JP2023034603A - Control system for one-sided alternate passage and traffic light - Google Patents

Abstract

To provide a control system capable of performing traffic control in a one-sided alternate passage zone safely and more efficiently than the prior arts, and a traffic light.SOLUTION: A control system for one-sided alternate passage comprises: a first camera 2a for capturing an image from a far side to a stop position of one lane; a second camera 2b for capturing an image from a far side to a stop position of the other lane; and a control computer 1 for acquiring a state of a vehicle in each lane using images captured by the cameras 2a and 2b and controlling determination logics of signal lights 3 and 4. The control computer 1 stores at least a control program defining the number of vehicles C which have passed in a one-sided alternate passage zone per unit time as a positive commission, defining an average stop time of vehicles stopped at the stop position as a negative commission, and learning a determination logic in which the commissions become maximum. A traffic light acquires states of the pair of signal lights 3 and 4 disposed on the respective lanes and the vehicles C on the respective lanes.SELECTED DRAWING: Figure 6

Description

The present invention relates to a control system and a traffic signal for efficiently performing alternating one-way traffic. In particular, the present invention relates to a control system for actively switching traffic lights by grasping the number of vehicles entering a one-way alternate traffic section and their running conditions. The present invention also relates to a traffic signal capable of efficiently performing alternating one-way traffic.

When the width of the road is insufficient due to road construction or the like and vehicles cannot pass each other, one-way alternate traffic is performed in which vehicles alternately pass by using only one lane. As a method for alternately passing vehicles on both lanes, there are a method using a traffic signal for construction and a manual method by a traffic director. Traffic signals for construction work, which are currently the most popular, employ passive control that automatically switches signals at a predetermined time. With such a traffic signal, when the traffic volume increases, optimal control is not performed, and a problem may arise that only one lane is congested. In contrast, a skilled traffic director should consider a lane-switching method that maximizes the number of vehicles crossing the alternating one-way section per unit time, or minimizes the average waiting time for both lanes, while keeping safety in mind. We have learned empirically how to switch between lanes in a timely manner, and in some cases it is possible to realize a smoother one-way alternate traffic than traffic lights.

On the other hand, it takes a considerable amount of time to train traffic directors, and there is a chronic shortage of skilled traffic directors. Currently, it is difficult to allocate traffic guides to all alternating one-way sections, and traffic signals are used everywhere. Therefore, there is a demand for a technology that optimizes the time interval between the red and green lights of a traffic signal and controls traffic in a safe and short waiting time.

In Patent Document 1, a vehicle detector, a visible camera, or an infrared camera is provided to detect vehicles waiting for a signal at a stop line provided for road construction, and traffic congestion is alleviated based on this information. A road construction traffic signal controller is disclosed that adjusts the display time of a traffic light to allow for traffic lights. The signal control device of Patent Literature 1 measures the vehicle occupation range occupied by the vehicle in the image of the camera and uses it as an indicator of traffic congestion. However, the driving conditions of individual vehicles and the distance between vehicles during driving are not used as indicators for signal control. Possibly bad control.

In order to realize more efficient one-way alternating traffic with traffic lights, it is important to grasp the running state of vehicles approaching from a long distance. Conventionally, a technology has been used in which images of a road and its surroundings are acquired by a camera, and vehicles are detected by a program that has undergone deep learning. Non-Patent Document 1 discloses a technique for detecting a vehicle at a short or middle distance that occupies a relatively large area in an image. However, there has not been much research on how to quickly detect a vehicle that is far away and appears small in an image.

JP-A-2000-172987

M. Koziarski and B. Cyganek, “Impact of low resolution on image recognition with deep neural networks: An experimental study,” Int. J. Appl. Math. Comput. Sci., vol. 28, no. 4, pp. 735 -744, 2018, doi: 10.2478/amcs-2018-0056.

There is a need for a control system that can safely and more efficiently direct traffic on alternating one-way sections to minimize congestion. In particular, there is a demand for a control system that analyzes the number of vehicles entering a one-way alternate traffic section and the running conditions, and actively switches signals based on the obtained analysis results. Further, there is a need for a traffic signal that incorporates an active control system to provide safer and more efficient alternating one-way traffic.

The present invention has been made in view of the above problems to be solved, and aims to provide a control system and a traffic signal that can control traffic in alternating one-way traffic sections more safely and efficiently than before. This was made as an issue that should be addressed.

The present invention relates to a control system for alternating one-way traffic. The control system of the present invention includes a first camera that captures an image of one lane from a distance to a stop position, a second camera that captures an image of the other lane from a distance to the stop position, the first camera and the second camera. and a control computer for acquiring the state of the vehicle in each lane using the image captured by the camera, and for controlling the decision logic of the traffic light. In the alternate one-way traffic control system of the present invention, the control computer uses at least the number of vehicles passing per unit time passing through the alternate one-way traffic section as a positive reward, and the average stop time of vehicles stopping at a stop position. is a negative reward, and a control program that has learned a judgment logic that maximizes the reward is stored.

In the alternate one-way traffic control system of the present invention, learning of the control program includes the probability of vehicle appearance time intervals in each lane, the probability density distribution of vehicle speed when there is no vehicle in front of each vehicle, and the probability density distribution of vehicle speed when there is no vehicle in front of each vehicle. By repeatedly evaluating the reward when the predetermined judgment logic is continued for a certain period of time for the data generated by the simulator whose parameter is the probability density distribution of the inter-vehicle distance according to the vehicle speed when there is a car, It is preferable to learn the judgment logic that maximizes the .

The control computer of the control system of the present invention preferably includes an image processing section that extracts a distant vehicle using an image captured by the first camera or the second camera. In this case, the image processing unit determines the singular point of the vehicle (the vanishing point specified by the image processing unit of the present invention may be at a different position from the vanishing point specified by a general method. The vanishing point specified by the method of the invention is hereinafter referred to as a "singular point".

When the image processing unit extracts a vehicle existing in the distance, the singular point of the vehicle is determined by the image processing unit as a region of interest where the histogram of the motion vector direction in the image forms two peaks shifted by 180 degrees. , derive streamlines of motion vectors in the region of interest, and are estimated from the positions of the end points of the derived streamlines.

The present invention also provides a traffic signal for alternating one-way traffic. The traffic signal of the present invention acquires the state of the vehicle in each lane using a pair of traffic signals placed in each lane and an image captured by a camera, and a control computer that controls the judgment logic of the traffic signal. and have. The images used by the control computer are images captured by a first camera that captures an image of one lane from a distance to the stop position, and an image captured by a second camera that captures an image of the other lane from a distance to the stop position. . In the traffic signal of the present invention, the control computer gives a positive reward for at least the number of vehicles passing through the one-way alternate traffic section per unit time, and gives a negative reward for the average stopping time of the vehicles stopping at the stop position. , the control program that learned the judgment logic that maximizes the reward is stored. The pair of traffic lights are characterized by changing their display contents and display time under the control of a control program.

The control system for alternating one-way traffic of the present invention acquires the vehicle status in each lane by photographing vehicles approaching the stop position from a distance, thereby grasping the traffic status of the vehicle with higher accuracy than before, It is possible to efficiently guide one-way alternate traffic.

The control system of the present invention uses a learning model that rewards the number of traffic per unit time and the average stopping time of stopped vehicles for learning the judgment logic of the control program, thereby performing machine learning of the judgment logic. It is possible to acquire judgment logic comparable to that of a skilled traffic controller.

Since the control system of the present invention includes an image processing unit, it becomes possible to identify a vehicle approaching a one-way alternate traffic section from a farther distance than the conventional one. By identifying vehicles from a distance, the running conditions of each vehicle can be grasped more accurately, and more efficient one-way alternate traffic control becomes possible.

The traffic signal of the present invention efficiently switches signals for alternating one-way traffic under the control of a control program that acquires the state of vehicles in each lane from images captured by a camera of vehicles approaching from a distance to a stop position. It can be performed.

FIG. 1 is a block diagram showing the configuration of the control computer of the one-way alternate traffic control system of the present invention. FIG. 2 is a diagram showing an outline of a traffic simulator used for learning control programs. FIG. 3 is a block diagram showing a control program learning method. FIG. 4 is a diagram showing traffic conditions of vehicles when a learned control program actively controls a traffic simulator. FIG. 5 is a diagram showing the traffic state of vehicles when passive control is performed to switch signals at fixed intervals in the traffic simulator. FIG. 6 is a diagram conceptually showing the configuration of a traffic signal provided with the control computer of the present invention. FIG. 7 is a drawing-substituting photograph showing an example of a road image input from a camera to the image processing unit of the control computer. FIG. 8 is a drawing-substituting photograph of the estimation result of the region of interest in the image by the image processing unit. FIG. 9 is a diagram comparing the position of the singular point of the vehicle specified by the image processing unit of the present invention with the vanishing point specified by the conventional method. FIG. 10 is a diagram showing a vehicle being tracked by the image processing unit. FIG. 11 is a diagram showing an example of a two-dimensional analysis of a two-dimensional flow field on the ground surface in a three-dimensional flow field created by a vehicle in a region of interest in an image. FIG. 12 is a diagram showing an example of two-dimensional analysis of a three-dimensional flow field created by a vehicle in a region of interest in an image. FIG. 13 is a diagram showing an example of two-dimensional analysis of a three-dimensional flow field created by a vehicle in a region of interest in an image. FIG. 14 is a diagram showing an example of a three-dimensional analysis of a three-dimensional flow field created by a vehicle, in which the area of the vehicle is added to the two-dimensional region of interest in the image.

Hereinafter, a control system for alternating one-way traffic and a traffic signal according to the present invention will be described in detail with reference to the drawings. In this embodiment, first, an overview of the configuration of the control system will be described, and then the configuration of a traffic signal equipped with a control computer will be described.

The control system for alternating one-way traffic of this embodiment includes a first camera that captures an image of one lane from a distance to a stop position, a second camera that captures an image of the other lane from a distance to the stop position, and a first camera. and a control computer that acquires the vehicle state in each lane using the images of the second camera and the second camera and controls the decision logic of the traffic light.

FIG. 1 shows a block diagram of the configuration of a control computer 1 of the control system. The control computer of the present invention includes a control program and an image processing section. The control computer of this embodiment is composed of one computer, and aggregates both the traffic conditions of one lane and the other lane in which alternating traffic is performed.

Normally, one person per lane is in charge of one-way alternate traffic guidance by traffic guides, and guidance is provided while communicating with communication means such as walkie-talkies. With such a method, there is a problem that it is difficult to sufficiently obtain information such as the number of vehicles stopped on the opposite side and the stop time. On the other hand, in this embodiment, since the conditions of both lanes are integrated into one control computer 1, more efficient control becomes possible.

Efficient traffic control of alternating one-way traffic should preferably be performed by considering the following characteristics of alternating one-way traffic sections.
・ Traffic volume in both lanes (number of traffic per unit time)
・The average vehicle speed of each vehicle before reaching the one-way alternate traffic section ・The average acceleration of each vehicle before reaching the one-way alternate traffic section The probability of each vehicle appearance time interval when arriving in the state of .
- Probability density distribution of vehicle speed when there is no vehicle in front of each vehicle.
• Probability density distribution of acceleration when there is no vehicle in front of each vehicle.
・ Probability density distribution of inter-vehicle distance according to vehicle speed when there is a vehicle in front of each vehicle.

In consideration of these characteristics, the control program of the control computer of the present embodiment uses a traffic simulator to perform deep learning and various It learns optimal control that can respond to various traffic conditions.

Figure 2 shows the concept of a traffic simulator for reinforcement learning of control programs. In the figure, each square with a symbol C represents a vehicle traveling in each lane. A circle marked with a number 3 is a traffic signal for controlling one lane, and a circle marked with a number 4 is a traffic signal for controlling the other lane. On the simulator, the alternating one-way traffic section is represented by the area that crosses both lanes, and this section means that when one lane is traveling, the other does not enter this area. defined.

For each vehicle C, the traffic simulator provides coordinates with the center of the traffic lights 3 and 4 as the origin, speed, and acceleration data. For each vehicle C, the following three conditions are defined.
・ Accelerate if the speed is less than the specified speed.
・ If the traffic lights 3 and 4 in front of the lane you are traveling in are red, slow down.
・ When the distance to the preceding vehicle C approaches, the vehicle decelerates.

を生成し、制御プログラム内のＡＩに与える。 FIG. 3 is a block diagram showing a method of reinforcement learning performed by the control program using the traffic simulator. In the traffic simulator, the control program includes arrays having coordinate, speed, and acceleration values, which are data for each vehicle C,

is generated and given to AI in the control program.

The control program receives data for each vehicle C and determines actions within the traffic simulator to control the traffic lights 3, 4 to either red or green. In the initial stage of traffic control, only the simplest rule is established: when the opposite lane signal is red and there is no vehicle C in the one-way alternate traffic section, the signal is switched to green. In the following description, the state in which both the traffic lights 3 and 4 are controlled to be red will be abbreviated as action 1 "red". Also, the state where the traffic light 3 is red and the traffic light 4 is controlled to be green is abbreviated as action 2 "red blue", and the state where the traffic light 3 is green and the traffic light 4 is controlled to be red is abbreviated as action 3 "blue red".

A positive reward for the result of the action of the control program is defined as the number of vehicles C passing per unit time through the alternating one-way section. On the other hand, the average stop time of vehicle C stopped at the stop position is defined as a negative reward. In addition to this, when a sudden brake is generated, a negative reward according to the magnitude and rate of deceleration acceleration is stipulated.

The control program performs traffic control for a certain period of time using vehicle C data given from the traffic simulator. Then, the reward at the time of completion of control is evaluated, the judgment logic is corrected each time, and learning is repeated so as to maximize the reward. In the present embodiment, the ε-greedy method, in which a random action is taken according to the value of ε, is used to continuously search for the optimum action. . Furthermore, the traffic simulator provides the control program with data of the vehicle C based on various possible traffic attributes, and by repeating traffic control in the same way, the optimum traffic control logic that can adapt to various traffic conditions is acquired. be.

FIG. 4 shows the traffic state of the vehicle C when the learned control program actively controls the traffic simulator. FIG. 5 is a diagram showing the traffic state of vehicles when passive control is performed to switch signals at fixed intervals in the traffic simulator. In the control by the learned control program, every time the signal is switched from "red-green" to "blue-red", it is shown that all the stopped vehicles C can pass through the one-way alternate traffic section. there is On the other hand, in the passive control shown in FIG. 5, many vehicles have already stopped when the signal changes from green to red on the side of the traffic light 3, and they cannot pass by switching the signal once. It turns out that the vehicle is congested.

In this way, by using a control program that has learned the determination logic of the present embodiment, it becomes possible to control a traffic light that can properly control traffic.

In order to apply the control program of this embodiment to a traffic signal in a one-way alternate traffic section, it is necessary to grasp the traffic conditions of vehicles on the road where the traffic signal is actually installed. Moreover, for more appropriate control, it is important to grasp the running state of a vehicle approaching from a long distance. The control computer 1 of this embodiment includes a control program for controlling the traffic lights 3 and 4, and an image processing section for extracting a vehicle from an image of a vehicle traveling on a road.

FIG. 6 shows the configuration of a pair of traffic lights 3 and 4 for alternating one-way traffic, which are equipped with the control computer 1 of this embodiment. The control computer 1 has a first camera 2a for capturing images from a distance up to a stop position on one lane and a camera 2a for capturing a vehicle C from a distance on the other lane, in order to grasp the running state of the vehicle C approaching from a long distance. A second camera 2b is connected to take pictures up to the position. In addition, in this embodiment, a third camera 2c is provided for capturing an image of a one-way alternate traffic section and confirming that no vehicle is present in the section.

The control computer 1 has a communication function for capturing images captured by the first camera 2a, the second camera 2b, and the third camera 2c, and a communication function for transmitting control commands to the traffic lights 3 and 4. I have. The traffic lights 3 and 4 change their red or green display content and display time by the control program of the control computer.

The image processing section of the control computer identifies a region of interest in which a regularly moving object appears from the images captured by the cameras 2a and 2b. In addition, a singular point where a vehicle, which is a moving object, appears or disappears (as mentioned above, it may be a different position from a general vanishing point, so it is called a “singular point”, which is different from a “vanishing point”). ) and early detection of vehicles approaching from a distance. The image detection method executed by the image processing unit will be described in detail below.

FIG. 7(a) shows an example of a road image input from the first camera 2a to the image processing section of the control computer. FIG. 7(b) shows an example of a road image input from the second camera 2b to the image processing section of the control computer. The control computer records moving images for several minutes immediately before starting control of traffic lights 3 and 4, and displays video frames for each time included in the moving images and a background in which moving objects including vehicle C are deleted from the video frames. A region of interest is extracted by a background subtraction method of subtracting from an image.

A video frame I (x, y, t) is taken out from a moving image captured by the first camera 2a and the second camera 2b at the number of frames per second V (fps). At this time, the pixel value for each video frame is
I(x, y, t) ∈ {0, 1, ..., 255},
x∈{1, 2, . . . , N_x},
y∈{1, 2, . . . , N_y},
tε{1, 2, . . . , T}
can be expressed as Here, x is the horizontal coordinate in the image, y is the vertical coordinate in the image, t is the frame number, N_x is the number of pixels in the horizontal direction, N_y is the number of pixels in the vertical direction, and T is the total number of frames. Note that in this embodiment, the image is captured in grayscale.

A background image in which no moving object exists is estimated from the frame of frame number t. The median value of r video frames before and after is used for estimation. The background image I _BG (x, y, t) is estimated by the following equation. In this embodiment, 150 video frames of 5 seconds before and after are used.

ここで、関数H(x;th)は、閾値関数、あるいは活性化関数と呼ばれるものであり、

を表す。閾値Th1は、小さな値に設定すると移動物体を検知しやすくなるが、ノイズにも反応することになる。本実施形態では、Ｔｈ１＝５０としている。 Whether the pixel indicates a moving object, depending on whether the difference in pixel value between the video frame I(x, y, t) and the background image I _BG (x, y, t) is greater than or equal to a reference value (Th1) It is possible to determine whether or not When determining whether the moving object is a moving object, it is determined that the pixel is . If the reference value Th1 is set to a small value, it becomes easier to detect a moving object, but it also reacts to noise, so in this embodiment, Th1=50 is set. Then, a binary image I _FO (x, y, t) where 1 is assigned to a pixel determined to be a moving object and 0 to a pixel determined not to be a moving object is obtained by the following equation.

Here, the function H(x;th) is called a threshold function or an activation function,

represents If the threshold value Th1 is set to a small value, it becomes easier to detect moving objects, but it also reacts to noise. In this embodiment, Th1=50.

For the image of the road shown in FIG. 7A and the image of the road shown in FIG. A value image is shown in FIG.

で表すことができる。 Next, the region of interest of the resulting binary image I _FO (x,y,t) is divided into individual moving objects, each of which is indexed k (k=1, 2, . . . ). At this time, if the moving object is a rigid body such as a vehicle, in the next video frame, as the coordinates move, the size increases when the object approaches, and decreases when the object moves away. Denote the moving object with index k by I _FO,k (x,y,t), and the moving object in video frame t I _FO,k (x,y,t), k=1,2, . is represented by MO _k (x, y, t), and the scale transformation parameters are a and b.

_MOk (ax+ _bx ,ay+ _by ,t)=MOk'(x,y,t+1)

satisfy the relationship At this time, if the movement vector is represented by (b _x , b _y ), the movement direction φ is

can be expressed as

By observing about 20 to 100 moving objects, it is possible to obtain the distribution of the movement directions φ of the moving objects at each coordinate (x, y) in the image. When a vehicle is shown at the coordinates (x, y), the distribution of the moving direction φ has two sharp peaks shifted by 180 degrees due to the existence of the opposing vehicle. On the other hand, in places such as grass swaying in the wind and leaves of trees, such a characteristic distribution does not occur, so they can be clearly identified. In this way, by specifying a moving object in an image that has two peaks in which the distribution of the moving direction φ is shifted by 180 degrees, it is possible to specify a region of interest in which the vehicle to be tracked is photographed. .

Next, a method for estimating a point where an approaching vehicle appears or a singular point where a moving vehicle disappears in a moving image will be described.

In the region of interest, a number of moving objects assumed to be vehicles and their moving direction vectors are observed. Let (c _x , c _y ) be the barycentric coordinates in the image of each moving object, and (b _x , b _y ) be its moving direction vector. An example of the flow field created by the vehicle in the region of interest is shown in FIG. 11(a). In this figure, the starting point of each arrow represents the barycentric coordinates (c _x , c _y ) of the moving object, and the size and direction of the arrow represent the movement direction vector (b _x , b _y ). To obtain singular points from these, the following algorithm used in flow field analysis is used.
(Step 1) Set one point in the region of interest as an initial position.
(Step 2) Find the moving object closest to this initial position.
(Step 3) The next point is obtained by adding the moving direction vector of the moving object to this initial position.
(Step 4) Find the closest moving object to this point.
(Step 5) The next point is obtained by adding the moving direction vector of the moving object to this point.
(Step 6) The above processing is continued until no movement occurs. Here, the trajectory from the initial position to the final position is called a streamline. Also, the final position is called the terminal point.
(Step 7) The above processing is performed while changing the initial position to obtain a plurality of streamlines and end points.
(Step 8) Let the center of the terminal point be a singular point.
However, in steps 2, 3, 4 and 5, the moving direction vector of that point is actually estimated by interpolating the moving direction vectors of surrounding moving objects.

Here, although the height of the vehicle was not taken into consideration in FIG. 11(a), it actually becomes as shown in FIGS. 12-13. Fig. 12(a) shows the case where the road is straight, Fig. 12(b) shows the case where the road curves to the right, Fig. 13(a) shows the case where the road curves to the left, Fig. 13(b) shows If the road curves to the left, but the camera position is set higher than the height of all vehicles on that road. In these figures, the size of the vehicle (especially the vehicle height) is fixed, but in reality there are vehicles with various vehicle heights, so the arrows appear at various y-coordinates. For example, in the case of FIGS. 12(a), (b), and 13(b), it is possible to obtain a singular point by performing flow field analysis in the same manner as in FIG. 11(a). In this case, as shown in FIG. 11(b), there is a singular point completely inside the region of interest, so the singular point cannot be obtained by applying a normal flow field analysis. Therefore, in the present invention, as shown in FIG. 14, in addition to the x and y axes of the image, the area A of the moving object is added to a new axis to form a three-dimensional space (x, y, A). Consider a flow field by a three-dimensional vector (b _x , b _y , dA) obtained by adding a change amount dA in the area of the moving object in addition to the moving direction vector. As can be seen from FIG. 14, by considering the space to which the area of the moving object is added as a new axis, the singular point can be positioned at the end point of the region of interest. It becomes possible to estimate points.

In FIG. 9, the peculiar point specified by the processing of the image processing unit of the present embodiment is indicated by P, and the vanishing point defined by the conventional perspective drawing method is indicated by Q. Although the difference is small on a road that is straight to a long distance as in (a), in the case of a road that curves along the way as in (b), the peculiar It is clear that the points are located closer to the actual appearance and disappearance positions of the moving object than the vanishing points defined by the conventional perspective drawing method.

Normally, the singular point is far away, so even if a vehicle appears from there, the change from the background is slight and is almost at the noise level. Even the traffic director cannot recognize the appearance of the vehicle at the moment of its appearance, but recognizes the appearance of the approaching vehicle by its regular movement thereafter. The image processing unit of this embodiment follows this idea and tracks the movement of a moving object over several frames. As a result, even if the probability that a vehicle can be detected from noise in one still image is 0.1=10% (0.9=90% is noise), for example, one second, that is, 30% By using frames, the probability that all 30 frames are noise is (1-0.1) ³⁰ , so the probability that a vehicle can be detected is 1-(1-0.1) ³⁰ =0.96=96%. becomes. As a result, it became possible to easily detect the vehicle.

が一定の閾値Ｔｈ３以上の時、接近車両の出現と判断する。ここで、Ｎは大きいほど検知が遅れる一方で、確実な検知が可能になる。本実施形態では、道路状況に応じ、０．３ｓ～１．０ｓ、つまりＮ＝１０～３０の範囲を規定することで、接近車両を確実に検知することができる。また、車両の速度により、存在領域Ａ_ｎ，ｎ＝０，１，・・・，Ｎ－１は若干変わってくるが、車両が出現する特異点付近では、車両の速度の影響は小さく、統計的にはほとんど影響しないとみなすことができる。 A specific method of the image processing unit is as follows. An area A ₀ where the vehicle first appears (a set of image coordinates), an area A ₁ where the vehicle is located in the next frame, and an area A _n after n frames are converted into binary images by statistical processing. I _FO (x, y, t) is sought. Then, the image processing unit

is equal to or greater than a certain threshold value Th3, it is determined that an approaching vehicle has appeared. Here, the larger N is, the more the detection is delayed, but the more reliable the detection becomes. In this embodiment, an approaching vehicle can be reliably detected by defining a range of 0.3 s to 1.0 s, that is, N=10 to 30, according to road conditions. In addition, although the presence area A _n , n=0, 1, . can be assumed to have little effect.

After the vehicle is detected near the singularity, the vehicle can be tracked by a Kalman Filter and the coordinates at each frame can be estimated. Matlab (Motion-Based Multiple Object Tracking, manufactured by Mathworks) can be used for implementation in the image processing unit. The distance, velocity, and acceleration data to the stop line can be obtained by geometric transformation that converts the camera coordinates and the real coordinates from the coordinates in each frame. FIG. 10 shows a drawing substituting photograph of the vehicle tracked by the image processing unit. The image processing unit recognizes the portion indicated by the rectangle in FIG. 10 as a vehicle and tracks it.

The early vehicle detection capability of the image processing unit of this embodiment, ie, how quickly an approaching vehicle can be detected, was compared with a method using a convolutional neural network (R-CNN), which is often used for object shape recognition and detection. Matlab was used for implementation. A convolutional neural network is a network formed from an image input layer, a 2D convolutional layer of the convolutional neural network, a rectified linear unit (ReLU) layer, a max pooling layer, a fully connected layer, a Softmax layer, and a classification output layer. A convolutional neural network was utilized that was trained using the CIFAR-10 dataset containing 50,000 training images. These training images have 10 categories, including cars, and detection of these objects is possible. For learning, stochastic gradient descent (SGDM) with an initial learning rate of 0.001 was used, and the initial learning rate was decreased every 8 epochs, and learning was performed up to a total of 40 epochs. After confirming that the convolutional neural network worked well on the CIFAR-10 dataset, we additionally trained it on the video images collected in this embodiment. For training, the same stochastic gradient descent method (SGDM) was used with an initial learning rate of 0.001 and training was performed for 100 epochs.

A verification experiment for detecting an approaching vehicle was performed on moving images collected by the camera of the present embodiment, using processing by the image processing unit of the present embodiment and a technique using a convolutional neural network. For the four cases, the video was reversed and the frame detected by the vehicle was carefully determined by the naked eye (true appearance frame number), the frame number at the point where the stop line was reached, and the image processing unit first Table 1 shows the frame number in which the vehicle was detected and the frame number in which the convolutional neural network first detected the vehicle. All of the image processing units of this embodiment were able to detect the vehicle within 4 to 7 frames (within 0.23 seconds) after the true appearance point. In addition, it was found that the detection is faster by 11 to 91 frames (2 seconds on average) compared to the convolutional neural network. As a result of repeated demonstration experiments, qualitative results similar to those in Table 1 were obtained.

By the image processing by the image processing section described above, the control computer uses the images of the first camera 2a and the second camera 2b to extract data of the running state of the vehicle C approaching from a long distance. can be done. The extracted driving condition data of the vehicle C includes driving condition data equivalent to the data provided by the traffic simulator. can do.

The control computer of the alternate one-way traffic control system of this embodiment is equipped with a control program and an image processing unit, making it possible to identify vehicles approaching the alternate one-way traffic section from a farther distance than in the past. . By identifying vehicles from a distance, the running conditions of each vehicle can be grasped more accurately, and more efficient one-way alternate traffic control becomes possible.

The traffic signal of this embodiment can perform appropriate signal switching so that vehicles can safely and efficiently perform one-way alternate traffic under the control of the control program of the control system.

1 Control computer 2a First camera 2b Second camera 2c Third camera 3, 4 Traffic light c Vehicle P Singularity Q Vanishing point of conventional technology

Claims (5)

A control system for alternating one-way traffic,
a first camera that captures one lane from a distance to a stop position;
a second camera that captures an image of the other lane from a distance to the stop position;
A control computer that acquires the state of the vehicle in each lane using the images captured by the first camera and the second camera, and controls the judgment logic of the traffic light;
, and
The control computer treats at least the number of vehicles passing through the one-way alternate traffic section per unit time as a positive reward and the average stop time of vehicles stopping at a stop position as a negative reward, and the reward is the maximum. A control system for alternating one-way traffic, characterized in that it stores a control program that learns the judgment logic to be. The learning of the control program includes the probability of vehicle appearance time intervals in each lane, the probability density distribution of vehicle speed when there is no vehicle in front of each vehicle, and the distance between vehicles according to the vehicle speed when there is a vehicle in front of each vehicle. Learning the judgment logic that maximizes the reward by repeatedly evaluating the reward when the predetermined judgment logic is continued for a certain period of time for the data generated by the simulator with the probability density distribution of the distance as a parameter. The one-way alternate traffic control system according to claim 1, characterized in that: The control computer includes an image processing unit that extracts a vehicle existing in the distance using an image captured by the first camera or the second camera,
3. The control system for alternating one-way traffic according to claim 1, wherein the image processing unit performs processing for extracting, as a vehicle, an object that moves regularly from a singular point of the vehicle. The singular point of the vehicle is determined by the image processing unit assuming that a region in which the histogram of the motion vector direction in the image forms two peaks that are shifted by 180 degrees as a region of interest, and the streamline of the motion vector in the region of interest. 4. The control system for alternating one-way traffic according to claim 3, which is derived and estimated from the positions of the end points of the derived plurality of streamlines. A traffic signal for alternating one-way traffic,
a pair of traffic lights arranged in each lane;
Using the images taken by the first camera, which captures the distance from one lane to the stop position, and the second camera, which captures the distance from the other lane to the stop position, the state of the vehicle in each lane is acquired. and a control computer that controls the decision logic of the traffic light;
, and
The control computer treats at least the number of passing vehicles per unit time passing through the one-way alternate traffic section as a positive reward and the average stop time of vehicles stopping at a stop position as a negative reward, and the reward is the maximum. It stores a control program that learns the judgment logic that becomes
A traffic signal for alternating one-way traffic, wherein the pair of traffic signals change their display contents and display time according to the control program.

Images