JPH07254099A - Sudden event detector in road traffic - Google Patents

Abstract

PURPOSE:To highly accurately detect sudden events such as accidents ad fault vehicles, etc., in road traffic. CONSTITUTION:This detector is provided with a traffic flow data detection means 14 for detecting traffic flow data 13 relating to the road traffic of vehicles in the plural divided sections 12 of a road 10 where the vehicles pass through in a fixed cycle T, a neural network 16 for inputting the respective traffic flow data 13 in the continuous plural cycles T in the continuous plural sections 12 detected in the traffic flow data detection means 13 to respective input units constituting an input layer and outputting the generation presence/ absence information 15 of the sudden events relating to the vehicle traffic in a measurement object section 12a from an output unit constituting an output layer and a neural network learning means 19 for making the neural network 16 learn with proven result generation presence/absence information 17 in the measurement object section 12a corresponding to the generation presence/absence information 15 outputted from the output unit as teaching information 18.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a traffic control system for effectively using roads, and in particular, a neural network is used to objectively detect a traffic accident, a sudden event such as a broken vehicle, or the like, and an unexpected event in road traffic. Regarding a detection device.

[0002]

2. Description of the Related Art In a traffic monitoring system for monitoring and controlling the running state of a vehicle on a highway or a highway, the road to be monitored is divided into a plurality of sections, and an average of vehicles passing through the section is averaged. A vehicle detector is provided to detect the traffic volume indicating the number of vehicles passing through the relevant section per unit time or speed.

[0003] Then, each speed (section speed) and section traffic volume from the vehicle detectors installed in each section are monitored and displayed by a monitoring device arranged at a road management center, and the monitoring staff is in charge of natural congestion and It monitors the occurrence of accidents and broken cars. When they are found, they are reported to the media as road traffic information or displayed on electronic display boards provided at a plurality of locations on the road to alert the driver.

In addition, when a road is blocked due to an accident such as an accident or a broken vehicle, or when the traffic lane is restricted, the information is not only notified to each driver as soon as possible, but also the corresponding section. It is necessary to enforce one-sided lane regulation for vehicles passing by in the case of highways, and in the case of highways, approach regulation at adjacent interchanges.

On the other hand, in the case of natural traffic congestion, if the driver is aware that natural traffic congestion is occurring, it is not necessary to particularly enforce lane regulation or entry regulation at an interchange. Therefore, when a traffic jam is detected, the observer determines whether the traffic jam is caused by natural traffic jams or accidents such as the above-mentioned accidents or broken cars, and processes corresponding to the type. Need to run.

An incident detecting device for detecting an accident such as an accident or a broken car at the management center is constructed as shown in FIG. 6, for example. Average speed V of each vehicle passing through the corresponding section 1a in each section 1a of the road 1 and the corresponding section 1
A vehicle detector 2 is installed to measure a traffic volume Q indicating the number of vehicles passing through a at regular intervals T. Velocity V for each period T in each section 1a measured by each vehicle detector 2
And the traffic volume Q are transmitted to the unexpected event detection device 3 incorporated in the monitoring system provided in the management center.

The outbreak event detecting device 3 is composed of a kind of computer, and is internally provided with a judgment value memory 4, a measurement value buffer 5, a judgment processing section 6, a flag memory 7 and an output section 8.

Then, the sudden event detection device 3 has a fixed period T.
The section speed V and the section traffic volume Q of each section 1a, which are input for each section, are read, and it is judged from these values whether or not an unexpected event such as an accident or a broken vehicle has occurred in each section 1a, and the output unit 8 To display.

In the judgment value memory 4, a traffic jam judgment speed Vpf indicating a speed at which a traffic jam can be judged in the designated measurement target section 9, and a traffic jam judgment indicating a traffic volume at which a traffic jam can be judged. Traffic volume Qpf, sudden traffic congestion occurrence determination traffic volume Qpj that indicates the traffic volume that can be determined to be traffic congestion due to the above-mentioned sudden event, and speed change amount that can be determined to be the result of the accident event when the speed change during a certain period of time The speed change determination value Vpt shown, the speed difference determination value Vpl which shows the speed difference between the downstream section 10 and the downstream section 10 that can be determined to be the result of an unexpected event, and the like are preset.

Further, the measured value buffer 3 is read at the present time (the current cycle T) in the measurement target section 9 for measuring (determining) whether or not an unexpected event has occurred this time among the sections 1a. The present (current) speed Vp, t, the previous speed Vp, t-1 read one cycle before, the two-previous speed Vp, r-2 read two cycles before, the current traffic volume Qp, the moving average speed Vp3
1. The downstream section speed Vp + 1, t of the downstream section 10 read in the current cycle is stored.

The moving average speed Vp31 is the current cycle in the section 9 to be measured. One cycle ago. Each speed Vp two cycles ago,
It is the average velocity of t, Vp, t-1, and Vp, r-2. The content stored in the measurement buffer 5 is updated with new data every time one cycle T elapses. Further, it is also updated when the measurement target section 9 moves to the next section 1a.

In the flag memory 7, seven kinds of flags indicating the state of the immediately preceding cycle in the relevant measurement section 9 are set as required. Flag 0 indicates "natural flow", flag 1 "watch out for sudden traffic jam", flag 2 "starts for sudden traffic jam detection", flag 3 "continues for natural traffic jam", flag 4 "starts for natural traffic jam", flag 5 is "detection of sudden traffic jam", and flag 6 is "confirmation of unexpected traffic jam".

Then, when each of the plugs 5 and 6 for "detection of unexpected traffic jam" and "confirmation of unexpected traffic jam" is set, it is considered that an unexpected event has occurred in the section 9 to be measured, and the unexpected event occurrence information is output to the output unit 8. Also, the measurement target section is displayed and output.

Now, in the current cycle T, the measurement target section 9
When the respective speeds Vp, t and Vp + 1, t are input from the downstream section 10 and the current traffic volume Qp is input from the measurement target section 9, the respective speeds and traffic volumes of the measurement value buffer 5 are updated. To be done. Then, the determination processing unit 6 determines whether or not an unexpected event has occurred in the relevant measurement section 9 according to the flowcharts shown in FIGS. 7 and 8.

When the flow chart of FIG. 7 is started, if the current speed Vp, t of the measurement target section 9 is faster than the traffic congestion determination speed Vpf (step ST1), no traffic congestion has occurred,
The "natural flow" flag 1 is set in the flag memory 7 (S
T2). Then, the determination process in the current cycle T ends. If the current speed Vp, t is slower than the traffic congestion judgment speed Vpf, the flag set in the flag memory 7 in the previous cycle is checked (ST3).

"Natural flow" flag 0, "Be careful of sudden traffic congestion"
When the flag 1 or the "first occurrence of sudden traffic congestion detection" flag 2 is set, it is checked whether or not the current traffic volume Qp is greater than the traffic congestion traffic volume Qpf (ST4). If the number is large, the flag "natural traffic congestion first issue" 4 is set (ST5). Then, the determination process this time is ended.

If the difference is small, it is judged whether the difference speed between the moving average speed Vp31 and the current speed Vp, t is larger than the speed change judgment value Vpt (ST6). If it is smaller, the "outbreak situation caution" flag 1 is set (ST7). Then, the determination process this time is ended.

When the differential speed is large, the "first occurrence of sudden traffic congestion detection" flag 2 is set (ST8), and the secondary detection processing shown in FIG. 8 is executed (ST9). Further, in ST3, if the “burst congestion detection” flag 5 or the “burst congestion confirmation” flag 6 is set in the flag memory 7 of the previous cycle,
The "outbreak congestion detection" flag 5 is set (ST12). Then, in ST9, the secondary detection process is executed.

Further, in ST3, if the "natural traffic continuation" flag 3 or the "natural traffic first departure" flag 4 is set in the flag memory 7 of the previous cycle, in ST10, the current traffic volume Qp suddenly causes traffic congestion. Judgment traffic Qpj
Check if more. If the number is large, the "continuation of natural traffic" flag 3 is set (ST11), and the determination process this time is ended. In addition, if the number is small, "the first occurrence of sudden traffic congestion detection"
The flag 2 is set (ST8), and the secondary detection process shown in FIG. 8 is executed (ST9).

When the secondary detection process of FIG. 8 is started, if the "burst congestion detection" flag 5 is already set in the flag memory 7 (ST13), the "burst congestion detection" flag 5 is set.
Is set (ST14). If the difference speed between the current speed Vp, t and the downstream section speed Vp + 1, t is larger than the speed difference determination value Vpl, the "burst congestion confirmation" flag 6 is set.
If the difference is small, the secondary detection process is executed again (ST18).

Further, when the "first occurrence of sudden traffic congestion detection" flag 2 is already set in the flag memory 7 (ST1
3), only when the same flag 2 is set twice in the past three cycles (ST17), the process proceeds to ST14. Two
If it is less than the number of times, the secondary detection process is performed again (ST1
8).

In such an unexpected event detecting device,
The current speed Vp, t in the measurement target section 9, the moving average speed Vp31, the current traffic volume Qp, and the current speed Vp + 1, t in the downstream section 10 are compared with each judgment value set in the judgment value memory 4 in advance. By doing so, it is possible to grasp the unexpected traffic congestion caused by the accident event such as an accident or a broken car, while distinguishing it from the natural traffic congestion. Therefore, the observer can detect a sudden event while staying at the management center, and can implement appropriate traffic regulation at the optimum timing.

[0023]

However, the above-mentioned sudden event detection device still has the following problems to be solved. That is, the "natural flow" flag 1, the sudden traffic congestion caution flag 2, which is set in the flag memory 7
The conditions for setting each flag such as the sudden traffic congestion confirmation flag 6, that is, the condition for determining that an unexpected event has occurred are:
All judgments are made based on the judgment values set in the judgment value memory 4.

All of these judgment values are artificial values obtained by the observer or engineer based on his many years of experience and intuition.
Therefore, there is an individual difference of the person who sets it, and it cannot be said that the value is always accurate. As a result, an incorrect judgment may be output.

Further, each value set in the judgment value memory 4 is a fixed value. However, on an actual road, even if the road has a relatively uniform shape such as a highway without traffic lights and intersections, natural congestion is likely to occur locally near the top of the uphill. This is because the driver is required to approach an uphill and depress the accelerator, but since the driver stays in the same accelerator state, he does not notice that the speed naturally decreases. Furthermore, natural congestion is likely to occur locally near the entrance of the tunnel. This is because the driver's instinct loosens the accelerator because the field of view suddenly gets dark. In addition, there is a large speed difference between before the uphill and above the uphill, although there is no traffic jam.

Therefore, there is a concern that ordinary natural traffic congestion may be determined as a traffic congestion due to a traffic accident, or that there is no traffic congestion at all, but it may be determined that the traffic congestion should be overlooked. In order to prevent such a situation, it is necessary to set a judgment value in the judgment value memory 4 for each section according to the road condition of the section. As described above, it is very difficult for the supervisor or the technician to set the optimum determination value which is different for each section, based on the intuition and experience.

The present invention has been made in view of the above circumstances, and detects a plurality of types of traffic flow data such as a section speed and a section traffic volume of each section of a road at a constant cycle, and this condition is satisfied. By judging the occurrence of a sudden event in the measurement target section using a neural network method from different types of traffic flow data, without automatically setting a large number of judgment values, by automatic learning,
An object of the present invention is to provide a device for detecting an unexpected event in road traffic that can accurately detect only an unexpected event.

[0028]

In order to solve the above-mentioned problems, in the accidental event detection device according to claim 1 of the present invention, as shown in FIG. 1, each of the roads 10 through which a vehicle passes is divided into a plurality of roads. Traffic flow data detecting means 14 for detecting the traffic flow data 13 related to the road traffic of vehicles in the section 12 at a constant cycle T, and each of a plurality of continuous cycle T in a plurality of continuous sections 12 detected by the traffic flow data detecting means 13. The neural network 1 in which the traffic flow data 13 is input to each input unit forming the input layer, and the output unit forming the output layer outputs the occurrence / non-occurrence information 15 about the occurrence of a sudden event related to vehicle traffic in the measurement target section 12a.
6 and a neural network learning means 19 for learning the neural network 16 by using the proven performance occurrence information 17 in the measurement target section 12a corresponding to the occurrence information 15 output from the output unit as the teaching information 18.

Further, in the invention of claim 2, the traffic flow data 13 in the sudden event detecting device of claim 1 passes through the section 12 corresponding to the section speed Vp indicating the vehicle speed in each section 12 and each cycle T. Section traffic volume Q indicating the number of vehicles
It consists of and.

Further, in claim 3, the above-mentioned traffic flow data detecting means 14 is arranged in each section of the road,
It is composed of a plurality of vehicle detection means for detecting the section speed and the section traffic volume in each section in a constant cycle.

According to another aspect of the present invention, there is provided a plurality of vehicle detecting means for detecting a section speed and a section traffic volume of a vehicle in each of a plurality of sections of a road on which the vehicle travels at a constant cycle. Each section speed and each section traffic in each of a plurality of consecutive sections including a measurement target section out of each section speed and each section traffic volume output from each vehicle detection unit constitute an input layer Multiple neural networks that are input to the unit and output from the output units that make up the output layer information about the occurrence or non-occurrence of a sudden event related to vehicle traffic in the section to be measured, and the occurrence information about the output from each output unit of the neural networks Corresponding to each of the measurement target sections corresponding to the And a plurality of neural networks learning means for learning neural networks.

Further, in the invention of claim 5, as input data of each input unit of each neural network in claim 4, the section speed in the current cycle of the measurement target section and the current cycle of the corresponding measurement target section are set. Section traffic volume, section speed in the current cycle of the downstream section of the measurement target section, section traffic volume in the current cycle of the downstream section of the measurement target section, and section in the cycle immediately before the measurement target section Speed, section traffic volume in the cycle immediately before the applicable measurement target section, section speed in the cycle immediately before the downstream section of the applicable measurement target section, and immediately before the downstream section of the applicable measurement target section The section traffic volume in the cycle is adopted.

According to another aspect of the present invention, there is provided a plurality of vehicle detecting means for detecting a section speed and a section traffic volume of a vehicle in a plurality of sections of a road on which the vehicle travels at a constant cycle. Each section speed and each section traffic volume in a plurality of consecutive cycles output from each vehicle detection means are input to each input unit constituting the input layer, and the vehicle passage in each section is performed from each output unit constituting the output layer. One neural network that outputs information on whether or not a sudden event has occurred, and a neural network that is trained by using the proven actual occurrence information on each section, which is output from each output unit and corresponds to the information on whether or not each section has occurred, as learning information. And a neural network learning means.

Further, in the invention of claim 7, as input data of each input unit of the neural network in claim 6, the section speed in the current cycle of each section and the section traffic volume in the current cycle of each section. , The section speed in the cycle immediately before each section and the section traffic volume in the cycle immediately before each section are adopted.

[0035]

In the apparatus for detecting an unexpected event according to claim 1 configured as described above, the traffic flow data detecting means detects the traffic flow data relating to the road traffic of the vehicle in each section 12 constituting the road at regular intervals T. To be done.

Each traffic flow data in each detected section and each cycle T, that is, a plurality of types of traffic flow data having different measurement conditions in a plurality of cycles from the past including the current cycle in a plurality of sections including the measurement target section Is obtained. Therefore, it is sufficient to determine whether or not an unexpected event such as an accident or a broken vehicle has occurred in the measurement target section by using these plural types of traffic flow data.

In the present invention, a neural network method is used as a means for this determination. Specifically, each traffic flow data under each of the above measurement conditions is input to each input unit of the input layer of the neural network, and the occurrence information of the occurrence of the unexpected event in the corresponding measurement target section is obtained from the output unit of the output layer. .

Of course, in the neural network, each weighting coefficient connecting each input unit of the input layer and each intermediate unit of the intermediate layer, and each weighting coefficient connecting each intermediate unit of the intermediate layer and the output unit of the output layer. Must be set to the correct value.

In view of this, in the present invention, information on whether or not an actual event such as an accident or a broken car actually occurred in the relevant measurement target section is obtained later, and this is used as teaching information in the neural network. Apply to train the neural network. Specifically, the weighting factors described above are corrected each time.

Since the information about the occurrence or non-occurrence of the results can be easily grasped by the monitor, it does not particularly impose a heavy burden on the monitor or the engineer. That is, it is not necessary to set a large number of determination values as in the conventional device.

In the second aspect, the section speed V and the section traffic volume Q are adopted as the traffic flow data. In the third aspect, vehicle detection means for detecting the section speed V and the section traffic volume Q is provided for each of the divided sections.

In the fourth aspect, a neural network is provided for each section. In the sixth aspect, one neural network is provided for the entire road.

Finally, it is sufficient to obtain information on the occurrence / non-occurrence of an unexpected event such as an accident or a broken vehicle for each section of the road.
Therefore, it depends on whether the neural network processing capability and the correct output are obtained depending on whether the occurrence information for each section is realized by one neural network or by plural independent neural networks for each section. It may be appropriately selected according to the learning ability.

In the case of realizing with one neural network, all the sections to be measured are to be input, and it is necessary to input the section speed V and the section traffic Q in a plurality of cycles in all the sections.

Further, in the case of realizing with a plurality of neural networks for each section, the section to be measured is one section, and the section speed V and the section traffic in a plurality of cycles of a predetermined plurality of sections including the corresponding section to be measured. It is only necessary to enter the quantity Q.

What kind of data is input to each input unit of the input layer of the neural network has an important influence on the accuracy of the information on the occurrence / non-occurrence of the sudden event in the measurement target section and the learning efficiency in learning. give.

Therefore, by specifying the type of traffic flow data as described in claim 5, the best result is obtained when a plurality of neural networks are used. Moreover, by specifying each data type as described in claim 7, the best result in the case of using one neural network is obtained.

[0048]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings. FIG. 3 is a block diagram showing a schematic configuration of an unexpected event detection device in road traffic that is configured by using a plurality of neural networks.

The road 21 to be monitored is divided into a plurality of sections 22. A vehicle detector 23 as a dedicated vehicle detecting means is provided in each section 22. In order to simplify the description, this road 21 is a highway in which the up lane and the down lane are separated, and each vehicle detector 23 is located in one of the lanes in which the vehicle travels in the direction indicated by the arrow in the figure. Suppose you want to detect the flow of a vehicle.

Each vehicle detector 23 has a section speed Vp, t indicated by the average speed of each vehicle traveling in the section 22 under its control and a vehicle passing through the corresponding section 22 at regular intervals T such as one minute. The section traffic volume Qp, t indicated by a number is detected.

The section speed Vp, t and the section traffic Qp,
The suffix [p] of t indicates the section number that specifies the section 22, and the suffix [t] indicates the measured period. For example, Vp, t and Qp, t indicate the section speed and section traffic volume measured at the current (current) cycle T in the section 22a to be measured. Further, Vp + 1, t-1 and Qp + 1, t-1 indicate the section speed and section traffic volume measured in the immediately preceding cycle T in the downstream section 22b adjacent to the downstream side of the measurement section 22a. .

Further, a surveillance camera is incorporated in each vehicle detector 23, and an image of the running condition of the vehicle in each section 22 corresponding to the self is captured and sent to the network learning device 25 corresponding to the section 22. To do.

Each vehicle detector 23 is equipped with a measurement buffer that stores and holds the section speed Vp, t and the section traffic volume Qp, t measured for each one-minute cycle T for two cycles. Is started, the section speed Vp, t measured in the cycle T this time
, The section traffic volume Qp, t, the section speed Vp, t-1, and the section traffic volume Qp, t-1 measured in the previous cycle T are sent to the neural network 24 corresponding to the section 22a to be measured.

At the same time, when the next period T is started, the vehicle detector 23 receives the above-mentioned four types of measurement data stored in the measurement buffer, that is, the four types of traffic flow data Kp, from its own measurement target section 22a. With respect to the neural network 24 corresponding to the upstream side section 22c adjacent to the upstream side, each section speed Vp + 1, t, Vp + 1, t-1 and each section traffic volume Qp + 1, t, Qp + 1, Send as t-1.

Therefore, this time, the neural network 24 corresponding to the measurement target section 22a has a total of eight traffic flow data Kp shown below from the vehicle detector 23 of the measurement target section 22a and the vehicle detector 23 of the downstream side section 22b.
Is entered.

Section speed Vp, t in the current cycle of the measurement target section 22a, section traffic volume Qp, t in the current cycle of the measurement target section 22a
t Section speed Vp + 1 in the current cycle of the downstream section 22b, t Section traffic volume Qp + 1, in the current cycle of the downstream section 22b
t Section speed V in the cycle immediately preceding the measurement target section 22a
p, t-1 Section traffic volume Qp, t-1 in the cycle immediately before the measurement target section 22a Section speed Vp + in the cycle immediately before the downstream section 22b
1, t-1 Section traffic volume Q in the cycle immediately preceding the downstream section 22b
p + 1, t-1 FIG. 4 is a block diagram showing a schematic configuration of each neural network 24.

As shown, the neural network 24 includes an input layer 26 including h input units 26a.
And an intermediate layer 27 including m intermediate units 27a,
The output layer 28 is composed of n output units 28a.

In FIG. 4, the output unit 28
The case where a is one (n = 1) is illustrated. Further, in the embodiment apparatus, the number of the input units 26a installed is eight corresponding to the above-mentioned traffic flow data Kp (h =
8). Therefore, each traffic flow data Kp described above is input to each of the first to eighth input units 26a.

Traffic data K for each input unit 26a
Instead of directly inputting p, each traffic flow data Kp is input after being normalized to a value of 0 to 1. That is, input values u1 (t), ..., Ui (t), ..., uh to each input unit 26a
(t) is a value obtained by dividing the maximum value (constant) expected over the entire section 22 by each traffic flow data Kp value. Time t at each input value ui (t) indicates a time series change of each cycle T described above. In the embodiment, since the cycle T is 1 minute, the unit of time t is minute (min).

Eight input values u1 (t) to uh (t) obtained by normalizing each traffic flow data Kp are called input patterns. Input layer 26
Each of the input units 26a constituting the above does not have any arithmetic function by itself. That is, each function fA indicating the operation of each input unit 26a is all one. Therefore, each function value fA1 (t) output from each input unit 26a,
, FAi (t), ..., fAh (t) are equal to the respective input values u1 (t), ..., ui (t), ..., uh (t) as shown in the equation (1).

FAi (t) = ui (t) (1) Each input unit 26a of the input layer 26 and each intermediate unit 27a of the intermediate layer 27 are connected with their own weighting factors. For example, the i-th input unit 26a and the j-th intermediate unit 27a have their own weighting factors a
connected by ij.

Each input unit 26 constituting the input layer 26
Each function value fA1 (t), ..., fAi (t), ..., Output from a
fAh (t) is transmitted to each intermediate unit 27a of the intermediate layer 27 as a multiplication value [fAi (t) · aij] obtained by multiplying the corresponding weighting factors ai1, ..., Aij, ..., Ahm.

A value obtained by adding a multiplication value [fAi (t) .aij (= ui (t) .aij)] for the number h of the input units 26a installed to each of the m intermediate units 27a constituting the intermediate layer 27. x
(t) is input. That is, the jth intermediate unit 27
a

[0064]

[Equation 1] Is entered.

Function f indicating the operation of each intermediate unit 27a
B is, for example, a sigmoid function f having the characteristics shown in FIG.
(x). f (x) = 1 / [1 + exp (-x)] (3) This sigmoid function f (x) is a monotonically increasing function, is differentiable, and satisfies the following conditions.

| F (x) | <α α: a finite value (α = 1 in the case of FIG. 2) Therefore, the function values fB1 (t), ..., FBj (t) output from the m intermediate units 27a. ), ..., fBm (t) is
It is the value obtained by substituting the input value xj (t) obtained by the equation (2) into the sigmoid function f (x) of the equation (3).

Each of the m intermediate units 27a of the intermediate layer 27
Are connected to n (n = 1) output units 28a of the output layer 28 with their own weighting factors. For example, j
Th intermediate unit 27a and kth output unit 28
They are connected to a with a weighting coefficient cjk peculiar to both.

The function values fB1 (t), ..., FBj (t), ..., fBm (t) output from the m intermediate units 27a are assigned to the output unit 28a of the output layer 28 by the corresponding weighting factors c1k. , ..., cjk, ..., cmk multiplied value [fBj
(t) · cjm].

In each of the n (n = 1) intermediate units 28a constituting the output layer 28, the number m of the intermediate units 27a is set.
A value x0 (t) obtained by adding the product value [fBj (t) · cjk] of minutes is input. That is, the kth output unit 28a

[0070]

[Equation 2] Is entered.

A function fC indicating the operation of the output unit 28a
Also has a sigmoid function f having the characteristics shown in FIG.
(x). Therefore, one function value fck (t) having a value of 0 to 1 output from n (n = 1) output units 28a is an output value of the neural network 24, that is, a measurement target section of the road 21. The information Xk (t) indicates whether or not an unexpected event has occurred at time t in 22a.

Each occurrence presence / absence information Xk (t) at each time t (each cycle T) output from the output unit 28a is
Since it is the value of the sigmoid function f (x) shown in FIG. 2,
The value is close to [1] or [0]. That is, [1]
The case indicates that an unexpected event has occurred, and the case of [0] indicates that no unexpected event has occurred.

Next, the operation of each neural network learning device 25 for learning each neural network 24 shown in FIG. 3 will be described. Each neural network learning device 25 receives the image of the traffic situation of the vehicle in each section 22 received from each vehicle detector 23 and displays it on the monitor. An observer or technician at the monitoring center constantly monitors the monitor screen, and when an unexpected event such as an accident or a broken car is displayed on the monitor screen, the screen is confirmed and the occurrence time t and the occurrence section p are checked. The neural network learning device 2 corresponding to the occurrence section 22 at the time of confirming the actual occurrence information Yp, t
Input to 5. Alternatively, the generated information is recorded in a notebook or the like for a certain period of time, and is collected and input to the corresponding neural network learning device 25 after the daily monitoring work is completed or before the monitoring work is started.

The record occurrence information Yp, t is binary data which is [1] when a sudden event has occurred and [0] when it has not occurred. p-th section 22
When a learning command is input from the outside, the neural network learning device 25 corresponding to 1 first obtains the error evaluation function E by the following equation (5).

[0075] E = [Xp (t) -Yp (t)] 2/2 ... (5) It should be noted that the actual occurrence or non-occurrence information Yp (t) is, wardens or technician has operated the input of the p-th of section 22 Each performance occurrence information Y
Based on p, t, with each time t as a variable, each presence / absence detailed information Xk (t) (= Xp output from each output unit 28a).
[1] or [0] of the measurement target section 22 corresponding to (t))
Is the value of.

Generally, the learning in the neural network is to determine the values of the respective weighting coefficients aij and cjk between the respective units so that the error evaluation function E becomes the minimum value. Therefore, each weighting coefficient aij, c in use
The correction amounts Δaij (t) and Δcjk (t) of jk can be obtained by the expression (6) by partially differentiating the error evaluation function E.

[0077]

[Equation 3]

The time Δt corresponds to the above-mentioned constant period T (= 1 minute) in the embodiment. The second term in the equation (6) is a term for stabilizing the convergence of learning, and ε and α are learning parameters for the weighting coefficient. That is, if the values of the parameters ε and α are excessively large, there is a concern that the parameters diverge and the occurrence information Xp, t of occurrence of a sudden event obtained at every constant period T changes. On the contrary, even if the value is excessively small, it takes a long time until the accuracy of the information on the occurrence or non-occurrence of the unexpected event obtained by the neural network 24 is improved.

Therefore, it is necessary to set the optimum value. Now, when the correction amounts Δaij (t) and Δcjk (t) of the weighting factors aij and cjk are calculated by the equation (6), the weighting factors aij (t) and cjk (t) currently in use are calculated. To the next cycle T (time t
+ Δt), each weighting coefficient aij (t + Δ shown in Eq. (7)
t), cjk (t + Δt).

Aij (t + Δt) = aij (t) + Δaij (t) cjk (t + Δt) = cjk (t) + Δcjk (t) (7) Therefore, the result generation presence / absence information Yp, t for a plurality of cycles is learned by the neural network. When inputting operation to the device 25,
This neural network learning device 25 has been described above.
The error evaluation function E is further improved by repeatedly executing the respective arithmetic processes from the equations (5) to (7) for the plurality of cycles.
Becomes smaller and the detection accuracy improves.

In the breakthrough event detecting apparatus shown in FIG. 3, in each section 22, each neural network 24 corresponding to the section 22 outputs the occurrence information Xp of the occurrence of the unexpected event in the section 22a. Therefore,
An observer resident at the monitoring center can surely know which section 22 of the road 21 the unexpected event has occurred.

When the unexpected event occurrence detecting device configured as described above is newly installed, the observer or engineer resident at the monitoring center outputs the occurrence of the unexpected event from each neural network 24. The actual state p of the section p and the time t corresponding to the section p and the occurrence time t is judged, and the information of the occurrence or non-existence of the actual result of [1] or [0] is collected, and the information of the actual occurrence or non-existence of the actual result is neural network learning. Input operation to the device 25. If this operation input work is carried out, for example, for one week or one month, the learning effect will work and the information Xp, t of the occurrence of the unexpected event output from each neural network 24 will be confirmed by the observer later, for example. The actual occurrence information Yp, t will be matched with the information Yp, t.

Further, if the device is operated for a long period of time, the accuracy of detecting the occurrence of a sudden event can be further improved.
Therefore, unlike the conventional device shown in FIG. 6, it is not necessary for a supervisor or an engineer to set each judgment value in the judgment value memory 4.
Of course, it is not necessary to set different determination values for each section 22. Therefore, not only the burden on the observer and the technician can be reduced, but also individual differences between setting persons and artificial setting errors can be prevented.

FIG. 5 is a schematic diagram showing a schematic structure of an unexpected event selecting device using one neural network according to another embodiment of the present invention. The same parts as those of the embodiment apparatus shown in FIG. 3 are designated by the same reference numerals. Therefore, detailed description of the overlapping portions is omitted.

Similar to the embodiment of FIG. 3, in each section 22 of the road 21 divided into n sections, the section speed Vp, t and the section traffic volume Qp, t of the vehicle passing through the section 22 are set to a constant period T. A vehicle detector 23a for detecting each of them is provided. Also,
A monitoring camera is incorporated in each vehicle detector 23a, and captures an image of the traveling state of the vehicle in the section 22 corresponding to the vehicle detector 23a and sends it to one network learning device 25a.

Each vehicle detector 23a is equipped with a measurement buffer that stores and holds the section speed Vp, t and the section traffic volume Qp, t measured for each one-minute cycle T over two cycles. When started, the section speed Vp, t measured in the cycle T this time
, The section traffic volume Qp, t, the section speed Vp, t-1 and the section traffic volume Qp, t-1 measured in the previous cycle T are sent to the neural network 24a.

Therefore, the neural network 24
Each time a new cycle T arrives at a, the vehicle detectors 23 of all sections 22 of the road 21 show a total (4 ×
n) Types of traffic flow data Kp are input.

Section speed V in the current cycle of each section 22
1, t… Vp, t… Vn, t Section traffic volume Q1, t… Qp, in the current cycle of each section 22
t ... Qn, t Section speed V1, t-1 in the cycle immediately preceding each section 22.
Vp, t-1 ... Vn, t-1 Section traffic volume Q1, t-1 in the immediately preceding cycle of each section
... Qp, t-1 ... Qn, t-1 The neural network 24a of this embodiment is the same as the neural network 24 of the previous embodiment shown in FIG. 4, except that the input layer 26 has (4 × n) input units 26a. (H = 4 × n). The intermediate layer 27 is composed of m intermediate units 27a as in the previous embodiment.
Further, the output layer 28 is composed of n output units 28a as shown by the dotted line.

Then, each of the (4 × n) input units 2
6a includes the above-mentioned (4 × n) types of traffic flow data K
p is normalized and input. Further, each function value fC1 (t), ..., FCk (t) calculated from the n output units 28a by the sigmoid function f (x) shown by the above-mentioned equation (3).
, ..., fCn (t) are output. Each output unit 2
8a, each function value fC1 (t), ..., fCk (t),
, FCn (t) corresponds to each section 22 from the 1st to the nth, occurrence information X 1, ..., Xp ,.
It becomes Xn.

Therefore, in the apparatus of this embodiment,
With one neural network 24a, each occurrence information X1 (t) ... Xn (t) approximated to [1] or [0] of an unexpected event in each n-divided section 22 can be grasped.

Next, one neural network 2 having n output units 28a constructed in this way is provided.
Neural network learning device 25 for learning 4a
The operation of a will be described.

The monitor or engineer resident at the monitoring center observes the monitor screen of the neural network learning device 25a, similarly to the device of the previous embodiment, and the occurrence section p in which the unexpected event actually occurred and the occurrence After confirming the time t, these values are input to the neural network learning device 25a as the performance occurrence information Yp, t of [1] or [0].

The neural network learning device 25a receives n pieces of occurrence information X1 (t) at the input time t,
, X1 (t), ..., Xn (t) are used to obtain the error evaluation function E by the following equation (8).

[0094]

[Equation 4]

Next, the weighting factors aij and cjk between the above-mentioned units are set so that the error evaluation function E becomes the minimum value.
Determine the value of. For this. Similar to the above, the correction amounts Δaij (t) and Δcjk of the weighting factors aij and cjk in use are used.
(t) is partially differentiated from the error evaluation function E by
Calculate with formula (9)

[0096]

[Equation 5]

Finally, the next cycle T (time t + Δ
Each corrected weighting coefficient aij (t + Δt), cjk in T)
Calculate (t + Δt) by the equation (10). aij (t + Δt) = aij (t) + Δaij (t) cjk (t + Δt) = cjk (t) + Δcjk (t) (10) Therefore, similar to the neural network learning device 25a of the above-described embodiment, the neural network 24
The weighting coefficients aij and cjk of a are corrected such that the respective output presence / absence information X1 (t) ... Xn (t) that are output match the actual values.

Therefore, it is possible to obtain substantially the same effect as that of the unexpected event occurrence detection device described above. further,
In this embodiment, one neural network 24a can obtain the occurrence information X1 (t) ... Xn (t) of occurrence of an unexpected event in all the sections 22, so that when a computer having a large arithmetic processing capability is used. The configuration can be greatly simplified as compared with the above-described embodiment apparatus.

The present invention is not limited to the above embodiments. In the embodiment apparatus shown in FIG.
Although the above-mentioned eight types of traffic flow data Kp are input to each neural network 24, the number and types of the traffic flow data Kp are not particularly limited. For example, the section speed Vp-1, t and the section traffic volume Qp-1, t detected in the section 22c on the upstream side of the section 22a to be measured may also be input. That is, the measurement target section 22 to be output
The occurrence information Xk (t) of the sudden event of a is the actual existence information Yp,
It suffices to select the traffic flow data Kp that accurately matches t.

[0100]

As described above, in the unexpected event detection device for road traffic according to the present invention, a plurality of traffic flow data such as a section speed and a section traffic volume of each section of a road are detected, Neural network method is used to detect the occurrence or non-occurrence of a sudden event in the measurement target section from a plurality of types of traffic flow data with different conditions.
Therefore, since it is not necessary to artificially set a large number of judgment values for judging the occurrence of an unexpected event, there is no individual difference or erroneous setting among the setters, and the accuracy of detecting the occurrence of an unexpected event can be improved. Furthermore, since road conditions such as uphills, downhills, curves, and tunnels on the road are automatically incorporated into the detection result, ideal sudden event detection is possible.

[Brief description of drawings]

FIG. 1 is a functional block diagram showing a schematic configuration of the present invention.

FIG. 2 is a diagram showing a sigmoid function adopted in the neural network of the sudden event detection device according to the embodiment of the present invention.

FIG. 3 is a block diagram showing a schematic configuration of the apparatus of the embodiment.

FIG. 4 is a schematic configuration diagram showing a neural network incorporated in the apparatus of the embodiment.

FIG. 5 is a block diagram showing a schematic configuration of an unexpected event detection device according to another embodiment of the present invention.

FIG. 6 is a block diagram showing a schematic configuration of a conventional unexpected event detection device.

FIG. 7 is a flowchart showing the operation of the conventional device.

FIG. 8 is a flow chart showing the operation of the same conventional device.

[Explanation of symbols]

21 ... Road, 22 ... Section, 22a ... Measurement target section, 22
b ... Downstream section, 23, 23a ... Vehicle detector, 24, 2
4a ... Neural network, 25, 25a ... Neural network learning device, 26 ... Input layer, 26a ... Input unit, 27 ... Intermediate layer, 27a ... Intermediate unit, 2
8 ... Output layer, 28a ... Output unit.

Claims (7)

[Claims] 1. A traffic flow data detecting means for detecting traffic flow data relating to road traffic of the vehicle in each of a plurality of sections of a road on which the vehicle travels, and a traffic flow data detecting means for detecting the traffic flow data. Each traffic flow data in a plurality of consecutive periods in a plurality of consecutive sections is input to each input unit that constitutes the input layer, and from the output unit that constitutes the output layer, information on the occurrence of sudden events related to vehicle traffic in the section to be measured And a neural network learning means for learning the neural network by using the proven performance occurrence information in the measurement target section corresponding to the occurrence information output from the output unit as teaching information. Device for detecting unexpected events in traffic. 2. The traffic flow data is composed of a section speed indicating a vehicle speed in each section and a section traffic volume indicating a number of vehicles passing through the section in each cycle. An unexpected event detection device in the road traffic described. 3. The traffic flow data detecting means is arranged in each section of the road, and is composed of a plurality of vehicle detecting means for detecting section speed and section traffic volume in each section at a constant cycle. The device for detecting an unexpected event in road traffic according to claim 2. 4. A plurality of vehicle detection means for detecting a section speed and a section traffic volume of the vehicle in each section divided into a plurality of roads on which the vehicle travels, and a plurality of vehicle detection means for outputting the output. Each section speed and each section traffic in a plurality of consecutive periods in a plurality of sections including the section to be measured of each section speed and each section traffic are input to each input unit that constitutes the input layer, and the output layer is A plurality of neural networks that output the presence / absence information of a sudden event related to vehicle traffic in the measurement target section from the output unit that constitutes the measurement object, and the respective measurement objects corresponding to the presence / absence information output from the output units of the plurality of neural networks A neural network that corresponds to each piece of proven performance information in the section as teaching information Unexpected incident detection device in the road traffic and a plurality of neural networks learning means for learning. 5. The input unit of each of the neural networks includes a section speed in a current cycle of a measurement target section of the corresponding neural network, a section traffic volume in a current cycle of the corresponding measurement target section, and the corresponding measurement. Section speed in the current cycle of the downstream section of the target section, section traffic volume in the current cycle of the downstream section of the applicable measurement target section, section speed in the cycle immediately preceding the applicable measurement target section, the applicable Traffic volume in the cycle immediately before the measurement target section, section speed in the cycle immediately before the downstream section of the relevant measurement target section, and cycle immediately before the downstream section of the relevant measurement target section 5. The sudden event detection device for road traffic according to claim 4, wherein the section traffic volume in is input. 6. A plurality of vehicle detection means for detecting a section speed and a section traffic volume of the vehicle in each section divided into a plurality of roads through which the vehicle travels, and a plurality of vehicle detection means for outputting the output. Each section speed and each section traffic volume in a plurality of consecutive cycles are input to each input unit that constitutes the input layer, and information about the occurrence or non-occurrence of an unexpected event regarding vehicle traffic in each section is output from each output unit that constitutes the output layer. One neural network for outputting, and a neural network learning means for learning the neural network by using, as teaching information, each proven actual occurrence information in each section corresponding to the occurrence information in each section output from each output unit. And an unexpected event detection device in road traffic. 7. The input unit of the neural network includes a section speed in a current cycle of each section, a section traffic volume in a current cycle of each section, and a section in a cycle immediately before each section. The sudden event detection device for road traffic according to claim 6, wherein the speed and the section traffic volume in the immediately preceding cycle of each section are input.