrix form, FIG. 12 (b) shows path data for each path, and FIG. 12 (c) shows a pre- .

13 is a flowchart of a model learning method according to a second embodiment of the present invention.

12 and Fig. 13 together.

In step S10, GPS data (= district location information) that a specific person can obtain in one movement can be collected and a 'GPS data set (= district location information set)' can be acquired. At this time, the specific person may be referred to as an 'object' or an 'experimenter'. The GPS data may be <latitude, longitude, time>. The number of GPS data that can be obtained in the single movement may be very large.

The 'GPS data set' may mean a set of GPS data obtained by moving a plurality of times.

In step S20, the 'GPS data set' can be converted into data having another type of data type.

In step S21, a whole map including a plurality of sub-areas y1x1 to y7x7 partitioned in a matrix form can be prepared. For example, the whole map may be a map including an area where the specific person moves.

The entire map can be divided and divided into a predetermined matrix form. And each of the divided sub regions may be labeled. For example, sub-regions can be distinguished in the form of a 7 * 7 matrix. Each sub-area may be assigned a label, for example, 1 to 49. [ A label of 1 may be assigned to the sub area y1x1, and a label of 7 may be assigned to the sub area y1x7.

In step S22, sub-areas located on the movement path of the specific person are determined by using the GPS data set to form an array (Array) having the labels of the sub-areas sequentially as its components, The formed array can be defined as path data.

For example, a movement path corresponding to the acquired GPS data set may be associated (e.g., drawn) on the global map.

The path data corresponding to the GPS data set can be generated as follows.

The array formed by sequentially arranging the sub areas that the specific person has passed through is used as the Path. (Actually, when processing by computer, the movement route of a specific person is represented by GPS data, so that the above-mentioned operation can be processed by using this data.)

For example, if a particular person sequentially passes through the sub-areas corresponding to labels 1, 2, 9, and 10 among the sub areas in the first movement experiment, the first pass data Path (1) 10}. In this case, the size of Path (1) is 4.

For example, if a particular person sequentially passes through the sub-areas corresponding to labels 1, 2, 9, 10, 11, and 18 among the sub areas in the second movement experiment, the second path data Path (2) 2, 9, 10, 11, 18}. At this time, the size of Path (2) is 6.

For example, if a particular person sequentially passes through the sub-areas corresponding to the labels 2, 3, 4, 11, and 12 among the sub areas in the third movement experiment, the third path data Path (3) 4, 11, 12}. At this time, the size of Path (3) is 5.

In step S30, the obtained 'path data' may be preprocessed to generate a 'preprocessed path set 30'.

At this time, a plurality of prediction positions 50 corresponding to the plurality of preprocessed passes 31 to 45 may be generated. Each preprocessed path is provided in an array form, and the lengths of the arrays (Array_P ()) may be equal to each other.

The length of the array (Array_P ()) may be equal to or greater than the size N of the array (Array ()) representing the path data. However, when the size of the array (Array ()) representing the obtained path data is N, the preprocessed data set may include N components.

For example, the first path data Path (1) has an array (Array ()) size of 4, and the preprocessed path set may include four components.

At this time, the first preprocessed path 31 is an array (Array_P ()) when moved to the first label of the array indicating the obtained path data (Path (1)), and Array_P 0, 0, 0}. The next predicted position corresponding to the preprocessed path 31 may be two.

The second preprocessed path 32 is an array (Array_P ()) when moved to the second label of the array representing the obtained path data (Path (1)), Array_P (1_2) = { 0, 0}. The next predicted position corresponding to the preprocessed path 32 may be nine.

The third preprocessed path 33 is an array (Array_P ()) when moved to the third label of the array representing the obtained path data (Path (1)), Array_P (1_3) = { 9, 0}. The next predicted position corresponding to the preprocessed path 33 may be ten.

The fourth preprocessed path 34 is an array (Array_P ()) when moved to the fourth label of the array indicating the obtained path data (Path (1)), Array_P (1_4) = { 9, 10}. The next predicted position corresponding to the preprocessed path 34 may be zero. At this time, the next prediction position is indicated as 0 because the destination is the label 10.

 When a plurality of path data (Path (1), Path (2), Path (3)) is preprocessed, one pre-processed path set 30 can be generated.

The one pre-processed path set 30 may include a plurality of preprocessed paths.

If a value obtained by adding together the lengths of the arrays (Array ()) representing the plurality of path data (Path (1), Path (2), Path (3)) is NT, the one pre- The number of preprocessed passes in the first and second paths 30, 30 may be NT. For example, since the lengths of Path (1), Path (2), and Path (3) are 4, 6, and 5, NT = 15. Thus, the number of preprocessed passes in one preprocessed path set may be fifteen.

As described above, each preprocessed path of the one pre-processed path set is provided in the form of an array. At this time, the lengths of the NT arrays (Array_P) may be determined to be the same value L_AP. At this time, L_AP can have the following values. If the length of the array having the longest value among the arrays (Path (1), Path (2), Path (3)) representing the plurality of path data is L_M, L_AP can be determined to be the same value as L_M. However, when L_AP is not the square of the natural number, it may be re-determined as the smallest value of the square of the natural number larger than L_AP. For example, when the lengths of the arrays representing the plurality of (e.g., three) path data are 4, 5, and 6, respectively, L_M = 6 and thus L_AP = L_M = 6. However, since 6 is not a square of a natural number, it can be recalculated to the smallest value (= 9) of the values (9, 16, ...) of the natural number squared larger than L_AP (= 6) ).

In step S31, when the obtained 'path data' is preprocessed and a 'preprocessed path set' is generated, the number of preprocessed paths of the 'preprocessed path set 30' It is possible to generate the predicted position set 50 having the same number.

In step S40, a plurality of preprocessed paths included in the 'preprocessed path set' and a set of predicted positions may be provided to a prediction model prepared in advance, thereby learning the prediction model.

That is, the prediction model can be learned by sequentially inputting the preprocessed paths, which are components of the preprocessed path set, one by one to the input layer of the prediction model. At this time, the model can be learned by using (1) a value output from the output layer of the predictive model, (2) and a predicted position corresponding to each input preprocessed path.

In step S50, the predicted movement path of the specific person can be reconfigured in the form of a path by receiving the current position of the specific person. For example, if it is assumed that the specific person exists in the current sub-area (y1x1) 1, the following procedures can be performed.

At this time, the step S50 is a step of predicting the next movement position of the specific person by using the values of the model and the sub-area learned in FIGS. 12 to 13 instead of the GPS type.

Since the specific person exists in the subarea 1, the subarea in which the specific person can move in the whole map (Fig. 12 (a)) is any one of 2, 8, and 9.

In order to use the learned model, data must be converted in the form of data that can be input to the input layer of the learned model.

Accordingly, each of the current movement positions of the specific person must be converted into a path data form having an array shape. At this time, the length of the array has the same size as the L_AP (for example, 9) which was determined when the learning model learned for the specific person was created.

For example, since the particular person is located in the subarea 1, the preprocessed path for the subarea 1 is Array_T {1} = {1, 0, 0, 0, 0, 0, 0, 0, 0}. The value output when the value of Array_T {1} is put into the input layer of the learned learning model can be the next expected moving sub-region.

For example, Array_T {2} = {1, 2, 0, 0, 0, 0, 0, 0, 0} if the particular person has reached the current sub- -> Array_T {2} value is input to the input layer of the learned learning model, the output value can be the next expected moving sub-region.

According to the above, one subsequent sub-region can be predicted from the actual movement path to the present, but if it is repeatedly applied, the expected moving sub-region of the future can be expanded and predicted.

For example, when the current person stays in the sub-region 1, Array_T {1} = {1, 0, 0, 0, 0, 0, 0, 0, 0} Can come out. That is, it is predicted that the current specific person will move from the sub region 1 to the sub region 2. Then, if the predicted value Array_T {2} = {1, 2, 0, 0, 0, 0, 0, 0, 0} created by using the motion path predictive value of the specific person is put back into the learning model, Can come out. That is, it can be predicted that a specific person in sub-area 1 will move to sub-area 9 via sub-area 2. [

&Lt; Third Embodiment >

FIG. 14 is a flowchart illustrating a motion path prediction method according to another embodiment of the present invention.

In the moving route predicting method according to this embodiment, the step (S110) of acquiring the geolocation information set by collecting the geolocation information obtained when the specific object moves is performed; Converting the set of geolocation information into path data in an array form (S120); A step (S130) of preprocessing the converted path data to generate a plurality of preprocessed paths (31 to 45) and a plurality of prediction positions (50) corresponding to the plurality of preprocessed paths; A step (S140) of inputting the plurality of preprocessed paths (31 to 45) and a plurality of prediction positions (50) into a prediction model prepared in advance and learning the prediction model; And transforming the first set of geolocation information acquired from the geolocation information according to another movement path of the specific object into first path data of an array type, inputting the first pass data to the learned prediction model, And determining a subsequent predicted position subsequent to the movement path (S150).

The geolocation information may include GPS information, for example. The one set of the geolocation information may be a set of geolocation information obtained from one independent movement path that passes when the specific object moves once. If the particular object performs independent movement a plurality of times, a plurality of sets of the geolocation information may be obtained.

One path data in the array form can be obtained from the one set of geolocation information. That is, N arrays representing N pass data from the N sets of geolocation information. For example, an example of the N path data is Path (1), Path (2), and Path (3) shown in FIG. 12 (b).

A plurality of preprocessed paths generated from specific path data may also have an array form. However, the length of the array representing the specific path data may be less than or equal to the length of the arrays representing the plurality of preprocessed paths. The arrays representing the plurality of preprocessed passes may all have the same length. The plurality of preprocessed passes may be given, for example, as the preprocessed passes shown at 30 in FIG. 12 (c).

When inputting a specific &quot; preprocessed path &quot; into the input layer of the predictive model at the time of learning the predictive model at the step S140, the output layer of the predictive model is set to &quot; Predicted position &quot;. For example, when inputting the 'preprocessed path' denoted by reference numeral 31 in FIG. 12 (c) to the input layer, the 'predicted position' of the reference numeral 31 may be input to the output layer.

The plurality of preprocessed paths may be input to the input layer of the predictive model and the plurality of predictive positions may be input to the output layer of the predictive model to learn the predictive model.

At this time, when the first pass data is input to the input layer of the learned prediction model, information output from the output layer of the learned prediction model can be determined as the subsequent prediction position.

At this time, as illustrated in (c) of FIG. 12, the plurality of preprocessed paths may all have an array shape having the same first length.

In the converting step, a plurality of path data may be provided. To this end, in step S110, a plurality of sets of geolocation information may be obtained by observing the movement of the specific object a plurality of times. When a plurality of path data are provided, a plurality of arrays representing the path data are also provided.

The moving path predicting method may further include converting the first path data into an array having the first length so as to input the first path data into the learned prediction model. At this time, one or more null data may be padded to the first pass data for the conversion. The null data may be '0', for example.

When a plurality of path data are provided, the plurality of path data each have an array shape, and at least two different path data among the plurality of path data may have different lengths.

Some or all of the plurality of preprocessed paths may be generated by padding null information (null data) to the information included in the path data.

The step of converting the path data into the path data includes preparing information on a global map in which sub-areas each having a label are assigned are defined; And sequentially observing the sub areas in which the specific object is located when the plurality of pieces of the geolocation information are arranged in accordance with the flow of time and sequentially arranging the labels of the sub area according to the observed result in an array And a step of generating the data. At this time, neighboring elements in the array representing the path data represent different sub-regions.

When a plurality of path data are provided, the number of the plurality of preprocessed paths may be equal to a sum of lengths of the respective arrays representing the plurality of path data.

The moving path predicting method comprising: generating second path data by padding the next predicted position to the end of the first path data; And inputting the second path data to the learned prediction model to determine a second subsequent prediction position following the next prediction position. That is, according to an embodiment of the present invention, it is possible to predict a position at a farther future time point as well as a near future time point.

When a plurality of path data are provided, the first length may be equal to or longer than the length of the longest path data among the plurality of path data, and the first length may be forced to have a value that is a square of a natural number .

The particular object may be a particular person or a particular measuring device.

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 (10)

Acquiring a geolocation information set by collecting geolocation information obtained when a specific object moves;
Converting the set of geolocation information into path data in array form;
Pre-processing the transformed path data to generate a plurality of preprocessed paths (31 to 45) and a plurality of predicted positions (50) corresponding to the plurality of preprocessed paths;
Inputting the plurality of preprocessed passes (31 to 45) and the plurality of prediction positions (50) into a prediction model prepared in advance, and learning the prediction model; And
The method comprising: converting a first set of geolocation information obtained from geolocation information according to a new moving path of the specific object to first pass data of an array type, inputting the first pass data to the learned prediction model, Determining a subsequent predicted position subsequent to the path;
/ RTI &gt;
A method of predicting a moving path. The method according to claim 1,
Wherein the plurality of preprocessed paths are input to the input layer of the predictive model and the plurality of predictive positions are input to the output layer of the predictive model to learn the predictive model,
And the information output from the output layer of the learned predictive model is determined as the next predicted position when the first pass data is input to the input layer of the learned predictive model.
A method of predicting a moving path. The method according to claim 1,
Wherein the plurality of pre-processed passes all have an array shape having the same first length,
In the converting step, a plurality of path data are provided,
Wherein the first length is equal to or longer than a length of the longest path data among the plurality of path data.
A method of predicting a moving path. The method according to claim 1,
Wherein the plurality of pre-processed passes all have an array shape having the same first length,
Further comprising converting the first pass data to an array having the first length to input the first pass data to the learned predictive model,
And one or more null data is padded to the first pass data for the conversion.
A method of predicting a moving path. The method according to claim 1,
In the converting step, a plurality of path data are provided,
Each of the plurality of path data having an array form,
Wherein at least two different path data among the plurality of path data have different lengths.
A method of predicting a moving path. The method according to claim 1,
The plurality of preprocessed passes each having an array shape,
Wherein some or all of the plurality of preprocessed paths are generated by padding null information on information included in the path data.
A method of predicting a moving path. The method according to claim 1,
The step of converting into the path data includes:
Preparing information on a global map in which sub-areas each having a label assigned thereto are defined; And
Sequentially observing the sub-areas where the specific object is located when the plurality of pieces of the geolocation information are arranged in accordance with the flow of time, and generating labels of the sub-areas according to the observed result in an array form according to the order of time ;
/ RTI &gt;
Characterized in that the elements adjacent to each other in the array representing the path data represent different sub-
A method of predicting a moving path. 5. The method of claim 4,
In the converting step, a plurality of path data are provided,
Wherein the number of preprocessed paths (31 to 45) is a sum of the lengths of the arrays representing the plurality of path data,
A method of predicting a moving path. The method according to claim 1,
Padding the next predicted position to the end of the first pass data to generate second pass data;
Inputting the second path data to the learned prediction model to determine a second subsequent prediction position following the next prediction position;
&Lt; / RTI &gt;
A method of predicting a moving path. The method of claim 3,
In the converting step, a plurality of path data are provided,
Wherein the plurality of pre-processed passes all have an array shape having the same first length,
Wherein the first length is equal to or longer than a length of the longest path data among the plurality of path data,
Wherein the first length is set to a value which is a square of a natural number and is reset to a smallest value among first natural numbers greater than the first length when the first length is not a square of a natural number,
A method of predicting a moving path.

Images