JP2012103902A - Action prediction method, device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To utilize position measurement data in a non-accumulated state in predicting actions.SOLUTION: An action prediction method includes: retrieving, when current position measurement data does not belong to any of plural accumulation locations extracted by analyzing past position measurement data, past position measurement data having geographical positions in the vicinity of the current position measurement data from among past position measurement data not belonging to any of the plural accumulation locations prerecorded in a recording unit in association with immediately preceding and immediately subsequent accumulation locations; extracting past position measurement data associated with an immediately preceding accumulation location that is the same as that of the current position measurement data from among the retrieved past position measurement data; and predicting the next accumulation location of a target based on an immediately subsequent accumulation location associated with the extracted past position measurement data.

Description

  The present invention relates to target behavior prediction.

  Position information sensors represented by GPS and the like have been reduced in size and price, and are being installed in various devices. For example, a communication service that detects position information of a mobile phone by GPS and provides information (for example, neighboring facility information) according to the current location to a user is operated. In the car navigation system, information corresponding to the current location is also provided.

  In addition to the information providing service corresponding to the current location, a service that predicts the next staying location of the target (such as a user) and provides information corresponding to the next staying location can be assumed. For example, a technique described in Non-Patent Document 1 is known as a technique for predicting a user's next staying place.

  The method described in Non-Patent Document 1 extracts a plurality of staying places by applying clustering to the history of target position measurement data. In this method, a Markov model expressing transitions between a plurality of staying places is created in advance, and the next staying place of the target is predicted based on the Markov model.

Daniel Ashbrook and Thad Starner, “Using GPS to Learn Significant Locations and Predict Movement across Multiple Users”, Personal and Ubiquitous Computing 7 (5), pp. 275-286 (2003) Martin Ester, Hans-Peter Kriegel, Jorg Sander, and Xiaowei Xu “A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise '', in proceedings of 2nd international conference on knowledge discovery and data mining, pp.226-231 , 1996.

  The Markov model defines the transition probability between staying places. Therefore, when the current position of the target does not belong to any staying place, it is difficult to predict with high accuracy. For example, when the target transitions from the staying place A toward the staying place B, the probability that the next staying place is B is expected to increase as the target approaches the staying place B. However, the prediction based on the Markov model does not intend to use position measurement data in such a non-staying state (during transition between staying places) for prediction. Therefore, when the target is moving between the staying places, the prediction accuracy may be substantially lowered. In addition, when the behavior prediction result is dynamically updated, a high update speed may not be achieved (for example, the behavior prediction result is not updated until the target staying place transitions).

  Therefore, an object of the present invention is to utilize position measurement data in a non-staying state in behavior prediction.

  If the current position measurement data does not belong to any of the plurality of staying places extracted by analyzing past position measurement data, the behavior prediction method according to one aspect of the present invention is described as follows: Search for past position measurement data having a geographical position near the current position measurement data from past position measurement data that does not belong to any of the plurality of staying places that are recorded in the recording unit in advance. And extracting the past position measurement data associated with the previous stay location same as the current position measurement data from the searched past position measurement data, and the extracted past position measurement data Predicting the next residence location of interest based on the immediate residence location associated therewith.

  According to the present invention, position measurement data in a non-staying state can be utilized in behavior prediction.

The block diagram which illustrates the action prediction device concerning one embodiment. The flowchart which illustrates the learning process by the action prediction apparatus of FIG. The figure which illustrates position measurement data. Explanatory drawing of a residence place. The figure which illustrates residence action data. The flowchart which illustrates the detail of step S210 of FIG. The figure which illustrates the position measurement data recorded on the position measurement data recording part of FIG. The flowchart which illustrates the prediction process by the action prediction apparatus of FIG. 9 is a flowchart illustrating details of step S310 in FIG. Explanatory drawing of the action prediction using the position measurement data in a non-staying state. The figure which illustrates the transition probability corresponding to FIG.

Embodiments of the present invention will be described below with reference to the drawings.
(One embodiment)
As shown in FIG. 1, the behavior prediction apparatus according to an embodiment includes a position measurement data input unit 100, a position measurement data analysis unit 110, a position measurement data recording unit 120, a staying behavior recording unit 130, and a staying place prediction unit 140. And a prediction result output unit 150.

The behavior prediction apparatus may be a mobile terminal (such as a mobile phone, a mobile PC, or a portable media player) having a position measurement function such as GPS, or a computer (server) connected to the mobile terminal via a network. It may be.
The position measurement data input unit 100 inputs position measurement data obtained by continuously measuring the target position with GPS or the like. The position measurement data includes information indicating measurement time, latitude, and longitude as shown in FIG. 3, for example. The position measurement data is input to the position measurement data analysis unit 110 or the stay location prediction unit 140 according to the processing mode of the behavior prediction apparatus. Specifically, if the behavior prediction apparatus is operating in the learning mode, the position measurement data 10 is input to the position measurement data analysis unit 110. On the other hand, if the behavior prediction apparatus is operating in the prediction mode, the position measurement data 13 is input to the staying place prediction unit 140.

  The learning mode is executed prior to the prediction mode. In the learning mode, residence location data, residence behavior data, location measurement data in a non-staying state, and the like are recorded based on the analysis of the location measurement data 10 in the learning period. Note that the length of the period for learning can be arbitrarily determined. Details of these various data will be described later. In the prediction mode, the next staying place of the target is predicted based on the position measurement data 13 and various data recorded in the learning mode.

  As described later, the position measurement data analysis unit 110 analyzes the position measurement data 10 to obtain stay location data, stay action data, position measurement data in a non-stay state, and the like. The position measurement data analysis unit 110 records the stay behavior data 11 in the stay behavior recording unit 130 and records the position measurement data 12 in the non-staying state in the position measurement data recording unit 120.

  The position measurement data recording unit 120 and the staying action recording unit 130 may be formed of any information recording medium (for example, HDD, SSD, flash memory, etc.). The position measurement data recording unit 120 records the position measurement data 12 from the position measurement data analysis unit 110. The stay location prediction unit 140 described later reads the position measurement data 15 recorded in the position measurement data recording unit 120 as necessary. The staying behavior recording unit 130 records the staying behavior data 11 from the position measurement data analysis unit 110. The stay location prediction unit 140 reads the stay behavior data 14 recorded in the stay behavior recording unit 130 as necessary.

  The stay location prediction unit 140 receives the position measurement data 13 from the position measurement data input unit 100 and reads the position measurement data 15 and the stay behavior data 14 from the position measurement data recording unit 120 and the stay behavior recording unit 130, respectively. Predict the next place of residence. Details of the prediction process will be described later. The stay location prediction unit 140 outputs the prediction result 16 to the prediction result output unit 150. The prediction result 16 may be information indicating a transition probability to each staying place, information indicating a staying place having the highest transition probability, or other information. The prediction result output unit 150 outputs the prediction result 16 from the stay location prediction unit 140 to the outside.

Hereinafter, the learning process by the behavior prediction apparatus in FIG. 1 will be described with reference to FIG.
When the process of FIG. 2 starts, the position measurement data analysis unit 110 inputs the position measurement data 10 from the position measurement data input unit 100 (step S201). The position measurement data 10 can be represented by a set Z _train = {z ₁ , z ₂ ,..., Z _M } composed of M (M is an integer of 2 or more) elements. Here, each element z _i is a two-dimensional vector (x _i , y _i ) composed of longitude x _i and latitude y _i measured at time t (z _i ). In the example of FIG. 3, i corresponds to the column of measurement IDs, y _i corresponds to the column of latitudes, x _i corresponds to the column of longitudes, and t (z _i ) corresponds to the column of times. . When i ≦ j, it is assumed that t (z _i ) ≦ t (z _j ) is satisfied, and t (z _i ) ≦ t (z _j ) ≦ t (z for any 1 ≦ i ≦ M−1 Assume that there is no j that holds _{i + 1} ).

  Next, the position measurement data analysis unit 110 applies a process such as clustering to the position measurement data 10 input in step S201, thereby extracting (a plurality of) stay locations where the target has stayed for a certain period of time or more ( Step S202). The stay location can be extracted using any conventional method. In the following description, as an example, each staying place is extracted as a cluster using the clustering algorithm described in Non-Patent Document 2 (hereinafter referred to as DBSCAN).

FIG. 4 shows an application example of DBSCAN. In FIG. 4, white circles and shaded circles represent the geographical positions (longitude and latitude) of individual position measurement data, and arrows represent the time-series order between the position measurement data. Area surrounded by a broken line represents the geographical range of residence location c ₁ and c _2, which are extracted as a cluster (longitude and latitude). In the following description, a set of staying places is represented by C = {c ₁ , c ₂ ,..., C _N } (N is an integer of 2 or more). (In other words, the position measurement data in the staying state) shaded circle position measurement data belonging to the residence where c ₁ or c ₂ representing the. On the other hand, the white circle represents position measurement data that does not belong to any of the staying places c ₁ and c ₂ (in other words, position measurement data in a non-staying state).

Next, the position measurement data analysis unit 110 extracts the stay behavior data by analyzing the position measurement data 10 based on the stay place extracted in step S202 (step S203). The extracted staying behavior data 11 is recorded in the staying behavior recording unit 130. The staying action data includes an identifier for identifying the staying action (column of staying action ID in FIG. 5), information indicating the staying place corresponding to the staying action (column of staying place in FIG. 5), and the staying place data. It includes a stay start time (start time column in FIG. 5) and a stay end time (end time column in FIG. 5). For example, if the geographic location of an arbitrary position measurement data z _i is included in the geographic area of residence location c _j, is defined as the target had been retained in the residence where c _j at time t (z _i) The Further, the object is staying at the staying place c _j at time t (z _i ), t (z _{i + 1} ),..., T (z _{i + α} ), and time t (z _i−1 ) and t ( If it is not staying at the staying location c _j at z _{i + α + 1} ), the staying start time is time t (z _i ), and the staying end time is time t (z _{i + α} ).

  Next, the position measurement data analysis unit 110 analyzes position measurement data (that is, non-staying behavior) that does not belong to any of the staying places extracted in step S202 (step S210). Details of the process in step S210 will be described later. After step S210, the process in FIG. 2 ends.

Hereinafter, the details of step S210 in FIG. 2 will be described with reference to FIG.
When the process of FIG. 6 starts, the position measurement data analysis unit 110 extracts position measurement data that does not correspond to any of the staying behaviors extracted in step S203 (step S211). For example, the position measurement data analysis unit 110 extracts all position measurement data z _i having a geographical position that is not included in the geographical range of any stay location c _j .

Next, the position measurement data analysis unit 110 searches for staying action IDs before and after the position measurement data z _i extracted in step S211 (step S212). Specifically, the position measurement data analysis unit 110 compares the remaining behavior before the time t (z _i ) with the closest temporal behavior (the previous residence behavior) and the time t (z _i ) later than the time t (z _i ). Each of the staying behaviors that are closest in time (the staying behavior immediately after) is searched for, and these staying behavior IDs are acquired. For example, for position measurement data P in FIG. 4, the residence action immediately before are those corresponding to the residence where c _1, retention behavior immediately after are those corresponding to the residence where c _2. If the position measurement data analysis unit 110 cannot search for at least one of the immediately preceding staying action and the immediately following staying action, it can discard the position measurement data.

  Next, the position measurement data analysis unit 110 records the position measurement data in the position measurement data recording unit 120 in association with the staying action ID immediately before and after the search in step S212 (step S213), and the process of FIG. 6 ends. . For example, the position measurement data 12 is recorded in the position measurement data recording unit 120 in a state where the immediately preceding and immediately following staying action IDs are added as shown in FIG.

Hereinafter, prediction processing by the behavior prediction apparatus in FIG. 1 will be described with reference to FIG.
When the process of FIG. 8 starts, the staying place prediction unit 140 inputs the (current) position measurement data 13 from the position measurement data input unit 100 (step S301). Next, the staying place prediction unit 140 determines whether the position measurement data 13 input in step S301 belongs to any staying place (step S302). That is, the staying place prediction unit 140 determines whether or not the geographical position of the position measurement data 13 is included in any geographical range of staying places extracted in the preceding learning process. If position measurement data 13 belongs to any staying location, the process proceeds to step S303; otherwise, the process proceeds to step S310.

In step S303, the stay location prediction unit 140 performs a prediction process based on a conventional Markov model (for example, described in Non-Patent Document 1), and the process proceeds to step S304. That is, the staying place prediction unit 140 predicts the next staying place based on the transition probability between staying places defined in the Markov model. The transition probability p (c _j ) to the staying place c _j satisfies the following conditional expression.

ステップＳ３１０において、滞留場所予測部１４０は後述される部分処理を実行することにより各滞留場所への遷移確率を予測し、処理はステップＳ３０４に進む。ステップＳ３０４において、滞留場所予測部１４０はステップＳ３０３またはステップＳ３１０において予測した遷移確率などの予測結果１６を出力し、図８の処理が終了する。

In step S310, the staying place prediction unit 140 predicts a transition probability to each staying place by executing a partial process described later, and the process proceeds to step S304. In step S304, the stay location prediction unit 140 outputs the prediction result 16 such as the transition probability predicted in step S303 or step S310, and the processing in FIG. 8 ends.

Hereinafter, the details of step S310 in FIG. 8 will be described with reference to FIG.
When the process of FIG. 9 starts, the staying place prediction unit 140 selects all of the past position measurement data 15 recorded in the position measurement data recording unit 120 that have a geographical position near the position measurement data 13. Search is performed (step S311). Here, the definition of the neighborhood is not particularly limited. For example, the distance from the geographical position of the position measurement data 13 can be set to a range equal to or less than a threshold value.

Next, the stay location prediction unit 140 matches the previous stay location corresponding to the previous stay action ID from the past position measurement data searched in step S311 with the stay location just before the position measurement data 13. All the items to be extracted are extracted (step S312).
Next, the staying place prediction unit 140 calculates a transition probability based on the staying place (immediately after) corresponding to the staying action ID immediately after being associated with the past position measurement data extracted in step S312, and the process of FIG. Ends. For example, the staying place prediction unit 140 can calculate the transition probability p (c _j ) by the following mathematical formula.

ここで、ｎ_ｊは、ステップＳ３１２において抽出された過去の位置計測データのうち直後の滞留行動ＩＤに対応する滞留場所がｃ_ｊであるものの総数である。但し、直後の滞留行動ＩＤの重複は許容されない。即ち、複数の過去の位置計測データの直後の滞留行動ＩＤが重複するならば、これらは同一視されて１つの位置計測データとして扱われる。このような取り扱いにより、特定の滞留行動に関して得られた位置計測データが遷移確率に過剰に影響を与える事態が回避されるので、より妥当な遷移確率が計算されやすくなる。

Here, n _j is the total number of the past location measurement data extracted in step S312 with the stay location corresponding to the stay action ID immediately after being c _j . However, duplication of the staying action ID immediately after is not allowed. That is, if the staying action ID immediately after a plurality of past position measurement data overlaps, they are identified and treated as one position measurement data. Such handling avoids a situation in which the position measurement data obtained with respect to a specific staying behavior excessively affects the transition probability, so that a more appropriate transition probability can be easily calculated.

A specific example of the partial processing in FIG. 9 will be described below with reference to FIGS. In the example of FIG. 10, the residence places c ₁ , c ₂ and c ₃ are defined. Among the position measurement data obtained during the learning process, to those measured before the transition from the retention location c ₁ to the residence where c ₂ is drawn in white circles, the transition from the residence where c ₁ to residence location c ₃ The measured value is drawn in a shaded circle. The position measurement data t1 and t2 obtained at the time of the prediction process are drawn with crosses. Further, the range in the vicinity of the position measurement data t1 and t2 obtained at the time of the prediction process is drawn with a circle centered on the x mark.

Two shaded circles and one white circle are included in the vicinity of the position measurement data t1. However, these two shaded circles are regarded as one position measurement data because they have the same staying action ID immediately after them and are identified. That is, n ₂ = n ₃ = 1. Therefore, according to the above equation, p (c ₂ ) = p (c ₃ ) = 0.5 as shown in FIG. On the other hand, one shaded circle is included in the vicinity of the position measurement data t2. That is, n ₂ = 1 and n ₃ = 0. Therefore, according to the above formula, p (c ₂ ) = 1 and p (c ₃ ) = 0 as shown in FIG. As described above, according to the partial processing in FIG. 9, the predicted transition probability changes every moment during the transition between the staying places.

  As described above, the behavior prediction apparatus according to the embodiment learns in advance the data measured during the transition between the staying locations in association with the immediately preceding and immediately following staying locations, and the data measured at the time of the prediction If it does not belong to the staying place, the next staying place is predicted using the previously learned data. Therefore, according to the behavior prediction apparatus according to the present embodiment, it is easy to increase the prediction accuracy during the transition between the staying places, and it is easy to achieve a high update speed when dynamically updating the behavior prediction result. .

  In addition, the said embodiment is not limited as the description, In the implementation stage, a component can be deform | transformed and embodied in the range which does not deviate from the summary. Moreover, various aspects can be formed by appropriately combining a plurality of components disclosed in the embodiment. Further, for example, a configuration in which some components are deleted from all the components shown in the above-described embodiment is also conceivable.

  The processing of the above embodiment can be realized by using a general-purpose computer as basic hardware. The program for realizing the processing of the above embodiment may be provided by being stored in a computer-readable storage medium. The program is stored in the storage medium as an installable file or an executable file. The storage medium may be a computer-readable storage medium such as a magnetic disk, optical disk (CD-ROM, CD-R, DVD, etc.), magneto-optical disk (MO, etc.), semiconductor memory, etc. Any form may be used. Further, the program for realizing the processing of the above embodiment may be stored on a computer (server) connected to a network such as the Internet and downloaded to the computer (client) via the network.

10, 12, 13, 15 ... Position measurement data 11, 14 ... Staying action data 16 ... Prediction result 100 ... Position measurement data input unit 110 ... Position measurement data analysis unit 120 ... Position measurement data recording unit 130 ... Residence action recording unit 140 ... Residence place prediction unit 150 ... Prediction result output unit

Claims (9)

If the current position measurement data does not belong to any of the plurality of staying places extracted by analyzing past position measurement data, the plurality of pieces recorded in the recording unit in advance in association with the immediately preceding and immediately following staying places Searching for past position measurement data having a geographical position in the vicinity of the current position measurement data from past position measurement data that does not belong to any of the stay locations;
Extracting past position measurement data associated with the immediately preceding staying place same as the current position measurement data from the searched past position measurement data;
Predicting the next staying location of the target based on the staying location immediately after being associated with the extracted past position measurement data.   The behavior prediction method according to claim 1, wherein the prediction result of the next staying location of the object is expressed using a transition probability based on the staying location immediately after being associated with the extracted past position measurement data. Analyzing the past position prediction data to extract the plurality of staying places;
2. The behavior prediction method according to claim 1, further comprising: past position measurement data that does not belong to any of the plurality of staying places is recorded in the recording unit in association with immediately preceding and immediately following staying places.   2. The behavior prediction method according to claim 1, wherein a distance from a geographical position of the current position measurement data is in a range of a threshold value or less near the current position measurement data. A recording unit in which past position measurement data that does not belong to any of the plurality of staying places extracted by analyzing past position measurement data is recorded in advance in association with the staying place immediately before and immediately after,
If the current position measurement data does not belong to any of the plurality of staying places, it has a geographical position near the current position measurement data from the past position measurement data recorded in the recording unit. The past position measurement data was searched, and the past position measurement data associated with the immediately previous staying location same as the current position measurement data was extracted from the searched past position measurement data and extracted. A behavior prediction apparatus comprising: a prediction unit that predicts a next stay location of a target based on a stay location immediately after being associated with past position measurement data.   The behavior prediction apparatus according to claim 5, wherein the prediction result of the next staying location of the object is expressed using a transition probability based on the staying location immediately after being associated with the extracted past position measurement data.   Analyzing the past position prediction data to extract the plurality of staying places, and past position measurement data not belonging to any of the plurality of staying places is associated with the immediately preceding and immediately following staying places in the recording unit The behavior prediction apparatus according to claim 5, further comprising an analysis unit for recording.   The behavior prediction apparatus according to claim 5, wherein the distance from the geographical position of the current position measurement data is in a range that is equal to or less than a threshold value in the vicinity of the current position measurement data. If the current position measurement data does not belong to any of the plurality of staying places extracted by analyzing past position measurement data, it is recorded in advance in the recording unit in association with the staying place immediately before and after Means for searching past position measurement data having a geographical position near the current position measurement data from past position measurement data not belonging to any of the plurality of staying places;
Means for extracting past position measurement data associated with the immediately preceding staying place same as the current position measurement data from the searched past position measurement data;
An action prediction program for functioning as a means for predicting a next staying location of an object based on a staying location immediately after being associated with extracted past position measurement data.

Images