JP5838879B2 - Prediction system - Google Patents

Description

  The present invention relates to a prediction system that predicts the possibility of a mobile object appearing.

  Conventionally, the object of technology development related to risk prediction is to predict an event within a few seconds from the driver's perspective, and most of the assumptions are that the movement of the vehicle substantially follows the equation of motion (for example, Patent Document 1). See).

JP 2009-276992 A

  However, the technique described in Patent Document 1 is based on the assumption that the sensing performance is highly accurate in order to predict the movement of the target vehicle or pedestrian under the condition of within a few seconds when viewed from the driver. Yes. However, with the current outside-vehicle environment recognition technology, the recognition rate deteriorates as the vehicle travels at a higher speed, and the recognition becomes more difficult for objects that are far away or hidden. Therefore, with a vehicle alone, pedestrians appearing from the blind spot even within a few seconds, pedestrians in hidden or unclear images, pedestrians several hundred meters or more away, and further traffic through a right or left turn at an intersection There was a problem that the event could not be handled at all.

  The present invention has been made in view of these problems, and an object of the present invention is to provide a technique that can deal with the appearance of a moving object with a sufficient margin.

In the prediction system of the first invention and the second invention made to achieve the above object, the acquisition means includes a movement path position indicating the position of the movement path in which the moving body has moved in the past, and the moving body at the movement path position. Learning in which spatiotemporal trajectory data indicating the spatiotemporal trajectory associated with the existing time and mobile object attribute data indicating the attributes of the mobile object are learned using a mobile object database stored for each of the plurality of mobile objects When a space-time indicating the time and position is specified when the prediction means acquires the result, the specified time-space is designated as the specified time-space, and the specified time-space is determined based on the learning result acquired by the acquisition means. Predict the possibility that a moving object will appear at
Further, in the prediction systems of the first and second inventions, the moving object is a human, and the learning result is a result of learning by regression analysis assuming that the certainty about human existence and attributes returns to time and place. The predicting means calculates at least a part of the planned travel route that is a route on which the vehicle is scheduled to travel, calculates a certainty factor as a possibility that a person appears at the specified spatiotemporal point, Calculates the certainty factor at a plurality of spatio-temporal points within a preset spatio-temporal range including the portion designated as the designated spatio-temporal point in the scheduled travel route, The sum of the calculated certainty levels is calculated as the risk level.
In the prediction system according to the first aspect of the present invention, when the number of moving objects to be learned is small in the spatiotemporal range, the acquisition means has a higher time density in the time zone in the spatiotemporal range than in other time zones. By performing sampling, learning is performed by reducing the state matrix of the regression analysis, and this result is acquired as a learning result.
In the prediction system according to the second aspect of the present invention, when the number of pedestrians to be learned is small in the spatio-temporal range, the acquisition means is in other areas where the pedestrian movement pattern is similar to the spatio-temporal range. Data is extracted from the mobile database, learning is performed using the extracted data, and the result is acquired as a learning result.

  The prediction system configured as described above learns the spatio-temporal trajectory data of the moving body in the past, and predicts the possibility of the moving body appearing at the specified time and position. Therefore, if the time and position are specified, the user of the prediction system can know the possibility that a moving object will appear at the specified time and position when the time and position are specified. Can deal with the appearance of the body with sufficient margin.

It is a block diagram which shows the structure of the danger prediction system. It is a flowchart which shows the first half part of a danger prediction process. It is a flowchart which shows the latter half part of a danger prediction process. It is a figure which shows the whole outline | summary of the process of appearance prediction. It is a figure which shows the sampling using a rectangular density function. It is a figure which shows the sampling using the density function of arbitrary shapes. It is a figure which shows the usage method of a similar area. It is a figure which shows the prediction method of a spatiotemporal orbit.

Embodiments of the present invention will be described below with reference to the drawings.
[1. Overview]
The danger prediction system 1 to which the present invention is applied learns past human flow data, and predicts the number and attributes of pedestrians encountered at a date and time place on a planned driving route. This is a system for predicting danger.

[1-1. Feature]
(1) Learning human flow data on the assumption that the certainty about the existence and attributes of a person returns to time and place, and predicting when and where the person will appear on the travel route.

  (2) The degree of danger is calculated by taking the sum of the certainty levels of people in a certain space-time range (called the danger range) on the travel route. This danger range is changed according to the class composed of (age, sex, means of transportation).

(3) When the number of persons to be learned in a certain area is small, the state matrix is reduced in order by changing the sampling time around the scheduled travel time, and the prediction calculation is performed smoothly.
(4) When the number of people to be studied in a certain area is small, the time variation pattern of the number of people in the area is similar for each pedestrian attribute (purpose, occupation, age, sex) even in different areas If it is, use it as learning data.

[1-2. Specific use]
The present invention is applied to a driving support device that predicts the possibility of encountering a pedestrian on a planned travel route and issues a warning to the driver.

[2. System configuration]
The risk prediction system 1 to which the present invention is applied includes a database unit 10, a learning unit 20, a condition setting unit 30, and a prediction unit 40, as shown in FIG.

The database unit 10 includes a car navigation device 11, a human flow database 12, a search interface 13, and a spatiotemporal trajectory data extraction / processing unit 14.
The car navigation device 11 inputs a position detector (not shown) that detects the current position of the vehicle, a data storage device (not shown) that can store map data and various information, and various instructions from the user. Operating switch (not shown) and a display device (not shown) for performing various displays such as a map display screen and a TV screen. The car navigation device 11 has a map display function for calculating the current position of the vehicle based on the detection result of the position detector and displaying a map and the like near the current position read from the data storage device on the display device. In addition, the car navigation device 11 includes a route guidance function that performs route guidance by specifying a destination according to the operation of the operation switch, and performing route calculation that automatically obtains the optimum planned travel route from the current position to the destination.

  The human flow database 12 is a database that stores human flow data. Human flow data is time-series data that is obtained by public survey and interpolated in space and time at intervals of 1 minute from “identification number, date / time, location, gender, age, address, occupation, purpose, and means of transportation”. . This was expanded to major cities nationwide, starting with Tokyo in 1998, and in recent years has also been expanded overseas. Currently, the Center for Spatial Information Science at the University of Tokyo provides a database of human flow data. For example, about 720,000 people in the Tokyo area in 1998 and about 260,000 people in the Chukyo area in 2001 are publicly available. The personal trip data (hereinafter referred to as PT data) that has been surveyed can be searched by the “Web API” provided by the University of Tokyo.

The search interface 13 is an interface for extracting from the human flow database 12 the human flow data that matches the specified condition.
The spatio-temporal trajectory data extraction / processing unit 14 extracts human flow data from the human flow database 12 and processes the extracted human flow data, thereby obtaining a state matrix (described later) and a pedestrian attribute certainty matrix ( Create later.

The learning unit 20 includes a data learning unit 21, a prediction characteristic database 22, and a prediction characteristic control unit 23.
The data learning unit 21 creates a prediction function indicating prediction characteristics by learning using the state matrix and the certainty factor matrix created by the database unit 10.

The prediction characteristic database 22 stores an existing prediction function.
The prediction characteristic control unit 23 selects either the prediction function created by the data learning unit 21 or the prediction function stored in the prediction characteristic database 22 as a prediction function used for calculating the pedestrian certainty factor.

The condition setting unit 30 includes a prediction condition setting unit 31 and a scheduled traveling track mapping unit 32.
The prediction condition setting unit 31 sets the scheduled travel route through the car navigation device 11 by the passing position and time.

The planned travel path mapping unit 32 sets a representative point Q _I (I = 1, 2,..., _IQ ) on the planned travel route.
The prediction unit 40 includes a pedestrian route estimation unit 41, a pedestrian certainty calculation unit 42, a spatiotemporal distance calculation unit 43, a risk calculation unit 44, and a notification unit 45.

The pedestrian route estimation unit 41 sets a pedestrian route that can pass through each representative point Q _I on the planned travel route.
The pedestrian certainty calculating unit 42 calculates the pedestrian certainty for all the routes set by the pedestrian route estimating unit 41 for each representative point Q _I set by the planned traveling trajectory mapping unit 32.

The spatiotemporal distance calculation unit 43 calculates a spatiotemporal distance between the pedestrian and the planned travel route.
The risk degree calculating unit 44 is based on the pedestrian certainty degree calculated by the pedestrian certainty degree calculating unit 42, the spatiotemporal distance calculated by the spatiotemporal distance calculating unit 43, and various information from the car navigation device 11. To calculate the risk level.

The notification unit 45 notifies the driver of the risk level calculated by the risk level calculation unit 44.
[3. Risk prediction process]
Next, an outline of the procedure of the danger prediction process executed by the danger prediction system 1 will be described.

  When this danger prediction process is executed, as shown in FIGS. 2 and 3, the danger prediction system 1 first sets a scheduled travel route through the car navigation device 11 by the passing position and time in S10. Thereafter, in S20, it is determined whether or not an existing prediction function is used. If an existing prediction function is used (S20: YES), the existing prediction function is loaded from the prediction characteristic database 22 in S30, and the process proceeds to S100. On the other hand, when the existing prediction function is not used (S20: NO), the human flow data in the spatiotemporal range corresponding to the planned travel route is extracted from the human flow database 12 in S40.

  In S50, it is determined whether or not the data extracted in S40 is sufficient as the learning data amount. If the amount of learning data is sufficient (S50: YES), the process proceeds to S90.

  On the other hand, if the amount of learning data is not sufficient (S50: NO), it is determined in S60 whether to use data in a similar area. Here, when the data of the similar area is used (S60: YES), the flow data of the person in the similar area is acquired from the flow database 12 in S70, and the process proceeds to S90. On the other hand, when the data of the similar area is not used (S60: NO), the time axis of the state matrix is reduced in dimension in S80, and the process proceeds to S90.

And if it transfers to S90, a human flow data will be learned, a prediction function will be created, and it will transfer to S100. In S100, a representative point Q _I (I = 1, 2,..., I _Q ) is set on the planned travel route. Thereafter, in S110, I = 0, and in S120, I = I + 1.

  Thereafter, in S130, it is determined whether or not there is pedestrian information (age, sex, occupation, purpose, etc.) in the learning data. Here, when there is no pedestrian information in learning data (S130: NO), it transfers to S150. On the other hand, if there is pedestrian information in the learning data (S130: YES), the pedestrian information is introduced and used in S140, and the process proceeds to S150.

In S150, it is determined whether or not there is pedestrian road information. Here, when there is road information pedestrian (S150: YES), in S160, select N _I present roads that connect to the representative point Q _I, walking range of M _I component along each road Then, R ₁ spatio-temporal points are set according to the class composed of (sex, age, means of transportation), and the process proceeds to S180. On the other hand, if there is no road information of the pedestrian (S150: NO), a motion vector on the route is calculated in S170, and an eight-direction route is assumed and the process proceeds to S180.

In S180, the class number j _c = 0 is set, and in S190, j _c = j _c +1 is set.
At subsequent S200, 4 categories based on R _I number of space-time point corresponding to the class number j _c set in S160 or S170, the pedestrian attributes as follows (A), (B), (C) , (D) is calculated.

(A) Calculate the sum of certainty levels by gender and age.
(B) Sum of certainty factors for each purpose.
(C) Sum of certainty by occupation.

(D) Sum of certainty factors for each means of transportation (walking, bicycle).
In S210, it is determined whether or not the pedestrian certainty has been calculated for all classes. Here, when the pedestrian certainty has not been calculated for all classes (S210: NO), the process proceeds to S190 and the above process is repeated. On the other hand, when calculating the pedestrian confidence for all classes (S210: YES), in S220, the integrated value of all the classes in the _{Q I (A), (B} ), (C), (D ) Category. Moreover, they are summed to obtain a sum for all categories, and this is defined as the total pedestrian certainty C _I in Q _I.

At subsequent S230, I is equal to or less than I _Q. Here, if I is less than I _Q (S230: YES), the process proceeds to S120, and the above processing is repeated. On the other hand, if I is equal to or higher than I _Q (S230: NO), C _I on the travel route is regarded as a simple risk (the risk that can be calculated most easily) in S240, and the degree of change and the maximum extracting the upper I _R-number of the hazard from the value.

In addition at S250, and the C _I value for the above I _R-number of the dangerous place, another pedestrian confidence category calculated in S220, road, information about the transportation facilities, information of the surrounding and geographical environment, vehicle speed, vehicle information Then, the composite risk degree is calculated by comprehensively determining the spatiotemporal distance from the past pedestrian trajectory.

Then, in S260, the expected appearance time, place, profile of the appearing pedestrian and the degree of danger are notified, and the danger prediction process is terminated.
[4. Appearance prediction]
As shown in FIG. 4, the danger prediction system 1 interprets human flow data as a spatiotemporal trajectory, and predicts the appearance of pedestrians in three stages: database construction, learning, and prediction. Details will be described below.

[4-1. Construction and learning of PT database]
[4-1-1. Select target area]
Select the area to be predicted. First, the position coordinates of the reference point are represented by longitude and latitude, and for example, a 1 km square centered on it is defined as a target area.

[4-1-2. Sample time setting]
First, a person ID in the target area (hereinafter referred to as PID) is searched from the PT database. The PT database to be searched is 01CHU surveyed in 2001 for 260,000 people in the Chukyo area. Now, October 1, 2001 is set as the date and time, and for example, a PID list in the target area every hour is created. Thereby, the time change of the number of pedestrians in the area in one day is known. For example, 8:00 am is selected as the sample time when commuting to school.

[4-1-3. Search PT data]
The PID list contains the PIDs of K people who were present in the area at 8:00:00 (among those who participated in the 01CHU survey). To 23:59 as a time series of PT data at intervals of 1 minute, and K space-time trajectory files are obtained. Here, not only the travel route but also the attributes of each person (age, gender, occupation, purpose) are described.

[4-1-4. Creation of state measurement matrix]
From the above K space-time orbit files, a matrix of K rows can be created with a time series of longitude and latitude as one row. At this time, the row direction represents time, and the longitude information in the T column and the latitude information in the T column are arranged in one row. Therefore, the size of this state matrix X is (K × 2T). This time, T = 480 was obtained by taking a representative point on the time axis every 2 minutes to reduce the calculation load and taking the average of the position data and the position data after 1 minute.

[4-1-5. Create confidence matrix]
The pedestrian attributes included in the PT data are 70 dimensions in total by age, gender, occupation, and purpose (in the case of 01 CHU). Assuming that this pedestrian attribute constitutes a pedestrian's certainty, consider a model in which the certainty is estimated using time and position as explanatory variables. That is, the certainty matrix Y (K × L) is generally expressed as Y = f (X), and in particular, if f is expressed by linear regression, it can be expressed as Y = XB. However, B is a matrix composed of linear regression coefficients, and the size is (2T × L). However, L = 70, and the value of confidence during learning is either 0 or 1.

[4-1-6. Estimation of regression coefficient matrix]
When a regression coefficient matrix B is obtained by least square estimation by a well-known method in multivariate linear regression analysis, B = (X ^T X) ⁻¹ X ^T Y. This B becomes the prediction characteristic of the pedestrian spatio-temporal distribution obtained by learning PT data.

[4-1-7. State matrix reduction]
When the row size K of the state matrix X is K <2T, that is, when the number K of spatiotemporal orbital data cannot be obtained sufficiently, the covariance matrix S = X ^T X (size is 2T in the least square estimation) For (× 2T), the inverse matrix S ⁻¹ cannot be calculated properly due to the rank problem. The following methods can be considered for coping with this case.

(Method 1): The row size (2T) of X is sufficiently lowered. That is, the number of times to be learned is reduced.
(Method 2): “Ogiki,“ System for predicting the spatio-temporal distribution of humans-Examination of feasibility-“, IEICE Technical Report ITS2010-76 (2011-03).” Principal component analysis is performed on the state, and least square estimation is performed in the eigenspace composed of the top K principal components.

(Method 3): Regularization is performed by adding a small diagonal component to the covariance matrix S = X ^T X.
Since the above (Method 2) and (Method 3) are known methods, they will be omitted, and the proposal regarding (Method 1) will be described below.

(A) “Let the time zone to be learned be only the time zone of the planned driving route”
In the above, it was assumed that T = 480, and position data for 16 hours (difference value from the center position of longitude and latitude) was learned at 2-minute intervals (see FIG. 5). However, collecting data for 960 people corresponding to the minimum number of data K _min = 2T for regularizing S in a 1 km square area at a certain time is very dense (for example, Shinjuku or Shibuya) It is difficult unless it is).

  Therefore, if learning is performed only in the time range corresponding to the required traveling time of the own vehicle, the number of learning data is minimized. For example, in an area passing in 2 hours, it is only necessary to collect space-time orbit data for 120 people at intervals of 2 minutes. At this time, since T = 60, the size of S is 120 × 120, which can be regular (see FIG. 6).

  On the other hand, if the time interval MD to be collected is larger than MD = 2, the time resolution in the prediction is reduced accordingly, so that it is not preferable from the viewpoint of danger prediction. In addition, if the passing time is small, it is not so that the number of data can be reduced indefinitely within a range that can be regularized. Originally, in order to make the most effective use of the PT database, it should be predicted after learning the movements of people outside the driving time zone.

  That is, there is a possibility that a more accurate prediction can be obtained if it is considered that the position information of each person for 24 hours including that is involved in the flow of the person at a certain time. Therefore, the following method (B) is reached.

(B) “Selecting the time to be learned as a variable density centered on the time zone of the planned travel route”
The trajectory data in a wider time range than (A) is learned by sampling the data in the traveling time zone at a high time density (ideally at an interval of 1 minute) and sampling at a low time density in other time zones. (See FIG. 5). However, the number of sample times in the travel time zone and the distribution of the sample time numbers other than the travel time zone (sampling density function) can be arbitrarily set based on the knowledge of the designer (see FIG. 6).

The restriction condition for regularization is to learn the trajectory data of the number of people equal to or greater than K _min = 2T with respect to the total number of samples T. For example, increase the number of sample times at the time when a large number of people are moving as much as possible, or the time that becomes a change point before and after the movement (station waiting time, etc.), and at a certain point (school or office) for a long time A measure such as lowering the sample density is conceivable at the location where it stays.

(C) “Integrating learning results obtained from multiple sample patterns”
In this method, a plurality of sample time patterns set in (B) are defined, and the learning result of each pattern is synthesized by the linearity of multivariate linear regression analysis. As a result, groups having different movement patterns can be learned with different sample patterns, and individual learning results can be superimposed, so that it is considered that adaptation can be made more flexible than in (B).

As described above, which of (A), (B), and (C) is selected is determined empirically according to each prediction case.
[4-1-8. Use of similar areas]
When there is little PT data in the target area A, there is a method of reducing the dimension of the state matrix as described above, but it is also conceivable to substitute the data of the similar area A ′. Whether the area is a similar area is determined based on the closeness of the time-series correlation value and the average value whether the temporal change pattern of the total number of persons for each pedestrian attribute in the area is similar to the target area. A and A 'are considered to be different in the road structure and the location of customer-collecting facilities, but the main road links (from the station to the school, to the bank, to the hospital, etc.) attract customers from the center of the area. A and A ′ are associated only when the difference in distance to the facility is a certain threshold value or less, and the progress r (0 to 1) of the flow of people existing at time t ′ on the link of A ′ is The pseudo-learning data is generated by generating pseudo data indicating that a pedestrian exists at time t ′ at the position coordinate x (r) corresponding to the progress r on the road link (see FIG. 7).

[4-2. Prediction of pedestrian confidence]
[4-2-1. Setting the driving route]
The travel route of the host vehicle can be freely set within a range that does not deviate much from the target area.

[4-2-2. Pedestrian route modeling]
A route of a pedestrian that can pass through each representative point on the traveling route of the vehicle is set, and a total sum of pedestrian certainty is taken for all the routes. When the pedestrian's movement route is set using such a model, the certainty factor within the danger range can be roughly calculated even when there is no road structure or geographical condition.

[4-2-3. Calculation of pedestrian confidence]
Based on the linearity of the regression coefficient matrix B, the degree of certainty that the pedestrian defined by each attribute for the 70-dimensional attribute passes through the assumed route is based on the linearity of the regression coefficient matrix B. It is obtained by applying a B partial matrix Bs to the set state vector Xs. If the sum of all attributes is taken, the total certainty (however, relative value) of a pedestrian encountered at a certain representative point can be calculated as a probabilistic appearance frequency. This is the simplest risk prediction value (simple risk).

[4-2-4. Adaptation of danger range according to pedestrian attributes]
The finite number of pedestrian route models centered on the representative points can be generated dynamically and adaptively by using pedestrian attributes. For example, by setting a difference in average walking speed according to age, such as walking speed 5.0 km / h for ages 20 to 25 and 2.5 km / h for ages 70 and over, the risk range for the same time range. Will change about twice.

[4-3. Risk calculation]
[4-3-1. Simple risk]
The risk level that the risk level is higher as the pedestrian appearance frequency is higher is referred to as a simple risk level. However, the simple risk cannot correctly represent the risk in the following cases.

・ When there is an unexpected jump out in an area where there are very few pedestrians ・ Distance is close to the roadway, but inside a guardrail, pedestrian bridge, underground passage, or building such as a building, or beyond the wall or fence as seen from the roadway When there are pedestrians on different floors, etc. ・ When the roadway is structured so that pedestrians cannot enter, such as highways, etc. ・ When transportation facilities such as pedestrian crossings and traffic lights are enriched Think about the degree.

[4-3-2. Risk level focusing on cognition and avoidance]
For example, when the appearance frequency of pedestrians is very close to 0 around the planned travel route, pedestrians may appear that the driver does not assume. If this “unexpected” is interpreted as cognitive difficulty, it is caused by the following factors.

(A) Statistical appearance characteristics that are difficult to imagine
(A1) When the driver does not know the appearance statistics at all
(A2) When the appearance statistics themselves cannot be modeled and is close to a random number For example, in a residential area or the like, it is difficult to specify the appearance location of a pedestrian, and caution is required.

(B) Ambient environment that is difficult to recognize (such as jumping out from blind spots, appearance from hiding)
(C) Pedestrian behavior that is difficult or difficult to recognize In other words, the situation of cognitive difficulty is the result of both the “recognition target side (pedestrian etc.)” and the “recognition side (driver)” acting. Conceivable. However, since most of these cognitive difficulties are actually treated as human error on the driver's side, much research on risk prediction has been focused on human error detection. However, there are actually situations that cannot be avoided even if the driver is perfect. For example, the child jumping out just before corresponds to that.

Therefore, the following items can be defined by expressing the possibility of recognition and avoidance from the three viewpoints of a pedestrian, an in-vehicle device, and a driver.
P _R : Pedestrian's perception of the vehicle P _A : Pedestrian can avoid the vehicle (assuming that the vehicle is fully recognized)
V _R : Possibility of in-vehicle device recognition for pedestrians V _A : Possibility that the vehicle can avoid pedestrians (assuming that pedestrians are fully recognized)
D _R : Driver's perception of pedestrians D _A : Driver's possibility of avoiding pedestrians (assuming that pedestrians are fully recognized)
When each of these factors is regarded as a probability value and expressed as a value between 0 and 1, for example, the following formula (1) is used to set the risk degree focusing on the recognition possibility and avoidance possibility to the minimum value 0 and the maximum value 100. It can be calculated. Here, w _X (X = P, V, D) is a load coefficient.

  This is called RA-PVD risk. The RA-PVD risk is merged with the simple risk R to form a composite risk. As the simplest example, it is conceivable that the following formula (2) is set as a composite risk.

R _all = R + R _RA-PVD (2)
Hereinafter, factors that change the RA-PVD risk are listed and briefly described.
[4-3-3. Vehicle factor]
(1) Self-vehicle speed Generally, the higher the speed of the vehicle, the lower the possibility of recognition and avoidance due to the above P, V and D, and the RA-PVD risk increases.

(2) Vehicle shape / size / weight The vehicle size / weight affects the degree of destruction of buildings and human damage and changes the risk of RA-PVD from the viewpoint of cross-sectional area and kinetic energy. . In addition, the following items tend to be similar.

-Body shape, tire size-Blind spot on vehicle structure (rear, measurement, etc.)
(3) Others The following items particularly affect the avoidability.

・ Improper maintenance, illegal modification, breakdown, equipment failure (chain etc.)
[4-3-4. Road environmental factors]
[4-3-4-1. Road factor]
(1) Road width In places where there is a large flow of people, the road is generally wide and there are many sidewalks, guardrails, traffic lights, etc. In such areas, the outlook is good and the probability of an accident based on cognitive difficulty is relatively low. However, roads with a large flow of people are also converging roads to facilities with high customer attraction (stations, schools, hospitals, etc.), so pedestrians who have difficulty in recognizing narrow roads with narrow branches, intersections with poor visibility, etc. Appearance can occur. Therefore, although the simple risk tends to increase when the road width is large, it is considered that the RA-PVD risk increases when the road width is small.

On the other hand, when viewed from the vehicle side, the width of the branch road on the travel route affects vehicle control and cognitive judgment, which also affects the RA-PVD risk.
In addition, obstacles, parked vehicles, oncoming vehicles, street trees, traffic signs, and the like can be cited as factors that narrow the equivalent effective road width.

(2) Road surface condition A bad road surface condition adversely affects both recognition and avoidance, and therefore the RA-PVD risk increases. The situations such as “slippery”, “not visible”, “driving difficulty” caused by snowy roads, rain, freezing, typhoon, landslide, flooding, flooding, etc. correspond to this.

(3) Intersection shape RA-PVD risk is considered to increase as the gradient increases.
The RA-PVD risk is considered to increase as the number of branches increases.

(4) Others The factors include curvature, road markings, gutters, and shoulder width.
[4-3-4-2. Transportation facilities]
When there is the following traffic facility, the combined risk can be set small even if the spatio-temporal distance between the pedestrian and the travel route is small.

(1) Traffic light When there is no traffic light, it is left to the pedestrian side crossing judgment. In this case, if it thinks simply, the probability of occurrence of an accident should be higher as the appearance frequency is higher. However, in most cases, the driver side has a certain degree of recognition judgment ability, so that even if the appearance frequency is high, a collision is avoided in most cases. On the contrary, the mechanism that alerts the driver's side that the appearance frequency is high also operates. That is, an increase in the simple risk level is a case where the RA-PVD risk level is decreased.

(2) Guardrail The possibility of avoidance increases, but the recognition possibility decreases. In general, the former effect is large, and it is considered that the RA-PVD risk is lowered overall. However, the degree of danger increases depending on the vehicle speed, road width, vehicle size, and human error.

(3) Sidewalk RA-PVD risk is lower when there is no sidewalk than when there is no sidewalk.
(4) Pedestrian bridge The risk is almost zero.

(5) Pedestrian crossing RA-PVD risk is lower when there is no pedestrian crossing.
(6) Underpass The risk is almost zero.

(7) Others The presence or absence of signs, mirrors, breakers, or defective conditions affect the risk level.
[4-3-5. Traffic condition factors]
Traffic volume, traffic jam, positional relationship in lane, presence / absence of ambulance, unreasonable overtaking, wobble, blind spot by large vehicle or appearance from blind spot by two-wheeled vehicle, meandering operation, presence / absence of crackdown, signal timing, etc. affect the risk level It is thought to give.

[4-3-6. Geographic factors]
Building, mountain, river, roadside tree, undulation, coast, lake shore, flatland, countryside, residential area, tourist destination, castle, government office, company, factory, station, port, park, hospital, school, school zone, shopping street, bank It is considered that the risk level is changed by affecting the possibility of recognizing, avoiding, and the flow of people using keywords such as shopping center and urban structure.

[4-3-7. Geographic factors]
(1) Appearance form Since this is dealt with in the above-mentioned "flow of people", it is omitted.

(2) Behavior / Behavior The risk is affected by keywords such as human movement, walking speed, movement, illegality, popping out, wandering, carelessness, behavioral characteristics (especially children and suspicious persons who are difficult to predict). .

(3) Characteristic as a group It affects the risk level by using keywords such as school groups, commuting groups, urban excursions, demonstrations, parades, queues, gatherings at events, and flocks at department stores and stations.

(4) Individual attributes (profile)
Attribute information describing the characteristics of pedestrians includes age, gender, address, occupation, purpose, means of transportation, visual acuity, hearing, physique, personality, hobbies and the like. All of these affect the risk, but the attributes that can be collected within a relatively small range of privacy are already described in the PT data as described above {age, gender, address, occupation, purpose, traffic Means}. These are already covered above and will be omitted.

[4-3-8. Human factors in the passenger compartment]
(1) Driver The human error of the driver is not dealt with directly in this embodiment because there are already many research achievements. Keywords related to risk are elderly, lack of information, failure to confirm safety, driving skill, unlicensed, failure to confirm mirror, illegal behavior, drunk driving, poor physical condition, lack of sleep, mental state, temper, rush, assumption , Accidental operation of the accelerator, steering wheel and brake. Furthermore, the same personal attribute (profile) as described in the above pedestrian factor also affects.

(2) Passenger Children, elderly people, seat belts, etc. are keywords related to the risk level. Furthermore, the same personal attribute (profile) as described in the above pedestrian factor also affects.

[4-4. Application to orbit prediction]
[4-1-6. If the regression coefficient matrix B obtained in [Estimation of regression coefficient matrix] is applied not only to the planned traveling route but also to the entire target area, T sample times assigned to half the column size 2T of the state matrix X A spatial distribution of a 70-dimensional pedestrian certainty vector is obtained. At this stage, since the individual confidence vectors represent a plurality of pedestrians in combination, the spatiotemporal trajectory of each pedestrian cannot be reproduced. However, 70 representing a pedestrian class such as “20s, women, shopping purpose, office worker” for a pedestrian certainty vector c (70 × 1) within a certain spatiotemporal range (for example, the danger range around the above-described travel route). Using the evaluation function p = c ^T c _q expressed by the inner product with the dimensional certainty factor vector c _q , the position coordinates at which p at each time takes the maximum value are connected, and a sequence of points that are continuous in time and space is extracted. If possible, it can be a candidate for a spatiotemporal trajectory of a pedestrian who is “20s, woman, shopping purpose, office worker” (see FIG. 8).

[5. effect]
In the danger prediction system 1 of the present embodiment, a time-space trajectory in which a movement path position indicating the position of a movement path where a pedestrian has moved in the past and a time when the pedestrian is present at the movement path position is correlated. Spatial trajectory data and pedestrian attribute data indicating pedestrian attributes acquire learning results learned using the human flow database 12 stored for each of a plurality of pedestrians (S40, S90), and on the planned travel route When a representative point Q _I (I = 1, 2,..., I _Q ) and a passing time are designated, the possibility of a pedestrian appearing at the representative point Q _I is shown based on the learning result. The total pedestrian certainty C _I is predicted (S100 to S220).

  The danger prediction system 1 configured as described above learns the pedestrian's spatio-temporal trajectory data in the past, and predicts the possibility of a pedestrian appearing at a specified time and position. For this reason, if time and a position are designated, the user of the danger prediction system 1 can know the possibility that a pedestrian will appear at the designated time and position when the time and position are designated. It is possible to cope with the appearance of pedestrians with a sufficient margin.

Further, the risk prediction system 1 assumes that the pedestrian attribute constitutes a pedestrian's certainty, learns by a model in which the certainty is estimated using time and position as explanatory variables, and all pedestrians at the representative point Q _I. The certainty factor C _I is calculated. For this reason, the driver can know the possibility of a pedestrian appearing on the planned travel route at the time when the representative point Q _I is designated. Can be dealt with.

The risk prediction system 1 selects N _I present roads that connect to the representative point Q _I, it calculates the confidence in R _I-number of space-time point in the walking range of M _I component along each road (S200 ), The sum of the certainty factors at the R _I spatio-temporal points is calculated as a simple risk (S220). Thereby, not only the representative point Q _I on the planned travel route, but also the presence of pedestrians around the planned travel route can be considered, and the possibility of appearance of pedestrians can be predicted.

The risk prediction system 1, gender, age, and the walking range of M _I fraction corresponding to the class composed of transportation R _I number of time to set the spatial points (S160). Thereby, the range around the planned travel route used for calculating the simple risk can be appropriately set according to the gender, age, and transportation means of the pedestrian.

  In addition, when the extracted data is not sufficient as the learning data amount (S50: NO), the risk prediction system 1 performs the regression analysis by sampling at a higher time density than the other time zones in the running time zone. Learning by lowering the matrix (S80) or extracting data from other areas where the time variation pattern of the total number of people for each pedestrian attribute in the area is similar from the human flow database 12, Learning is performed using the extracted data (S70). This makes it possible to perform prediction even when the learning data is not sufficient.

  In the embodiment described above, the danger prediction system 1 is the prediction system in the present invention, the human flow database 12 is the mobile database in the present invention, the processes in S10 to S90 are the acquisition means in the present invention, and the processes in S100 to S220 are in the present invention. The predicting means in step S220 is a risk degree calculating means in the present invention.

  As mentioned above, although one Embodiment of this invention was described, this invention is not limited to the said embodiment, As long as it belongs to the technical scope of this invention, a various form can be taken.

  DESCRIPTION OF SYMBOLS 1 ... Danger prediction system, 10 ... Database part, 11 ... Car navigation apparatus, 12 ... Human flow database, 13 ... Search interface, 14 ... Spatio-temporal orbit data extraction and processing part, 20 ... Learning part, 21 ... Data learning part, DESCRIPTION OF SYMBOLS 22 ... Prediction characteristic database, 23 ... Prediction characteristic control part, 30 ... Condition setting part, 31 ... Prediction condition setting part, 32 ... Planned travel path mapping part, 40 ... Prediction part, 41 ... Pedestrian route estimation part, 42 ... Walking Person reliability calculation unit, 43 ... spatiotemporal distance calculation unit, 44 ... risk level calculation unit, 45 ... notification unit

Claims (3)

Spatio-temporal trajectory data indicating a spatio-temporal trajectory in which a movement path position indicating a position of a movement path in which the moving body has moved in the past and a time when the moving body existed at the movement path position are associated with each other, and the moving body Acquisition means (S10 to S90) for acquiring the learning result obtained by using the moving object database (12) in which the moving object attribute data indicating the attribute of the moving object is stored for each of the plurality of moving objects;
When a spatiotemporal point indicating time and position is designated, the designated spatiotemporal point is designated as a designated spatiotemporal point, and based on the learning result obtained by the obtaining means, the designated spatiotemporal point Prediction means (S100 to S220) for predicting the possibility of appearance of a moving object ,
The moving body is a human;
The learning result is
It is the result of learning by regression analysis assuming that the certainty about the human presence and attributes regresses to time and place,
The predicting means calculates at least a part of a planned travel route, which is a route on which the vehicle is scheduled to travel, as the designated time-space point, and calculates the certainty factor as a possibility that the person appears at the designated time-space point. And
The predicting means calculates the certainty factor at a plurality of the spatiotemporal points within a preset spatiotemporal range including a portion designated as the designated spatiotemporal point in the planned travel route,
A risk degree calculation means (S220) for calculating a sum of the certainty degrees calculated at the plurality of spatio-temporal points as a risk level;
The acquisition means includes
When the number of moving objects to be learned in the spatiotemporal range is small, the regression analysis state is obtained by sampling at a higher time density than the other time zones in the time zone of the spatiotemporal range. A prediction system (1) , wherein learning is performed by lowering a matrix and the result is acquired as the learning result . Spatio-temporal trajectory data indicating a spatio-temporal trajectory in which a movement path position indicating a position of a movement path in which the moving body has moved in the past and a time when the moving body existed at the movement path position are associated with each other, and the moving body Acquisition means (S10 to S90) for acquiring the learning result obtained by using the moving object database (12) in which the moving object attribute data indicating the attribute of the moving object is stored for each of the plurality of moving objects;
When a spatiotemporal point indicating time and position is designated, the designated spatiotemporal point is designated as a designated spatiotemporal point, and based on the learning result obtained by the obtaining means, the designated spatiotemporal point Prediction means (S100 to S220) for predicting the possibility of appearance of a moving object,
The moving body is a human;
The learning result is
It is the result of learning by regression analysis assuming that the certainty about the human presence and attributes regresses to time and place,
The predicting means calculates at least a part of a planned travel route, which is a route on which the vehicle is scheduled to travel, as the designated time-space point, and calculates the certainty factor as a possibility that the person appears at the designated time-space point. And
The predicting means calculates the certainty factor at a plurality of the spatiotemporal points within a preset spatiotemporal range including a portion designated as the designated spatiotemporal point in the planned travel route,
A risk degree calculation means (S220) for calculating a sum of the certainty degrees calculated at the plurality of spatio-temporal points as a risk level;
The acquisition means includes
When the number of pedestrians to be learned in the spatio-temporal range is small, the movement pattern of the pedestrian is extracted from the mobile object database data in other areas similar to the spatio-temporal range, forecasting system that is characterized in that by using the extracted data learned, to obtain this result as the learning result. The prediction means includes
The prediction system according to claim 1 or 2 , wherein the spatiotemporal range is changed in accordance with a class constituted by sex, age, and means of transportation.